JP2001343992A - Method and device for learning voice pattern model, computer readable recording medium with voice pattern model learning program recorded, method and device for voice recognition, and computer readable recording medium with its program recorded - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve the problem that proper voice pattern models for conversa tional voices cannot be provided. SOLUTION: An m-phoneme set extraction part 10 and a model learning part 3 are provided, and the part 10 uses reading voice m-phoneme set models to recognize 3-phoneme sets held in a conversational voice learning data memory 8 and extracts m-phoneme sets having low recognition rates, and the part 3 uses a time series of feature vectors of tokens held in the memory 8 to learn conversational voice m-phoneme set models with respect to only m-phoneme sets extracted by the part 10.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice pattern model learning apparatus, a voice pattern model learning method, and a voice pattern model learning method capable of appropriately learning a voice pattern model for an utterance having a high utterance speed and an ambiguous voice such as an interactive voice. The present invention relates to a computer-readable recording medium storing a speech pattern model learning program. Further, the present invention relates to a voice recognition device, a voice recognition method, and a computer-readable recording medium storing a voice recognition program capable of accurately recognizing an ambiguous voice having a fast utterance speed such as a dialogue voice. Things.

[0002]

2. Description of the Related Art Generally, in speech recognition, pattern matching is performed between a time series of a feature vector of a speech obtained by acoustic analysis of a speech and a speech pattern model obtained by modeling a pattern of the time series of the feature vector. Is realized by: As the voice pattern model, HMM (Hi
Dden Markov Model, Hidden Markov Model) is often used.

When an HMM is used as a voice pattern model, a phoneme is often used as a unit of a voice pattern to be modeled. Phonemes are consonants (/ s /, / h /, /
f /, / p /, / t /, / k /, / z /, / b /, / z
/, / G /, / m /, / n /, / r /) and vowels (/ a
/, / I /, / u /, / e /, / o /) and the like. If all phonemes appearing in Japanese are modeled by the HMM, an arbitrary word or sentence can be modeled by connecting the phoneme HMMs, and word speech or continuous speech can be recognized.

When a phoneme is modeled by an HMM, the phoneme is often segmented and modeled as follows. For example, the first phoneme of syllables / ha / and / hi / is / h
/ Is a subsequent phoneme even if it is the same phoneme / a /, / i /
And / h / preceding / a / and / h / preceding / i / have different acoustic feature amounts (hereinafter referred to as feature vectors). Such a phoneme having a different feature vector is called an abnormal sound. It is assumed that abnormal sounds are mainly caused by the context in which the phoneme appears, that is, the difference between the succeeding phoneme and the preceding phoneme. So, each phoneme is 1
Instead of using one model, many models use different models depending on the context. Particularly in recent years, R.A. Schwartz, Y .; Chow, “Improved Hidden Markov
MODELING OF PHONEMES OF C
ONTINUOUS SPEECH RECOGNIT
ION ”, IEEE INTERNATIONAL C
ONFERENCE ONACOUSTICS, SP
EECH, AND SIGNAL PROCESSI
NG, Vol. 3, 35.6.1-35.6.4 "(hereinafter referred to as reference 1) and the like, and a three-phoneme set (triphone) considering both preceding and succeeding phoneme contexts
Models are often used. For example, / aki / of / k /
Is (a) k (i), / hako // k /
Is (a) k (o) in a three-phoneme set. Here, the parentheses indicate the preceding or succeeding phonemes. The above (a) k
(I) and (a) k (o) are different sets of three phonemes because the subsequent phonemes are different. By using this three-phoneme set model, higher recognition performance than a normal phoneme model can be obtained. Note that the above (a) k (i) and (a) k
The notation such as (o) is hereinafter referred to as m phoneme set notation.

Next, a method for creating a three-phoneme set model will be described. FIG. 23 is a block diagram showing a configuration of an example of a conventional voice pattern model learning apparatus for learning a three-phoneme set model disclosed in the above-mentioned Document 1, for example. Note that, although the description is given in reference 1 using English phonemes, the same technique can be used in Japanese, so the following description will be given using Japanese phonemes as an example. In FIG. 23, reference numeral 100 denotes a learning data memory in which learning data of a three-phoneme set model is stored;
Is a time series of a speech feature vector included in the learning data stored in the learning data memory 100;
A model learning unit 400 for learning a phoneme set model; 400, a parameter of the three phoneme set model learned by the model learning unit 300; 500, a three phoneme set model for storing the learned parameter 400 of the three phoneme set model; Memory.

Next, the operation will be described. The learning data stored in the learning data memory 100 can be obtained by acoustic analysis of a voice read by a number of speakers reading a word or a sentence including various contexts of three phonemes, a human-to-person conversation voice, and the like. , A set of tokens which is a phoneme set notation indicating a time series of feature vectors and utterance contents, specifically, a time series of feature vectors obtained by acoustic analysis of a speech waveform of learning data and cut out for each phoneme section. 6 is a three-phoneme set table for associating three-phoneme set notations of three-phoneme sets existing in the learning data. FIG. 24 shows an example of the three phoneme set table.

As an acoustic analysis, for example, LPC (Linea
r Predictive Coding (linear prediction analysis) is used, and the feature vector is an LPC cepstrum. The segmentation for each phoneme section is performed by, for example, a human observing a spectrogram. It is also assumed that each token is given a three-phoneme set notation in which the phoneme name, preceding phoneme name, and subsequent phoneme name of the token are described. FIG. 25 shows an example of the three-phoneme set notation.

The three-phoneme model is a continuous distribution type HM
Suppose M. In this case, the structure of each three-phoneme set model is a five-state left-to-
-Use the right model. In FIG. 26, state 1
Is the initial state, and state 5 is the final state. Each three-phone set model has a state transition probability a _ij and a label output probability b
_ij (x). Here, the subscript ij indicates a transition from the state i to the state j, and the state transition probability a
_ij is the probability of a transition from state i to state j.
The label output probability b _ij (x) is a continuous distribution type H
In MM, it is represented by a multidimensional normal distribution. State transition probability a
_ij and the label output probability b _ij (x) are referred to as HMM parameters. HM to determine the parameters of the HMM
It is called M learning.

Next, the model learning operation will be described. (1) Learning procedure 1: The model learning unit 300 reads the three-phoneme set table held in the learning data memory 100,
According to the description contents of the three phoneme set table, the three phoneme sets are selected as learning targets. When the three phoneme set table is described, for example, as shown in FIG.
First selects (a) a (a), which is the first set of three phonemes, as a learning target.

(2) Learning procedure 2: Next, the model learning section 3
00 reads from the learning data memory 100 a time series 200 of feature vectors of all tokens having a three-phoneme set notation that matches the three-phoneme set selected in the above-described learning procedure 1, and selects the time series 200 using, for example, a forward / backward algorithm. The model is learned for the set of three phonemes. Upon completion of learning, the state transition probability model learning unit 300 is a parameter of the model ended learning a _{i j} and label output probabilities b _{ij (x),} and the 3 phoneme set notation, the 3 phoneme sets model memory 500 Send out. The three-phoneme set model memory 500 holds the parameters of the model for which learning has been completed and the three-phoneme set notation.

(3) Learning procedure 3: Model learning section 300
Refers to the three-phoneme set table held in the learning data memory 100, and continues in the order described in the three-phoneme set table until model learning is completed for all three-phoneme sets existing in the learning data. Are selected as learning targets, and the above learning procedure 2 is repeated. In this way, the model learning unit 300 learns a model for all three phoneme sets existing in the learning data.

[0012]

Since the conventional speech pattern model learning apparatus is configured as described above, it uses a three-phoneme set model that considers both the preceding and succeeding phoneme contexts and uses the three-phoneme context model to determine the difference between phoneme contexts. Although a model was created in consideration of the resulting deformation of phoneme feature vectors to improve recognition performance, there was a problem that it was not possible to cope with differences in speech styles such as sentence utterance, reading style, and dialogue style. That is, the deformation of the feature vector of the phoneme is caused not only by the phoneme context but also by the difference between the case of uttering as a word and the utterance style such as sentence utterance, reading tone, dialogue tone and the like. For example, when the word "reservation" is uttered as a word alone, when the text "I want to make a reservation tomorrow" is read out, and when this text is spoken to a person, the deformation state of the feature vector Will be different. Therefore, the voice pattern model learning apparatus has a problem that it is not possible to provide a voice pattern model appropriate for a dialogue voice when using only voices that read text as conventional learning data.

When learning a speech pattern model by simultaneously using speech data of various utterance styles, such as text-to-speech speech and dialogue speech with a person, various deformation modes of feature vectors differ. Since the feature vector is represented by one model, there is a problem that the accuracy of the voice pattern model is reduced.

Further, when learning a voice pattern model using learning data for each voice in various utterance styles, such as text-to-speech voices and dialogue voices with people, it is necessary to avoid a decrease in accuracy of the voice pattern model. However, there is a problem that the number of voice pattern models increases in proportion to the number of utterance styles to be learned.

[0015] In addition, in the case of a speech having a fast and ambiguous speech rate such as a dialogue voice, the feature vector is affected not only by the preceding and succeeding phonemes but also by the preceding and succeeding phonemes. There was a problem that sufficient learning could not be performed with the phoneme set model.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problem, and efficiently learns a voice pattern model even for an interactive voice without greatly increasing the number of voice pattern models. A voice pattern model learning device, a voice pattern model learning method, and a computer readable recording medium storing a voice pattern model learning program are provided.

Further, the present invention can efficiently perform, without a large increase in the number of voice pattern models, a voice having a fast and ambiguous utterance speed, such as a dialog voice, which cannot be sufficiently learned by a three-phoneme set model. It is another object of the present invention to provide a voice pattern model learning device, a voice pattern model learning method, and a computer readable recording medium storing a voice pattern model learning program for learning a voice pattern model in consideration of a longer phoneme environment.

Further, the present invention provides a speech recognition device, a speech recognition method, and a computer-readable recording medium on which a speech recognition program is recorded, for accurately recognizing speech having a fast and ambiguous speech rate such as conversational speech. The purpose is to obtain.

[0019]

A speech pattern model learning apparatus according to the present invention uses a read-aloud m-phoneme set model trained using text-to-speech voices, and a recognition rate of a predetermined threshold is determined from interactive voice learning data. It comprises m phoneme set extraction means for extracting the following m phoneme sets, and model learning means for learning a dialogue speech m phoneme set model for each extracted m phoneme set by using the dialogue speech learning data. is there.

In the speech pattern model learning apparatus according to the present invention, the m phoneme group extracting means includes a m phoneme group having the same m phoneme group notation in the dialog speech learning data.
A phoneme set is selected, the selected m-phoneme set is recognized using an m-phoneme set model of the read-aloud voice, and the selected m-phoneme set is extracted if the recognition rate is equal to or less than a predetermined threshold.

A speech pattern model learning apparatus according to the present invention uses a read-aloud m-phoneme set model learned using a text-to-speech voice, and a recognition rate based on conversational voice learning data is equal to or less than a first predetermined threshold. m phoneme set extraction means for extracting m phoneme sets, dialogue speech m phoneme set model learning means for learning a dialogue speech m phoneme set model using the dialogue speech learning data for each extracted m phoneme set,
An n-phoneme set extraction unit that extracts an n-phoneme set whose recognition rate is equal to or less than a second predetermined threshold from the dialogue speech learning data using the read-aloud m-phoneme set model and the dialogue m-phoneme set model; A dialogue speech n phoneme set model learning means for learning a dialogue speech n phoneme set model using the dialogue speech learning data for each extracted n phoneme set.

In the speech pattern model learning apparatus according to the present invention, the n phoneme group extracting means may include an n phoneme group extracting unit in which the number of data having the same n phoneme group notation is equal to or more than a predetermined number from the conversation speech learning data.
Selecting a phoneme set, recognizing the selected n phoneme set using the read-aloud speech m-phoneme set model and the dialogue speech m-phoneme set model, and selecting the n-phoneme set if the recognition rate is equal to or less than a second predetermined threshold value; This is for extracting n phoneme sets.

A speech recognition apparatus according to the present invention recognizes by connecting in parallel a m-phoneme set model, an m-phoneme set model, and an n-phoneme set model learned by the above-mentioned speech pattern model learning apparatus. A recognition target vocabulary model creating means for creating a speech pattern model for the target vocabulary; and a recognition means for recognizing the input speech using the speech pattern model for the recognition target vocabulary created by the recognition vocabulary model creating means. Things.

A speech pattern model learning method according to the present invention uses a m-phoneme set model of a read-aloud speech trained using a text-to-speech speech, and recognizes m-phoneme sets whose recognition rate is equal to or lower than a predetermined threshold from conversational speech learning data. Is extracted, and for each of the extracted m phoneme sets, a dialogue speech m phoneme set model is learned using the above dialogue speech learning data.

In the speech pattern model learning method according to the present invention, when extracting m phoneme sets, m phoneme sets in which the number of data having the same m phoneme set notation is equal to or more than a predetermined number are selected from conversational speech learning data. The selected m-phoneme set is recognized by using the m-phoneme-set model of the read-out voice, and the selected m-phoneme set is extracted if the recognition rate is equal to or less than a predetermined threshold.

A speech pattern model learning method according to the present invention uses a read-aloud m-phoneme set model learned using a text-to-speech voice, and a recognition rate is less than or equal to a first predetermined threshold value based on interactive voice learning data. An m-phoneme set is extracted, and for each of the extracted m-phoneme sets, a dialogue m-phoneme set model is learned using the dialogue speech learning data. The n phoneme sets whose recognition rate is equal to or less than a second predetermined threshold value are extracted from the conversational speech learning data using
For a phoneme set, a dialogue speech n phoneme set model is learned using the above-mentioned dialogue speech learning data.

In the speech pattern model learning method according to the present invention, when extracting n phoneme sets, an n phoneme set in which the number of data having the same n phoneme set notation is equal to or more than a predetermined number is selected from the interactive learning speech data. Recognizing the selected n-phoneme set using the read-aloud m-phoneme set model and the dialogue m-phoneme set model, and extracting the selected n-phoneme set if the recognition rate is equal to or less than a second predetermined threshold. Is what you do.

The speech recognition method according to the present invention provides a read-out speech m learned by a speech pattern model learning method.
A speech pattern model for the vocabulary to be recognized is created by connecting the phoneme set model, the dialogue speech m phoneme set model, and the dialogue speech n phoneme set model in parallel, and the input speech is generated using the created speech pattern model for the recognition target vocabulary. This is to recognize.

A computer-readable recording medium on which a speech pattern model learning program according to the present invention is recorded uses a read-aloud m-phoneme set model trained using text-to-speech voices, and a recognition rate is obtained from interactive voice learning data. M to extract m phoneme sets that are less than or equal to a predetermined threshold
The method includes a phoneme set extracting step and a dialogue m phoneme set model learning step of learning a dialogue m m phoneme set model using the dialogue speech learning data for the extracted m phoneme set.

[0030] In the computer-readable recording medium storing the speech pattern model learning program according to the present invention, the m-phoneme group extraction step includes the step of: if the number of data having the same m-phoneme group notation is more than a predetermined number from the interactive speech learning data. A m-phoneme group is selected, the m-phoneme group selected using the m-phoneme-speech model is read, and the m-phoneme group selected is extracted if the recognition rate is equal to or less than a predetermined threshold. .

A computer-readable recording medium on which a speech pattern model learning program according to the present invention is recorded uses a read-out speech m-phoneme set model learned by using text-to-speech speech, and a recognition rate is obtained from interactive speech learning data. An m-phoneme set extracting step of extracting m-phoneme sets that are equal to or smaller than a first predetermined threshold value; and a dialogue speech m for learning a dialogue speech m-phoneme set model using the dialogue speech learning data for each of the extracted m-phoneme sets. Using the phoneme set model learning step, and using the read-aloud m-phoneme set model and the dialogue m-phoneme set model, an n-phoneme set whose recognition rate is equal to or less than a second predetermined threshold is extracted from the dialogue speech learning data. an n-phoneme group extraction step, and a learning step for learning a dialogue n-phoneme group model for each of the extracted n-phoneme groups using the above dialogue voice learning data. Those having an audio n phoneme sets model learning step.

[0032] In the computer readable recording medium storing the speech pattern model learning program according to the present invention, the n phoneme group extracting step is such that the number of data having the same n phoneme group notation is more than a predetermined number from the interactive speech learning data. Selecting a certain n phoneme set and recognizing the selected n phoneme set using the read-aloud m-phoneme set model and the dialogue m-phoneme set model, and if the recognition rate is equal to or less than a second predetermined threshold, The selected n phoneme sets are extracted.

The computer-readable recording medium on which the voice recognition program according to the present invention is recorded is a read-out voice m learned by the voice pattern model learning method.
A recognition target vocabulary model creating step of creating a speech pattern model for the recognition target vocabulary by connecting the phoneme group model, the dialogue m m phoneme group model, and the dialogue n n phoneme group model in parallel; A recognition step of recognizing the input voice using the generated voice pattern model for the recognition target vocabulary.

[0034]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a block diagram showing a configuration of a speech pattern model learning device according to Embodiment 1 of the present invention. Referring to FIG. 3, reference numeral 3 denotes a m-phoneme group extracting unit (m-phoneme group extraction unit) which learns a m-phoneme group extraction model using a voice read out of a text with respect to each m-phoneme group stored in the read-aloud speech learning data memory 6. Means) a model learning section (model learning means) for learning a dialogue m m phoneme set model using the dialogue voice learning data stored in the dialogue voice learning data memory 8 for each m phoneme set extracted by 10; The time series of the feature vector of the reading voice included in the reading voice learning data memory 6, the time series 9 of the feature vector of the dialogue voice included in the dialogue learning data memory 8, and the 11 are extracted by the m phoneme set extraction unit 10. m phoneme set notation of m phoneme set, 12 is an extracted m phoneme set notation memory, 13 is a parameter of m-phoneme set model and m phoneme set Serial, the reading voice m phoneme sets model memory 14, 15 parameters and m phonemes sets representation of interactive voice m phoneme sets model 16 is an interactive voice m phoneme sets model memory. In the following, a description will be given of a three-phoneme set where m = 3 as an example. Also, typically, the read-aloud m m phoneme set model and the dialogue m m phoneme set model used in the first embodiment are both continuous distribution HMMs.

The read-aloud speech learning data memory 6 stores a time series of feature vectors and utterance contents obtained by acoustically analyzing a speech read out by a number of speakers from words and sentences including various m-phoneme contexts. The phonetic speech learning data includes a phoneme set notation shown, and specifically, the speech training data is a phoneme based on a time series of a feature vector obtained by acoustically analyzing a speech waveform obtained by reading a text. It is an m phoneme set table that associates a set of tokens extracted for each section with a set of m phoneme set notations of m phoneme sets. This m phoneme set table is described, for example, as shown in FIG. Here, as the acoustic analysis method, for example, LPC analysis is used as in the related art, and the feature vector is an LPC cepstrum. The segmentation for each phoneme section is performed by, for example, a human observing a spectrogram. In addition, the reading voice learning data memory 6
Are assigned m-phoneme set notation in which the phoneme name, preceding phoneme name, and subsequent phoneme name of each token are described. Each m-phoneme set notation is described as shown in FIG.

The dialogue speech learning data memory 8 stores dialogue speech learning data obtained by acoustic analysis of a person-to-person dialogue voice in various scenes, including a time series of feature vectors and a phoneme set notation indicating utterance contents. Specifically, the dialogue speech learning data includes a set of tokens obtained by extracting a time series of feature vectors obtained by acoustically analyzing a person-to-person dialogue speech waveform for each phoneme section, and m phonemes. It is an m phoneme set table which associates a set of m phoneme set notations with a set. This m-phoneme set table has the same format as the m-phoneme set table in the reading voice learning data memory 6. In addition, as a method of acoustic analysis, for example, LPC analysis is used in the same manner as the read-out speech learning data, and the feature vector is LP
C cepstrum. It is assumed that, for example, a human observes a spectrogram as in the case of the read-out voice learning data, and cuts out each phoneme section. The phoneme name of each token (each token is assigned a token number) is also stored in each token held in the dialogue voice learning data memory 8.
It is assumed that m phoneme group notation in which the preceding phoneme name and the subsequent phoneme name are described is given. Each m-phoneme set notation is the same as the m-phoneme set notation in the reading voice learning data memory 6.

The read-aloud speech learning data is learning data relating to relatively polite and clear utterances, such as a text-to-speech voice, whereas the interactive speech learning data is learning data relating to a natural human-to-person interactive voice. The feature is that the feature vector of the phoneme is severely deformed.

Next, the operation will be described. The speech pattern model learning device according to the first embodiment of the present invention creates a read-aloud m m-phoneme set model as follows, and stores it in the read-aloud m-phoneme set model memory 14. In this case, the speech pattern model learning device connects the input terminal A of the model learning unit 3 to the output terminal B1 of the speech learning data memory 6 so that the data held in the speech learning data memory 6 is stored in the model learning unit 3. Set to input to. Further, the output terminal C of the model learning unit 3 is connected to the input terminal D1 of the read-aloud m-phoneme set model memory 14. In this connection state, the speech pattern model learning device according to the first embodiment learns the read-aloud m m phoneme set model according to the following procedure.

(1) Reading voice model learning procedure 1: The model learning unit 3 reads the m phoneme set table held in the reading voice learning data memory 6, and according to the description contents of the m phoneme set table, the first m phoneme set. Is first selected as a learning target. In this case, if the m phoneme set table in which m = 3 is described as in the prior art, for example, as shown in FIG. 24, the model learning unit 3 first determines the first m phoneme set as (a) a (a ) Is selected as a learning target.

(2) Speech model learning procedure 2: The model learning section 3 has m phoneme set notation that matches the m phoneme set selected in the above read speech model learning procedure 1 or the following read speech model learning procedure 3. The time series 7 of the feature vector of the token is read out from the read-out speech learning data memory 6, and the model is learned for the m phoneme set selected using, for example, a forward / backward algorithm. When learning is completed, the model learning unit 3
Sends the state transition probability and label output probability, which are model parameters, and the m-phoneme set notation 13 thereof to the m-phoneme set model memory 14 for read-out speech. Reading voice m
The phoneme set model memory 14 holds the parameters of the model for which learning has been completed as described above and the m phoneme set notation 13 thereof.

(3) Reading voice model learning procedure 3: Thereafter, the model learning unit 3 refers to the m phoneme set table held in the reading voice learning data memory 6 and reads all m phonemes present in the reading voice learning data memory 6. Until the learning of the model for the set is completed, the next m-phoneme set is selected as a learning target according to the order described in the m-phoneme set table, and the above-described reading voice model learning procedure 2 is repeated to obtain all m-phoneme sets. The learning of the read-aloud m-phoneme set model for the set is ended.

Next, the model learning section 3 cooperates with the m-phoneme set extracting section 10 to learn the m-phoneme set model of the dialogue speech, and stores the result obtained by the learning in the dialogue m-phoneme set model memory 1.
6 is stored. Before starting the learning, the speech pattern model learning device connects the input terminal A of the model learning unit 3 to the output terminal B2 of the dialogue learning data memory 8 to thereby store the data held in the dialogue learning data memory 8. It is set so as to be input to the model learning unit 3. Further, the output terminal C of the model learning unit 3 is connected to the input terminal D2 of the conversational speech m phoneme group model memory 16. In this connection state, the speech pattern model learning apparatus according to the first embodiment learns the m-phoneme set model of the dialogue speech according to the following procedure.

The learning procedure of the dialogue m m phoneme set model is performed by using each of the tokens stored in the dialogue speech learning data memory 8 using the readout m m phoneme set model stored in the readout m m phoneme set model memory 14. And a procedure for extracting a m-phoneme set having a low recognition rate, and a procedure for learning a dialogue m-phoneme set model for each m-phoneme set extracted in this manner.

First, a procedure for extracting m phoneme sets having a low recognition rate will be described. (1) m-phoneme-set extraction procedure 1: The m-phoneme-set extraction unit 10 reads the parameters of all the read-aloud m-phoneme-set models and the m-phoneme-set notation from the read-aloud m-phoneme-set model memory 14.

(2) m phoneme set extraction procedure 2: The m phoneme set extraction unit 10 refers to the m phoneme set table held in the conversational speech learning data memory 8 and reads the m phoneme set table according to the description contents of the m phoneme set table. The m phoneme sets are selected as recognition targets. When the m-phoneme set table of m = 3 is described, for example, as shown in FIG. 24, the m-phoneme set extraction unit 10 first selects the first m-phoneme set (a) a (a) as a recognition target.

(3) m-phoneme-set extraction procedure 3: The m-phoneme-set extraction unit 10 extracts the m-phoneme-set notation that matches the m-phoneme-set selected in the above-mentioned m-phoneme-set extraction procedure 2 or the following m-phoneme-set extraction procedure 4. Time series 9 of feature vectors of all tokens
Is read from the conversational speech learning data memory 8, and for each of the read tokens, the likelihoods of all the read-out speech m-phoneme set models read in the above m-phoneme set extraction procedure 1 are calculated, and the highest likelihood is calculated. The notation of the m phoneme set of the indicated m phoneme set model is used as the recognition result of the token. The likelihood calculation uses, for example, a Viterbi algorithm.
m phoneme set extraction unit 10, after obtaining the recognition result for all tokens read, calculates the recognition rate R _t in accordance with the following equation (1).

R _t = C _t / N _t * 100.0 (1)

[0048] However, (1) subscript t denotes the m phoneme sets of type selected in formula, N _t is the number of tokens m phoneme pairs notation is m phoneme sets of types t, C _t is This is the number of tokens that were recognized correctly. Here, the correct recognition is defined as a case where the m phoneme set notation of each read token matches the m phoneme set notation of the m phoneme set model showing the highest likelihood.

The m phoneme set extraction unit 10 compares the recognition rate _Rt with a predetermined threshold value _Tr, and if it is equal to or smaller than the threshold value _Tr ,
The m phoneme set notation of the m phoneme set is sent to the extracted m phoneme set notation memory 12. The extracted m phoneme set notation memory 12 holds the input m phoneme set notation.

(4) m phoneme set extraction procedure 4: The m phoneme set extraction unit 10 refers to the m phoneme set table held in the dialogue speech learning data memory 8 and reads all m-phoneme set data existing in the dialogue speech learning data memory 8. In order to extract a low recognition rate from a phoneme set, the next m phoneme set is selected according to the order described in the m phoneme set table, and the m phoneme set extraction procedure 3 is repeated.

As described above, the m-phoneme set extraction unit 10 extracts all the m-phoneme sets whose recognition rate _Rt is equal to or less than the threshold value _Tr by performing the m-phoneme set extraction procedures 1 to 4. The m phoneme group notation is stored in the extracted m phoneme group notation memory 12.

Next, the procedure for learning the m-phoneme group model for dialogue speech for each m-phoneme group extracted as described above will be described.

(1) Extracted m-phoneme set model learning procedure 1: The model learning section 3 reads the m-phoneme set notation held in the extracted m-phoneme set notation memory 12 and holds the m-phoneme set notation memory 12. First, the first m phoneme set is selected as a learning target. If the contents of the extracted m-phoneme set notation memory 12 are as shown in FIG. 2, for example, the model learning unit 3 selects the first m-phoneme set (a) a (u) as a learning target.

(2) Extracted m phoneme set model learning procedure 2: The model learning unit 3 learns the extracted m phoneme set model learning procedure 1
Alternatively, the time series 9 of the feature vectors of all tokens having the m phoneme set notation that matches the m phoneme set selected in the learning procedure 3 of the extracted m phoneme set model is read from the interactive speech learning data memory 8 and forward-backed, for example. A dialogue m-phoneme set model for the selected m-phoneme set is learned using a word algorithm. Then, the model learning unit 3 sends the learned model parameters and the m-phoneme set notation to the dialogue-phone m-phoneme set model memory 16. The conversational speech m-phoneme set model memory 16 holds the received model parameters and m-phoneme set notation.

(3) Extracted m-phoneme set model learning procedure 3: Next, the model learning section 3 holds the extracted m-phoneme set notation memory 12 in the order stored in the extracted m-phoneme set notation memory 12. Then, the next m phoneme set is selected, and the above-described extracted m phoneme set model learning procedure 2 is repeated.

Next, a method for learning the m phoneme set model used by the voice pattern model learning apparatus according to the first embodiment will be described in detail. FIG. 3 is a flowchart showing the procedure of the voice pattern model learning method according to Embodiment 1 of the present invention. As shown in FIG. 3, in the speech pattern model learning apparatus according to the first embodiment, the learning procedure of the m phoneme set model is roughly divided into three steps.

First, in step ST101, which is the first step, the model learning unit 3 learns model parameters and m-phoneme set notation obtained as a result of learning and learning the read-out m-phoneme set model, and reads out the m-phoneme set model memory. 14 is stored.

Next, in step ST102, which is the second step, the m-phoneme set extraction unit 10 reads out the read speech m
Each token stored in the conversational speech learning data memory 8 is recognized by using the m-phoneme group model of the read-out voice stored in the phoneme group model memory 14 and an m-phoneme group having a low recognition rate is extracted.

Thereafter, in step ST103, which is the third step, the model learning section 3 uses the token stored in the conversational voice learning data memory 8 to store the second
For the m phoneme set extracted in the step, a m-phoneme set model of conversational speech is learned.

Next, the first to third steps will be described in detail. FIG. 4 is a flowchart showing a learning procedure of the read-aloud m-phoneme set model as the first step. Referring to FIG. 4, the learning procedure of the m-phoneme set model of the read-aloud speech will be described in detail.

In step ST201, the model learning section 3 reads the m-phoneme set table in the read-aloud speech learning data memory 6, and selects the first m-phoneme set as a learning target according to the description contents of the m-phoneme set table. As shown in FIG.
In this case, the model learning unit 3 selects the head m phoneme set (a) a (a) as a learning target.

The model learning unit 3 then proceeds to step ST2
02, a time series 7 of feature vectors of all tokens having m phoneme set notation that matches the m phoneme set selected in step ST201 or ST206.
Is read from the reading voice learning data memory 6.

Then, the model learning section 3 determines in step ST
At 203, a m-phoneme set model to be read aloud is learned for the m-phoneme set selected at step ST201 or ST206 using, for example, a forward / backward algorithm.

After that, the model learning section 3 executes step ST
At step 204, when the learning is completed, the above-described step ST2 is performed.
The state transition probability and the label output probability, which are the parameters of the model obtained as a result of the learning in step 03 and the m-phoneme group notation 13, are sent to the read-out m-phoneme group model memory 14. The read speech m-phoneme set model memory 14 holds the parameters of these models and the m-phoneme set notation 13 that have been received.

Next, the model learning section 3 determines in step ST2
At 05, the m-phoneme set table held in the read-speech learning data memory 6 is referred to, and it is determined whether or not the learning of the read-speech m-phoneme set model has been completed for all the m-phoneme sets existing in the read-speech learning data memory 6. Judge,
If learning has not been completed for all m phoneme sets,
In step ST206, the next m phoneme set is selected as a learning target according to the order described in the m phoneme set table, and the process returns to step ST202. On the other hand, when the learning is completed for all the m phoneme sets, the model learning unit 3 ends the reading voice m phoneme set model learning procedure.

Next, in a second step, the m-phoneme set extracting unit 10 stores the m-phoneme set data in the conversational speech learning data memory 8 using the read-out m-phoneme set model stored in the m-phoneme set model memory 14. Each token is recognized, and m phoneme sets having a low recognition rate are extracted. FIG.
Is a flowchart showing the extraction procedure of the second step. Hereinafter, this extraction procedure will be described in detail with reference to FIG.

First, the m-phoneme set extraction unit 10 determines in step S
At T301, the parameters of all the read m-phoneme group models and the m-phoneme set notation 13 are read from the read m-phoneme group model memory 14.

Next, the m-phoneme set extraction unit 10 determines in step S
At T302, the m phoneme set table stored in the conversational speech learning data memory 8 is read, and the first m phoneme set is selected as a recognition target according to the description contents of the m phoneme set table. The m phoneme set table is shown in FIG.
4, the m-phoneme set extraction unit 10 selects the first m-phoneme set (a) a (a) as a recognition target.

Then, in step ST303, the m-phoneme set extraction unit 10 interacts with the time series 9 of the feature vectors of all the tokens having the m-phoneme set notation that matches the m-phoneme set selected in step ST302 or ST308. Read from the voice learning data memory 8.

Thereafter, in step ST304, the m-phoneme set extraction unit 10 calculates the likelihood of each of the read tokens with all the read-out speech m-phoneme set models read in step ST301. The m phoneme set notation of the m phoneme set model showing high likelihood is
This is the recognition result of the token. As described above, for example, the Viterbi algorithm is used for the likelihood calculation.
m phoneme set extraction unit 10, after obtaining the recognition result for all tokens read, calculates the recognition rate R _t in accordance with the equation (1).

Next, the m-phoneme set extraction unit 10 executes step S
In T305, it is compared with a threshold value _{T r} which defines a recognition rate _{R t} obtained in step ST304 in advance, if less than the threshold value _{T r,} the process proceeds to step ST 306, the m phoneme sets of m phonemic sets notation 11 selected extracted It is sent to the m phoneme set notation memory 12. The extracted m phoneme set notation memory 12 holds the input m phoneme set notation 11. On the other hand, if the recognition rate _Rt is greater than the threshold value _Tr , the m-phoneme-set extracting unit 10 does not send anything to the extracted m-phoneme-set notation memory 12 and returns to step S
Proceed to T307.

At step ST307, the m-phoneme set extraction unit 10 refers to the m-phoneme set table stored in the dialogue speech learning data memory 8, and checks all m-phoneme sets existing in the dialogue speech learning data memory 8. It is determined whether or not the recognition rate _Rt has been calculated. If the recognition has not been completed for all m phoneme sets, the process proceeds to step ST308, where the next m phoneme set table is processed according to the order described in the m phoneme set table.
The phoneme set is selected as a recognition target, and the process returns to step ST303. On the other hand, if the m-phoneme-set extracting unit 10 has finished recognizing all the m-phoneme-sets, the m-phoneme-set extraction procedure ends.

As described above, the m phoneme set extraction unit 10 executes the m phoneme set extraction procedure (steps ST301 to ST3 in FIG. 5).
08), all m phoneme sets whose recognition rate R _t is equal to or less than the threshold value _Tr are extracted, and the m phoneme set notations 1
1 can be stored in the extracted m phoneme set notation memory 12.

Finally, the model learning section 3 learns the dialogue m m phoneme set model for each m phoneme set extracted in the second step using the token stored in the dialogue speech learning data memory 8 in the third step. I do. FIG. 6 is a flowchart showing the learning procedure of the third step. Hereinafter, the details of the learning procedure will be described with reference to FIG.

First, the model learning section 3 first proceeds to step ST4
In step 01, the m phoneme group notation 11 stored in the extracted m phoneme group notation memory 12 is read, and the first m phoneme group is selected as a learning target according to the order stored in the extracted m phoneme group notation memory 12. I do. When the contents of the extracted m phoneme set notation memory 12 are as shown in FIG. 2, for example,
First, the model learning unit 3 is the first m phoneme set (a) a
(U) is selected as a learning target.

Next, the model learning section 3 determines in step ST4
02, a time series 9 of feature vectors of all tokens having the m phoneme set notation that matches the m phoneme set selected in step ST401 or ST406.
Is read from the conversation voice learning data memory 8.

Then, the model learning section 3 determines in step ST
At 403, the dialogue speech m for the m phoneme set selected using, for example, the forward-backward algorithm
Learn phoneme set models.

Thereafter, the model learning section 3 determines in step ST
In 404, the parameters of the model obtained as a result of the learning in step ST403 and the m-phoneme set notation 15 are sent to the dialogue speech m-phoneme set model memory 16.
The conversational voice m phoneme set model memory 16 holds the parameters of the received model and the m phoneme set notation 15.

Next, the model learning section 3 determines in step ST4
At 05, it is determined whether all m phoneme set models have been learned for all m phoneme sets held in the extracted m phoneme set notation memory 12, and learning has not been completed for all m phoneme set models. In this case, the process proceeds to step ST406, where the next m phoneme set is selected as a learning target according to the order described in the extracted m phoneme set notation memory 12, and the process returns to step ST402. On the other hand, if the learning has been completed for all the m phoneme sets, the model learning unit 3 ends the conversational speech m phoneme set model learning procedure.

When the voice pattern model learning method according to the first embodiment is realized by software, the read-aloud m m-phoneme set model is learned and stored in the read-aloud m-phoneme set model memory 14. The first step is to perform the recognition of each token stored in the conversational speech learning data memory 8 using the read-out speech m-phoneme set model stored in the read-out speech m-phoneme set model memory 14, and the recognition rate is low. a second step of extracting m phoneme sets, and a dialog voice m for each of all m phoneme sets extracted in the second step using the token stored in the dialog voice learning data memory 8.
And a third step of learning a phoneme set model, and a computer-readable recording medium storing a voice pattern model learning program for causing a computer to learn a voice pattern model is required.

As described above, according to the speech pattern model learning apparatus and the speech pattern model learning method of the first embodiment, the conversational speech learning data memory 8 holds the read-out speech m-phoneme set model. Recognition of all m phoneme sets is performed, m phoneme sets having a low recognition rate are extracted, and only the extracted m phoneme sets are extracted using the time series of the feature vectors of the tokens held in the conversational speech learning data memory 8. Since the m-phoneme set model of the dialogue speech is learned, it is difficult to recognize the m-phoneme set model of the read-out speech trained with the read-out speech without learning the m-phoneme set model of the dialogue speech for all the m-phoneme sets. This has the effect of efficiently learning a dialogue m m phoneme set model that can recognize dialogue voices. In the first embodiment, m = 3
However, it is also possible to select any integer other than 3 for m, and in that case, the same effect is obtained.

Embodiment 2 The speech pattern model learning device according to the second embodiment of the present invention is similar to the first embodiment.
M phoneme set extraction unit 1 that executes the following improved m phoneme set extraction procedures 1 to 4 in place of m phoneme set extraction procedures 1 to 4
0 is provided. The speech pattern model learning apparatus according to the second embodiment has the same configuration as that according to the first embodiment shown in FIG. 1, and the components other than the m phoneme set extraction unit 10 are the same as those in the first embodiment. 1 performs the same operation as the speech pattern model learning apparatus according to No. 1, and the description of other components will be omitted below. Also, in the second embodiment, a description will be given of m phoneme sets where m = 3.

Next, the operation will be described. (1) Improved m phoneme group extraction procedure 1: m phoneme group extraction unit 10
Are the parameters of all the read m-phoneme set models and the m
The phoneme set notation 13 is read.

(2) Improved m phoneme group extraction procedure 2: Next, m
The phoneme set extraction unit 10 reads the m phoneme set table stored in the dialogue speech learning data memory 8 and sets the first m phoneme set in the dialogue speech learning data as a recognition target according to the description contents of the m phoneme set table. select. When the m phoneme set table is described, for example, as shown in FIG. 24, the m phoneme set extraction unit 10 is the first m phoneme set (a).
a (a) is selected as a recognition target.

(3) Improved m-phoneme group extraction procedure 3: The m-phoneme group extraction unit 10 selects m m-phonemes that match the m-phoneme group selected in the above-mentioned improved m-phoneme group extraction procedure 2 or the following improved m-phoneme group extraction procedure 4. The time series 9 of the feature vectors of all the tokens having the set notation is read from the interactive speech learning data memory 8. Number of read tokens N _t (subscript t is selected m
If the phoneme set name is less than a predetermined threshold N, the m phoneme set extraction unit 10 extracts the m phoneme set notation memory 12
Nothing is sent, and the procedure moves to the following improved m phoneme group extraction procedure 4. On the other hand, if the threshold value N or more that N _t is predetermined, for recognizing as in the first embodiment. That is, for each of the read tokens, the likelihood of all the read m-phoneme set models read out in the improved m-phoneme set extraction procedure 1 is calculated, and the m-phoneme set of the m-phoneme set model showing the highest likelihood is calculated. The set notation is the recognition result of the token. The likelihood calculation uses, for example, a Viterbi algorithm. After obtaining the recognition result for all tokens read, m phoneme set extraction unit 10 calculates the recognition rate R _t in the same manner as the first embodiment by the expression (1). Then, the m phoneme set extraction unit 10 sets the recognition rate R _t to a predetermined threshold T
_r, and if it is less than or equal to the threshold value _Tr , m
The phoneme set notation 11 is sent to the extracted m phoneme set notation memory 12. The extracted m phoneme set notation memory 12 holds the input m phoneme set notation 11.

(4) Improved m-phoneme set extraction procedure 4: The m-phoneme set extraction unit 10 refers to the m-phoneme set table held in the dialogue speech learning data memory 8 and reads all m-phoneme set data present in the dialogue speech learning data. In order to execute the improved m phoneme set extraction procedure 3 for the phoneme set, the next m phoneme set is selected according to the order described in the m phoneme set table, and the improved m phoneme set extraction procedure 3 is repeated. In this way,
When the recognition rates are obtained for all the m phoneme sets existing in the conversational speech learning data, the m phoneme set extraction unit 10 ends the improved m phoneme set extraction procedure.

Next, a method for learning the m-phoneme set model used by the voice pattern model learning apparatus according to the second embodiment will be specifically described. In the speech pattern model learning device according to the second embodiment, the learning procedure of the m phoneme set model is roughly divided into three steps as in the speech pattern model learning device according to the first embodiment.

First, the first step is to store the parameters of the model obtained as a result of learning and learning the m-phoneme set model of the read-aloud speech m and the m-phoneme set notation 13 in the read-aloud speech m-phoneme set model memory 14. Sound m
This is a step of learning a phoneme set model.

In the second step, the m-phoneme set model stored in the conversational speech learning data memory 8 is described by using the m-phoneme set model of the read speech m stored in the m-phoneme set model memory 14. the number N _t of tokens from the m phoneme set is a step of threshold N or more and recognition rate R _t to extract m phonemes sets is equal to or less than the threshold T _r.

Then, the next third step is a step of learning a dialogue m m phoneme set model for each m phoneme set extracted in the second step using the token stored in the dialogue speech learning data memory 8. It is.

Of the first to third steps, the first and third steps have exactly the same procedure as in the first embodiment, and therefore will not be described below, and the second step of extracting m phoneme sets will be described. The procedure will be described in detail. FIG. 7 shows this second
8 is a flowchart showing a procedure for extracting steps. Hereinafter, the extraction procedure will be described in detail with reference to FIG.

The m-phoneme-set extracting unit 10 first executes step S
At T501, the parameters of all the read-out m-phoneme set models and the m-phoneme set notation 13 are read from the read-out m-phoneme set model memory 14.

Next, the m-phoneme set extraction unit 10 executes step S
At T502, the m phoneme set table stored in the conversational speech learning data memory 8 is read, and the m phoneme set described at the head of the m phoneme set table is selected as a recognition target. When the m phoneme set table is described, for example, as shown in FIG. 24, the m phoneme set extraction unit 10 selects the first m phoneme set (a) a (a) as a recognition target.

Then, in step ST503, the m-phoneme set extraction unit 10 interacts with the time series 9 of the feature vectors of all the tokens having the m-phoneme set notation that matches the m-phoneme set selected in step ST502 or ST509. Read from the voice learning data memory 8.

[0095] Thereafter, m phoneme set extraction unit 10, at step ST 504, compared to the number _{N t} (subscript t indicates the name of the m phoneme sets selected) threshold N a predetermined a read token, _{N t} If <N, nothing is sent to the extracted m phoneme set notation memory 12, and the routine goes to step ST508. On the other hand, if N _t > = N, the m phoneme group extraction unit 10 moves to step ST505.

In step ST505, the m-phoneme set extraction unit 10 calculates the likelihood of each of the read tokens with all the read-out speech m-phoneme set models read in step ST503, and determines the highest likelihood. M shown
The m phoneme set notation of the phoneme set model is set as the recognition result of the token. The likelihood calculation uses, for example, a Viterbi algorithm. After obtaining the recognition result for all tokens read, m phoneme set extraction unit 10 calculates the recognition rate R _t in accordance with the equation (1).

Next, the m-phoneme set extraction unit 10 executes step S
In T506, it is compared with a threshold value _{T r} which defines a recognition rate _{R t} determined in advance at step ST505, the threshold T
If it is equal to or smaller than _r , the process proceeds to step ST507, where the m phoneme set notation 11 of the m phoneme set is extracted and the m phoneme set notation memory 12
To send to. The extracted m phoneme set notation memory 12 holds the input m phoneme set notation 11. On the other hand, if the recognition rate _{R t} is larger than the threshold value _{T r,} m phoneme set extraction unit 10 proceeds to step ST 508.

Then, in step ST508, the m-phoneme set extraction unit 10 refers to the m-phoneme set table stored in the dialogue speech learning data memory 8 and retrieves all the m-phoneme sets existing in the dialogue speech learning data memory 8. It is determined whether or not the selection has already been completed. If there is an unselected m phoneme set, the process proceeds to step ST509, and the next m phoneme set is selected as a recognition target in accordance with the order described in the m phoneme set table. Then, the process returns to step ST503. On the other hand, if the m phoneme set extraction unit 10 has already selected all m phoneme sets, the m phoneme set extraction procedure ends.

When the speech pattern model learning method according to the second embodiment is realized by software, the read-out speech m-phoneme set model is learned, and the result obtained by the learning is stored in the read-out speech m-phoneme set model memory 14. A first step of learning a read-aloud m phoneme group model;
The number of tokens N from the m phoneme sets described in the m phoneme set table stored in the conversational speech learning data memory 8 using the read out voice m phoneme set model stored in the read out voice m phoneme set model memory 14. using a second step of extracting m phonemes sets _t is a threshold value N or more and recognition rate R _t is less than the threshold value T _r, the token stored in the interactive voice learning data memory 8 is extracted with the second step And a third step of learning the m-phoneme set model of the dialogue speech with respect to the m-phoneme set described above, and a computer-readable recording medium storing a program for causing a computer to learn a speech pattern model is required.

As described above, the speech pattern model learning apparatus according to the second embodiment executes the above-described improved m phoneme group extraction procedures 1 to 4 (steps ST501 to ST509 in FIG. 7), thereby obtaining the tokens. Number N
_t is extracted m phonemic sets representation 11 of every m phonemes sets the threshold value N or more and recognition rate R _t is equal to or less than the threshold value T _r, extracting m phonemes all m phoneme sets of m phonemic sets notation 11 extracted It is stored in the set notation memory 12. Therefore, the second embodiment
Speech pattern model learning device according to the extracted m number N _t is the threshold value N or more m tokens in the training of phoneme pairs model
Since learning model only phoneme set, reading of the m phoneme sets of low recognition rate interactive voice sound m phoneme sets model, a statistically of low reliability model learning number N _t is less than the threshold value N of tokens This has the effect of avoiding and efficiently learning only models with high statistical reliability. In the second embodiment, m is described as 3; however, m may be any integer other than 3, and the same effect is obtained in that case.

Embodiment 3 FIG. FIG. 8 is a block diagram showing a configuration of a speech pattern model learning device according to Embodiment 3 of the present invention. In the drawing, reference numeral 30 denotes a m-phoneme group extraction unit (m-phoneme group extraction unit) which learns a m-phoneme group extraction model by using a text-to-speech voice for each m-phoneme group stored in the speech-to-speech learning data memory 6. Means) For each m phoneme set extracted by 10,
The conversational speech m-phoneme set model is learned using the conversational speech learning data stored in the conversational speech learning data memory 80, and further, the n phonemes are set using the read-out speech m-phoneme set model and the dialogue m-phoneme set model. For each of the n phoneme sets extracted from the conversational speech learning data by the set extraction unit (n phoneme group extraction means) 17, a model learning unit (conversational speech learning) for learning a conversational speech n phoneme group model using the conversational speech learning data. m phoneme set model learning means, dialogue speech n phoneme set model learning means), 7 is a time series of the feature vector of the read speech included in the read speech learning data memory
Is a time series of the feature vector of the dialogue speech included in the dialogue speech learning data memory 80, 11 is an m phoneme set notation of the m phoneme set extracted by the m phoneme set extractor 10, and 12 is an extracted m
Phoneme set notation memory, 13 is a parameter of m-phoneme set model of read-out speech m and m-phoneme set notation, 14 is a read-out speech m-phoneme set model memory, 15 is a parameter of m-phoneme set model of dialogue speech and m-phoneme set notation, 16 is a dialogue Speech m phoneme group model memory, 18 n phoneme group notation of n phoneme group extracted by n phoneme group extractor 17, 19 extracted n phoneme group notation memory, 20 parameters of n phoneme model of conversational speech and n phonemes A set notation, 21 is a conversational speech n phoneme set model memory. In FIG. 8, the same reference numerals as those shown in FIG. 1 denote the same or corresponding components as those of the speech pattern model learning apparatus according to the first embodiment. In the following, description will be made assuming that m = 3 and n = 5. The speech pattern model used by the speech pattern model learning device according to the third embodiment is a continuous distribution type HMM as in the first embodiment.

The conversational speech learning data memory 80 includes n data existing in the conversational speech learning data in addition to the data held in the conversational speech learning data memory 8 according to the first embodiment.
It holds an n phoneme set table describing the types of phoneme sets.
Here, n is an integer satisfying n> m, and the n phoneme set is a set of n phonemes taking into account differences in phonemes in a range longer than the m phoneme set. For example, when n = 5, / saQp
oro (Sapporo) // p / is n (= 5) phoneme set (a
Q) It becomes p (or). This (aQ) p (or)
Will be referred to as pentaphone notation hereinafter. 5
FIG. 9 shows an example of the phoneme set table. In addition, each token (each token is assigned a token number) held by the dialogue speech learning data memory 80 includes a phoneme name of the token, a phoneme name preceding the phoneme, a preceding phoneme name, and a succeeding phoneme together with the three-phoneme set notation. A five-phoneme set notation in which a name and a subsequent phoneme name are described later is given. FIG. 10 shows an example of a five-phoneme set notation provided together with the three-phoneme set notation.

Next, the operation will be described. The speech pattern model learning apparatus according to the third embodiment includes five procedures divided as follows: (1) a procedure for learning a m-phoneme set model of a read-aloud speech, and (2) a conversational speech learning data memory 80.
, A procedure for extracting a m-phoneme set having a low recognition rate, (3) a learning procedure for a m-phoneme set model of the dialogue speech for the extracted m-phoneme set, and (4) an n having a low recognition rate held in the dialogue learning data memory 80 Model learning is performed by sequentially executing a phoneme group extraction procedure and (5) a learning procedure of the extracted dialogue speech n phoneme group model.

First, the procedure for learning the m-phoneme set model of the read-out speech will be described. The voice pattern model learning device connects an input terminal A of the model learning unit 30 shown in FIG. 8 to a terminal B1 connected to the read-out voice learning data memory 6, and receives data in the read-out voice learning data memory 6. Set to. Further, the speech pattern model learning device connects the output terminal C of the model learning unit 30 to the terminal D1 connected to the read-out speech m phoneme set model memory 14. First, the speech pattern model learning device learns the read-aloud m-phoneme set model in this connection state. The model learning unit 30 of the speech pattern model learning apparatus according to the third embodiment learns the read-out speech m phoneme set model according to the read-out speech model learning procedures 1 to 3 described in the first embodiment, and reads out the read-out speech m. The model parameters obtained as a result of learning and the m phoneme group notation are stored in the phoneme group model memory 14. Speech m for all m phoneme sets existing in the speech speech learning data memory 6
When the learning of the phoneme set model is completed, the model learning unit 30
Ends the learning procedure of the m-phoneme set model of the read-aloud voice.

Next, in the speech pattern model learning apparatus, the m phoneme set extraction unit 10 extracts m phoneme sets having a low recognition rate held in the conversational speech learning data memory 80. The m phoneme set extraction unit 10 executes the m phoneme set extraction procedure with a low recognition rate in accordance with the m phoneme set extraction procedures 1 to 4 described in the first embodiment, and executes m phoneme set extraction for all m phoneme sets. The set notation is stored in the extracted m-phoneme set notation memory 12.

Next, the model learning section 30 learns the m-phoneme set model of the dialogue voice. Before starting the learning, the speech pattern model learning apparatus connects the input terminal A of the model learning unit 30 to the output terminal B2 of the conversational speech learning data memory 80,
Further, another input terminal E of the model learning unit 30 is connected to the output terminal F1 of the extracted m phoneme set notation memory 12. Further, the voice pattern model learning device includes a model learning unit 30.
Is connected to the input terminal D2 of the conversational voice m phoneme group model memory 16. In this connection state, the model learning unit 30 learns the m-phoneme set model of the dialogue voice.

The model learning unit 30 learns the m-phoneme group model of the dialogue speech in accordance with the learning procedures 1 to 3 of the m-phoneme group model extracted in the first embodiment, and stores the learning in the dialogue m-phoneme group model memory 16. The resulting model parameters and their m phoneme set notation are stored. Then, the model learning unit 30 ends the learning procedure of the dialogue m m phoneme set model when the learning of the dialogue m m phoneme set model is completed for all the m phoneme sets stored in the extracted m phoneme set notation memory 12.

Next, in the speech pattern model learning apparatus, the n phoneme set extraction unit 17 extracts an n phoneme set having a low recognition rate held in the conversational speech learning data memory 80. The procedure for extracting the n phoneme sets is as follows.

(1) Procedure for extracting n phoneme sets 1: The n phoneme set extraction unit 17 reads the parameters of all read m speech phoneme models and the m phoneme set notation 13 from the read speech m phoneme set model memory 14. The n phoneme set extraction unit 17
Further, the parameters of all the m-phoneme set models of the dialogue speech and the m-phoneme set notation 15 are read from the dialogue m-phoneme set model memory 16.

(2) N phoneme set extraction extraction procedure 2: Next, the n phoneme set extraction unit 17 refers to the n phoneme set table held in the conversational speech learning data memory 80, and adds the description contents of this phoneme set table. Therefore, the first n phoneme sets are selected as recognition targets. When n = 5 and the n phoneme set table is described as shown in FIG. 9, for example, the n phoneme set extraction unit 17 first selects the first n phoneme set (ka) a (ai) as a recognition target. I do.

(3) N phoneme set extraction procedure 3: The n phoneme set extraction unit 17 extracts the n phoneme set notation that matches the n phoneme set selected in the above n phoneme set extraction procedure 3 or the following n phoneme set extraction procedure 4. Time series 9 of feature vectors of all tokens
Is read from the dialogue speech learning data memory 80, and for each of the read tokens, the likelihood of all the read-out m-phoneme set models and all the dialogue m-phoneme set models read in the n-phoneme set extraction procedure 1 is calculated. The m phoneme set notation of the m phoneme set model showing the highest likelihood is used as the token recognition result. The likelihood calculation uses, for example, a Viterbi algorithm. After obtaining the recognition result for all tokens read, n phoneme set extraction unit 17 calculates the recognition rate R _q in accordance with the following equation (2).

R _q = C _q / N _q * 100.0 (2)

Here, the subscript q indicates the type of the selected n phoneme set, _Nq is the number of tokens whose n phoneme set type is q in the n phoneme set notation, and C _q is positively recognized among them. The number of tokens. Here, correct recognition means m of the token
M of phoneme set model with phoneme set notation showing highest likelihood
The case where it matches the phoneme set notation is regarded as correct recognition. For example, a token whose n phoneme group notation (n = 5) is (ka) a (ai) has the highest likelihood because the m phoneme group notation (m = 3) is (a) a (a). If the m phoneme set notation of the m phoneme set model is (a) a (a), it is determined that the recognition is correct.

[0114] n phoneme set extraction unit 17 compares the threshold _{T q} determined in advance of the recognition rate _{R q,} equal to or less than the threshold value _{T q,}
The n phoneme set notation 18 of the n phoneme set is sent to the extracted n phoneme set notation memory 19. The extracted n phoneme set notation memory 19 holds the inputted n phoneme set notation 18.

(4) n Phoneme Group Extraction Procedure 4: The n phoneme group extraction unit 17 stores n
Referring to the phoneme set table, the dialogue speech learning data memory 8
In order to execute the above n phoneme group extraction procedure 3 for all n phoneme groups existing in 0, the next n phoneme group is selected according to the order described in the n phoneme group table,
The above n phoneme group extraction procedure 3 is repeated.

Thus, the n-phoneme-set extracting unit 17 ends the procedure for extracting the n-phoneme-set. n phoneme set extraction unit 17
Extracts the n phoneme set notations 18 of all n phoneme sets whose recognition rate R _q is equal to or less than the threshold value T _q by executing the above n phoneme set extraction procedures 1 to 4, and extracts the n phoneme set notation memory 1
9 can be stored.

Next, the model learning unit 30 learns a conversational speech n phoneme group model for each n phoneme group extracted as described above. Before starting the learning, the speech pattern model learning apparatus connects the input terminal A of the model learning unit 30 to the output terminal B2 of the conversational speech learning data memory 80, and connects another input terminal E of the model learning unit 30 to the input terminal E. It is connected to the output terminal F2 of the extracted n phoneme set notation memory 19. Further, the speech pattern model learning device connects the output terminal C of the model learning unit 30 to the input terminal D of the conversational speech n phoneme set model memory 21.
Connect to 3. In this connection state, the model learning unit 30 learns the conversational speech n-phoneme set model. The learning procedure is shown below.

(1) Extracted n phoneme set model learning procedure 1: The model learning unit 30 firstly outputs the extracted n phoneme set notation memory 19
Is read, and the first n phoneme set is selected as a learning target in accordance with the order held in the extracted n phoneme set notation memory 19. Extraction n
If the contents of the phoneme set notation memory 19 are, for example, as shown in FIG. 11, the model learning unit 30 is the first n phoneme sets (k
a) Select a (ai) as a learning target.

(2) Extracted n phoneme set model learning procedure 2: Next, the model learning unit 30 selects the n phoneme set selected in the extracted n phoneme set model learning procedure 1 or the extracted n phoneme set model learning procedure 3 described below. The time series 9 of the feature vectors of all tokens having the matching n phoneme set notation is read from the dialogue speech learning data memory 80, and a model for the selected n phoneme set is learned using, for example, a forward / backward algorithm. Then, the model learning unit 30 sends the parameters of the model obtained as a result of the learning and the n phoneme group notation to the n phoneme group model memory 21 of the dialogue speech. The conversational speech n phoneme set model memory 21 holds the parameters of the received model and the n phoneme set notation.

(3) Extracted n phoneme set model learning procedure 3: The model learning unit 30 executes the extracted n phoneme set model learning procedure 2 for all the n phoneme sets held in the extracted n phoneme set notation memory 19. To do so, the next n phoneme sets are selected according to the order stored in the extracted n phoneme set notation memory 19, and the extracted n phoneme set model learning procedure 2 is repeated. Thus, the model learning unit 30 ends the learning of the extracted n phoneme set model.

Next, a method of learning the m phoneme set model and the n phoneme set model used by the speech pattern model learning apparatus according to the third embodiment will be specifically described. FIG. 12 is a flowchart showing a procedure of a voice pattern model learning method according to Embodiment 3 of the present invention. As shown in FIG. 12, the read-out speech m phoneme set model and the dialogue speech m in the speech pattern model learning method according to the third embodiment.
The learning procedure of the phoneme set model and the dialogue n-phoneme set model is mainly divided into five steps.

That is, the first step (step ST601 in FIG. 12) is to learn the m-phoneme set model of the read-aloud voice and to store the model parameters obtained as a result of the learning and the m-phoneme set notation 13 in the read-aloud m-phoneme set model memory 14. This is the learning procedure for the m-phoneme set model of the reading voice to be stored.

The next second step (step ST6 in FIG. 2)
02) recognizes each token stored in the interactive speech learning data memory 80 using the read-aloud m-phoneme set model stored in the read-aloud m-phoneme set model memory 14, and performs m-phonemes with a low recognition rate. This is a procedure for extracting a set.

The next third step (step ST in FIG. 12)
603) learns the dialogue m m phoneme set model for each m phoneme set extracted in the second step using the token stored in the dialogue speech learning data memory 80, and obtains the parameters of the model obtained as a result of learning and m phoneme set notation 1
5 is a dialogue m-phoneme set model learning procedure for storing 5 in the dialogue m-phoneme set model memory 16.

The next fourth step (step ST in FIG. 12)
Reference numeral 604) denotes a reading voice m phoneme group model stored in the reading voice m phoneme group model memory 14 and a dialog voice m stored in the dialog voice m phoneme group model memory 16.
Interactive speech learning data memory 80 using phoneme set model
Is a procedure for recognizing each of the tokens stored in the, and extracting an n-phoneme set having a low recognition rate.

The next fifth step (step ST in FIG. 12)
605) is the n extracted in the fourth step using the token stored in the conversational speech learning data memory 80.
This is a dialog speech n phoneme group model learning procedure in which a dialog speech n phoneme group model for a phoneme group is learned, and the model parameters obtained as a result of the learning and the n phoneme group notation 20 are stored in the dialog speech n phoneme group model memory 21.

Of the first to fifth steps, the first, second, and third steps are exactly the same as those in the first embodiment, and thus description thereof will be omitted, and the fourth and fifth steps will be described below. Will be described. FIG. 13 is a flowchart showing the details of the fourth step. Hereinafter, the procedure of extracting the n phoneme sets having a low recognition rate, which is the fourth step, will be described in detail with reference to FIG.

First, the n phoneme group extraction unit 17 determines in step S
At T701, the parameters of all the read m-phoneme group models and the m-phoneme group notation 13 are read from the read m-phoneme group model memory 14. In step ST702, the n-phoneme set extraction unit 17 reads the parameters of all m-phoneme set models of the dialogue speech and the m-phoneme set notation 15 from the dialogue m-phoneme set model memory 16.

Next, the n-phoneme set extraction unit 17 determines in step S
At T703, the n phoneme set table held in the conversational speech learning data memory 80 is referred to, and the first n phoneme set is selected as a recognition target according to the description contents of the n phoneme set table. When the n-phoneme set table is described, for example, as shown in FIG. 9, the n-phoneme set extraction unit 17 first selects (ka) a (ai), which is the first n-phoneme set, as a recognition target.

Then, in step ST704, the n-phoneme set extraction unit 17 interacts with the time series 9 of the feature vectors of all tokens having the n-phoneme set notation that matches the n-phoneme set selected in step ST703 or ST709. It is read from the voice learning data memory 80.

Thereafter, in step ST705, the n-phoneme set extracting unit 17 calculates, for each token read,
The likelihood of all the read-aloud m m-phoneme set models and all the dialogue m-phoneme set models read in the above-mentioned steps ST701 and ST702 is calculated, and the m-phoneme of the m-phoneme set model showing the highest likelihood is calculated. The set notation is the recognition result of the token. The likelihood calculation uses, for example, a Viterbi algorithm. After obtaining the recognition results for all the read tokens, the n-phoneme group extraction unit 17 calculates the recognition rate _Rq by the above equation (2).

Next, the n-phoneme set extraction unit 17 determines in step S
In T706, the threshold T determined in advance of the recognition rate _{R q}
compared to _q, equal to or less than the threshold value _{T q,} step ST70
Then, the process proceeds to step S7, where the n phoneme group notation 18 of the n phoneme group is sent to the extracted n phoneme group notation memory 19. The extracted n phoneme set notation memory 19 holds the inputted n phoneme set notation 18. On the other hand, if the recognition rate _Rq is larger than the threshold _Tq , n
The phoneme set extraction unit 17 proceeds to step ST708.

In step ST708, the n-phoneme set extraction unit 17 refers to the n-phoneme set table held in the dialogue speech learning data memory 80, and calculates the recognition rate for all n-phoneme sets present in the dialogue speech learning data. It is determined whether or not it has been performed. If the recognition has not been completed for all n phoneme sets, the n phoneme set extraction unit 17 proceeds to step ST
In 709, the next n phoneme sets are selected as recognition targets in the order described in the n phoneme set table, and the process returns to step ST704. On the other hand, if the recognition has been completed for all the n phoneme sets, the n phoneme set extraction unit 17 ends the procedure for extracting the n phoneme sets.

As described above, the n phoneme set extraction unit 17 performs the n phoneme set extraction procedure (step ST70 in FIG. 13).
1 to ST709), the n phoneme set notations 18 of all n phoneme sets whose recognition rate R _q is equal to or less than the threshold value T _q can be extracted and stored in the extracted n phoneme set notation memory 19. it can.

Next, the model learning section 30 executes the fifth step, that is, the learning procedure of the dialogue n phoneme group model. Before performing the learning, the speech pattern model learning apparatus connects the input terminal A of the model learning unit 30 to the output terminal B2 of the conversational speech learning data, and extracts another input terminal E of the model learning unit 30. It is connected to the output terminal F2 of the phoneme set notation memory 19. Further, the speech pattern model learning device connects the output terminal C of the model learning unit 30 to the input terminal D3 of the conversational speech n phoneme set model memory 21. In this connection state, the model learning unit 30 learns the conversational speech n-phoneme set model.

FIG. 14 is a flow chart showing the details of the learning procedure of the n-phoneme group model of the dialogue speech.
The details of the learning procedure will be described with reference to FIG. First, in step ST801, the model learning unit 30 reads the n phoneme group notation 18 stored in the extracted n phoneme group notation memory 19, and reads the extracted n phoneme group notation memory 19
Are selected as learning targets in accordance with the order held in. If the contents of the extracted n-phoneme set notation memory 19 are, for example, as shown in FIG.
0 selects the first n phoneme set (ka) a (ai) as a learning target.

Next, the model learning section 30 executes step ST
In step 802, the time series 9 of the feature vectors of all tokens having the n phoneme set notation that matches the n phoneme set selected in step ST 801 or ST 806 is read from the interactive speech learning data memory 80. Then, in step ST803, the model learning unit 30 learns a model for the selected n phoneme set using, for example, a forward / backward algorithm.

Thereafter, the model learning section 30 determines in step S
At T804, the parameters of the model obtained as a result of the learning and the n phoneme group notation are sent to the dialogue voice n phoneme group model memory 21. Dialogue voice n phoneme set model memory 21
Holds the received model parameters and n phoneme set notation.

Next, the model learning section 30 determines in step ST
In 805, it is determined whether or not the conversational speech n phoneme set model has been learned for all n phoneme sets held in the extracted n phoneme set notation memory 19, and the learning has not been completed for all n phoneme set models. In the model learning unit 30,
In step ST806, the next n phoneme sets are selected as learning targets in the order described in the extracted n phoneme set notation memory 19, and the process returns to step ST802. On the other hand, if the learning has been completed for all the n phoneme sets, the model learning unit 30 ends the fifth step, which is the learning procedure of the dialogue n phoneme set model.

When the speech pattern model learning method according to the third embodiment is realized by software, a read-out speech m-phoneme set model memory is learned, and the result of the learning is stored in the read-out speech m-phoneme set model memory 14. A first procedure for learning the m phoneme set model, and recognition of each token stored in the conversational speech learning data memory 80 using the read out speech m phoneme set model stored in the read out speech m phoneme set model memory 14 And a second step of extracting m phoneme sets having a low recognition rate, and a dialogue speech of each m phoneme set extracted in the second procedure using the token stored in the dialogue speech learning data memory 80. A third procedure for learning the m phoneme set model, and a read speech m phoneme set model stored in the read speech m phoneme set model memory 14 Interactive voice m phoneme sets model memory 16
A fourth procedure for recognizing each token stored in the conversational speech learning data memory 80 using the conversational speech m phoneme set model stored in the dialogue box and extracting an n phoneme set with a low recognition rate; A fifth step of learning a dialogue n-phoneme set model for each n-phoneme set extracted using the token stored in the learning data memory 80. A computer-readable recording medium that records the recorded data is required.

As described above, in the third embodiment,
Pattern model learning device and voice pattern model
In the Dell learning method, the n phoneme set extraction procedure (step S
Steps ST801 to ST806)
Therefore, the recognition rate R_qIs the threshold T _qAll n phonemes that are
Dialogue phonetics for each of the extracted n phoneme groups
N phoneme group of dialogue speech using tokens held by learning data
Since the model is learned, the utterance speed is high like dialogue voice.
Pair with m-phoneme set model, which is read out quickly and vaguely
Sufficient recognition performance cannot be obtained with m-phoneme set model
Efficiently construct dialogue n phoneme set models for each n phoneme set
It has the effect of being able to learn. Note that this implementation
In the third embodiment, m = 3 and n = 5 have been described.
May select any set of integers such that m <n.
The same effect is obtained in such a case.

Embodiment 4 The speech pattern model learning device according to the fourth embodiment of the present invention is similar to the third embodiment.
N phoneme set extraction unit 1 that executes the following improved n phoneme set extraction procedures 1 to 4 instead of n phoneme set extraction procedures 1 to 4
7 is provided. The speech pattern model learning apparatus according to the fourth embodiment is different from the third embodiment shown in FIG.
And the components other than the n-phoneme set extraction unit 17 operate in the same manner as the speech pattern model learning device according to the third embodiment, so that the other components will be described below. Omitted. Also in the fourth embodiment, a description will be given of m phoneme sets of m = 3 and n phoneme sets of n = 5.

Next, the operation will be described. (1) Improved n phoneme group extraction procedure 1: n phoneme group extraction unit 17
Reads all the parameters of the m-phoneme set model of the read-aloud speech and the m-phoneme set notation 13 from the read-aloud m-phoneme set model memory 14. The n-phoneme group extraction unit 17 further stores the parameters of all the m-phoneme group models of the dialogue speech and the m-phoneme group notation 1
Read 5

(2) Improved n phoneme group extraction procedure 2: Next, n
The phoneme set extraction unit 17 reads the n phoneme set table stored in the dialogue speech learning data memory 80, and recognizes the first n phoneme set in the dialogue speech learning data according to the description contents of the n phoneme set table. select. If the n-phoneme set table is described, for example, as shown in FIG. 9, the n-phoneme set extraction unit 17 is the first n-phoneme set (k
a) Select a (ai) as a recognition target.

(3) Improved n phoneme set extraction procedure 3: The n phoneme set extraction unit 17 selects the n phonemes that match the n phoneme set selected in the improved n phoneme set extraction procedure 2 or the improved n phoneme set extraction procedure 4 described below. The time series 9 of the feature vectors of all tokens having the set notation is read from the dialogue speech learning data memory 80. If the number of read tokens N _q (subscript q indicates the name of the selected n phoneme set) is less than a predetermined threshold N, nothing is sent to the extracted n phoneme set notation memory 19 and the next The procedure moves to the improved n phoneme set extraction procedure 4. On the other hand, if the threshold value N or more N _q is predetermined, n phoneme set extraction unit 17 performs the recognition as in the third embodiment. That is, the n-phoneme set extraction unit 17 calculates the likelihood of each of the read tokens with all the read-aloud m-phoneme set models and all the dialogue m-phoneme set models read in the improved n-phoneme set extraction procedure 1. Then, the m phoneme set notation of the m phoneme set model showing the highest likelihood is set as the token recognition result. The likelihood calculation uses, for example, a Viterbi algorithm. After obtaining the recognition results for all the read tokens, the n-phoneme set extraction unit 17 calculates the recognition rate _Rq by the above equation (2). Then, the n phoneme set extraction unit 17
Compares the recognition rate _Rq with a predetermined threshold value _Tq, and if it is equal to or smaller than the threshold value _Tq , sends the n-phoneme set notation of the n-phoneme set to the extracted n-phoneme set notation memory 19. The extracted n phoneme group notation memory 19 holds the input n phoneme group notation.

(4) Improved n-phoneme set extraction procedure 4: The n-phoneme set extraction unit 17 refers to the n-phoneme set table held in the dialogue speech learning data memory 80, and retrieves all the n-phoneme set tables existing in the dialogue speech learning data memory 80. In order to execute the improved n phoneme set extraction procedure 3 for the n phoneme sets, the next n phoneme set is selected as a recognition target according to the order described in the n phoneme set table, and the improved n phoneme set extraction procedure is performed. Repeat 3. When the recognition rates are obtained for all the n phoneme sets existing in the conversational speech learning data in this way, the n phoneme set extraction unit 17 ends the improved n phoneme set extraction procedure.

Next, a method for learning the m phoneme set model and the n phoneme set model used by the speech pattern model learning apparatus according to the fourth embodiment will be specifically described. In the speech pattern model learning device according to the fourth embodiment, the learning procedure of the m phoneme set model and the n phoneme set model is roughly divided into five steps, similarly to the speech pattern model learning device according to the third embodiment.

First, the first step is to read out the m-phoneme set model of the read-aloud speech m and store the m-phoneme set notation 13 in the read-aloud m-phoneme set model memory 14. This is the step of learning the m phoneme set model.

In the next second step, each token stored in the conversational speech learning data memory 80 is recognized using the read-out speech m-phoneme set model stored in the read-out speech m-phoneme set model memory 14. , Extracting m-phoneme sets with a low recognition rate.

The next third step is to use the token stored in the conversational speech learning data memory 80 for the second step.
This is a step of learning an m-phoneme set dialogue speech model for the m-phoneme set extracted in the step.

In the next fourth step, the speech m m phoneme set model stored in the speech m m phoneme set model memory 14 and the dialog m m phoneme set model stored in the dialog m m phoneme set model memory 16 are stored. The n phonemes for which the number of tokens N _q is equal to or greater than the threshold N and the recognition rate R _q is equal to or less than the threshold T _q from among the n phoneme sets described in the n phoneme set table held by the dialogue speech learning data memory 80 This is the step of extracting sets.

In the next fifth step, the fourth step is performed by using the token stored in the conversational speech learning data memory 80.
This is a step of learning a dialogue speech n phoneme set model for each n phoneme set extracted in the step.

[0153] Of the first to fifth steps, the first, second, third and fifth steps are exactly the same as those in the third embodiment, so that the description thereof will be omitted below and the fourth step will be described. The extraction procedure of a certain n phoneme set will be described in detail. FIG. 15 is a flowchart showing the extraction procedure of the fourth step. Hereinafter, the extraction procedure will be described in detail with reference to FIG.

First, the n phoneme group extraction unit 17 first executes step S
At T901, the parameters of all the read-out m-phoneme set models and the m-phoneme set notation 13 are read from the read-out m-phoneme set model memory 14. Subsequently, in step ST902, the n-phoneme set extraction unit 17 reads the parameters of all the m-phoneme set models of the dialogue speech and the m-phoneme set notation 15 from the dialogue m-phoneme set model memory 16.

Next, the n-phoneme set extraction unit 17 determines in step S
At T903, the n phoneme set table held by the conversational speech learning data memory 80 is read, and the first n phoneme set is selected as a recognition target according to the description contents of the n phoneme set table. If the n phoneme set table is described, for example, as shown in FIG.
The phoneme set (ka) a (ai) is selected as a recognition target.

Next, the n-phoneme set extraction unit 17 determines in step S
In T904, n matching the n phoneme set selected in step ST903 or ST910 described above.
The time series 9 of the feature vectors of all the tokens having the phoneme set notation is read from the conversational speech learning data memory 80.

[0157] Then, n phoneme set extraction unit 17, at step ST 905, is compared with a threshold value N which defines the number _{N q} of the read tokens (subscript q represents the name of the n phonemes sets selected) beforehand, _{N q} If <N, nothing is sent to the extracted n phoneme set notation memory 19, and the routine goes to step ST909. On the other hand, if N _q > = N, the n phoneme group extraction unit 17 proceeds to step ST906.

Next, the n-phoneme set extraction unit 17 determines in step S
At T906, the likelihood of all the read-in speech m-phoneme set models and all the dialogue m-phoneme set models read in steps ST902 and ST903 is calculated for each of the read tokens, and the highest likelihood is indicated. The m phoneme set notation of the m phoneme set model is used as the recognition result of the token. The likelihood calculation uses, for example, a Viterbi algorithm. After obtaining the recognition results for all the read tokens, the n-phoneme group extraction unit 17 calculates the recognition rate _Rq by the above equation (2).

Next, the n-phoneme set extraction unit 17 determines in step S
In T907, it is compared with a threshold value _{T q} that defines a recognition rate _{R q} previously determined in step ST 906, the threshold value T
If not more than _q , the process proceeds to step ST908, where the n phoneme set notation 18 of the n phoneme set is extracted.
To send to. The extracted n phoneme set notation memory 19 holds the inputted n phoneme set notation 18. On the other hand, if the recognition rate R _q is larger than the threshold value T _q , the n-phoneme set extraction unit 17 proceeds to step ST909.

[0160] In step ST909, the n-phoneme set extraction unit 17 refers to the n-phoneme set table stored in the conversational speech learning data memory 80, and retrieves all n-phoneme sets existing in the conversational speech learning data memory 80. It is determined whether or not the selection has been completed. If there is an unselected n-phoneme set, the process proceeds to step ST910, and the next n-phoneme set is selected as a recognition target according to the order described in the n-phoneme set table. Then, the process returns to step ST904. On the other hand, n
The phoneme set extraction unit 17 ends the procedure for extracting the n phoneme sets if all the n phoneme sets have already been selected.

When the voice pattern model learning method according to the fourth embodiment is to be realized by software, the read-aloud m m-phoneme set model is learned and the result obtained by the learning is stored in the read-aloud m-phoneme set model memory 14. A first step of learning a read-aloud m m-phoneme set model, and using the read-aloud m-phoneme set model stored in the read-aloud m-phoneme set model memory 14 to store each data stored in the interactive voice learning data memory 80. A second step of recognizing tokens and extracting m phoneme sets having a low recognition rate; and a dialogue for each m phoneme set extracted in the second step using the token stored in the dialogue speech learning data memory 80. 3rd learning of phonetic m phoneme set model
Steps and the number N _q of tokens from among n phoneme sets described in n phoneme sets table held spoken dialogue learning data memory 80 is the threshold N or more and recognition rate R _q is a threshold value T _q
A fourth step of extracting the following n phoneme sets and a step of learning a dialogue n phoneme set model for each of the n phoneme sets extracted using the token stored in the dialogue learning data memory 80 are included. In addition, a computer-readable recording medium that records a program for causing a computer to learn an audio pattern model is required.

As described above, the speech pattern model learning apparatus according to the fourth embodiment uses the improved n phoneme set extraction procedure (step ST901 to step ST901 in FIG. 15).
910), n phoneme set notations of all n phoneme sets whose number of tokens N _q is equal to or greater than the threshold N and whose recognition rate R _q is equal to or less than the threshold T _q are extracted, and all extracted n
Extract n phoneme set notation 18 of phoneme set n phoneme set notation memory 1
9 is stored. Therefore, the speech pattern model learning apparatus according to the fourth embodiment learns only n phoneme sets whose number of tokens _Nq is equal to or greater than the threshold N in learning the extracted n phoneme set models. In the dialogue m phoneme set model, among the n phoneme sets of the dialogue speech having a low recognition rate, the number of tokens _Nq is less than the threshold N, and learning of a model having a statistically low reliability is avoided. This has the effect that only high models can be efficiently learned. Although the fourth embodiment has been described assuming that m = 3 and n = 5, m and n may be any set of integers such that m <n, and the same effect is obtained in that case. .

Embodiment 5 FIG. FIG. 16 is a block diagram showing a configuration of a speech recognition apparatus according to Embodiment 5 of the present invention. In the figure, reference numeral 14 denotes a read speech m phoneme set model memory, 16 denotes a dialogue speech m phoneme set model memory, 21 denotes a dialogue speech n phoneme set model memory, 22 denotes a speech signal input terminal, and 23 denotes a speech signal input terminal 22. The input speech signal, 24 is an acoustic analysis unit that calculates a time series of acoustic feature vectors of the audio signal 23, 25 is a time series of feature vectors output from the acoustic analysis unit 24, and 26 is a phoneme notation of the vocabulary to be recognized. A recognition target vocabulary memory to be stored; 27, a phoneme set notation of the recognition target vocabulary; 28, a read-out speech m-phoneme set model, a conversational speech m-phoneme set model, which is learned by the speech pattern model learning apparatus according to the third or fourth embodiment. A speech pattern model for the vocabulary to be recognized (ie, a vocabulary model to be recognized) by connecting the dialogue n-phoneme set models in parallel A recognition target vocabulary model creation unit (recognition target vocabulary model creation means), 29 is a recognition target vocabulary model parameter and phoneme set notation, 31 is a recognition target vocabulary model memory, and 32 is a recognition target vocabulary model creation unit 28. A recognition unit (recognition means) 33 for recognizing the input voice using the voice pattern model for the recognition target vocabulary is a recognition result. In FIG. 16, the same reference numerals as those shown in FIG. 8 denote the same or corresponding components as those of the speech pattern model learning apparatus according to the third embodiment.

Read-out speech m phoneme group model memory 14
Holds the parameters of all the read-aloud m-phoneme set models created by the speech pattern model learning apparatus according to Embodiment 3 or 4, and the m-phoneme set notation. Further, the dialogue voice m phoneme group model memory 16
Holds the parameters of all the dialogue m-phoneme set models created by the speech pattern model learning apparatus according to the third or fourth embodiment and the m-phoneme set notation. Further, the conversational speech n phoneme set model memory 21 holds the parameters of all the dialogue speech n phoneme set models created by the speech pattern model learning device according to the third or fourth embodiment and the n phoneme set notation. . In the following, description will be made assuming that m = 3 and n = 5. In the following, the dialogue m m phoneme set model memory 16 holds the parameters of all the dialogue m m phoneme set models created by the speech pattern model learning apparatus according to the fourth embodiment and the m phoneme set notation. It is assumed that the dialogue speech n phoneme set model memory 21 holds the parameters of all the dialogue speech n phoneme set models created by the speech pattern model learning device according to Embodiment 4 and the n phoneme set notation. I do.

Next, the operation will be described. The speech recognition apparatus according to the fifth embodiment creates a recognition target vocabulary model before performing recognition, and stores the created recognition target vocabulary model in the recognition target vocabulary model memory 31.

First, a method of creating a recognition target vocabulary model used by the speech recognition apparatus according to the fifth embodiment will be described. The recognition target vocabulary model creating unit 28 creates a recognition target vocabulary model stored in the recognition target vocabulary memory 26. The recognition target vocabulary memory 26 also describes the phoneme notation of the vocabulary to be recognized. FIG. 17 shows an example of the contents of the vocabulary memory 26 to be recognized. In this example, the utterance of the user assuming a hotel reservation is to be recognized, and the vocabulary of vocabulary number 1 is "Please make a reservation", vocabulary number 2 is "Are you free?""
It is. The recognition target vocabulary model creation unit 28 creates a recognition target vocabulary model as follows.

(1) Recognition target vocabulary model creation procedure 1: The recognition target vocabulary model creation unit 28
6. The recognition target vocabulary to be model-created is selected in the order of the vocabulary numbers described in No. 6, and the phoneme description 27 of the recognition target vocabulary is read. For example, if the contents of the recognition target vocabulary memory 26 are as shown in FIG. 17, the recognition target vocabulary model creating unit 28 first has the vocabulary number 1 / yoyakuoneg.
Select aisimasu /.

(2) Recognition-target vocabulary model creation procedure 2: Next, the recognition-target vocabulary model creation unit 28 reads out the read-out speech m from the phoneme set model memory 14 according to the phoneme notation 27 of the selected recognition-target vocabulary. The parameters 13 of the phoneme set model are read, the read-out speech m phoneme set models are connected in series, and a series connection model is created for the selected vocabulary to be recognized. For example, phoneme notation is / yoya
In the case of kuonegaishimasu /, since m = 3, / (#) y (o) /, / (y) o (y) /, /
(O) y (a) /, / (y) a (k) /, / (a) k
(U) /, / (k) u (o) /, / (u) o (n) /,
/ (O) n (e) /, / (n) e (g) /, / (e) g
(A) /, / (g) a (i) /, / (a) i (s) /,
/ (I) s (i) /, / (s) i (m) /, / (i) m
(A) /, / (m) a (s) /, / (a) s (u) /,
/ (S) u (#) /, a total of 18 m phoneme set models are connected. Here, / # / means a silent section before and after the utterance. In the fifth embodiment, each m-phoneme set model is shown in FIG.
It is assumed that the device has a five-state structure as shown in FIG. FIG.
, State 1 is an initial state and state 5 is a final state. FIG. 18 illustrates a series connection model for / yoyokuoneigashimasu /.

(3) Recognition target vocabulary model creation procedure 3: Next, the recognition target vocabulary model creation unit 28 refers to the m phoneme set notation held in the dialogue speech m phoneme set model memory 16,
/ (#) Y (o) /, / is an m (= 3) phoneme set that constitutes the above / yoyokuoneigaishimasu /
(Y) o (y) /, / (o) y (a) /, / (y) a
(K) /, / (a) k (u) /, / (k) u (o) /,
/ (U) o (n) /, / (o) n (e) /, / (n) e
(G) /, / (e) g (a) /, / (g) a (i) /,
/ (A) i (s) /, / (i) s (i) /, / (s) i
(M) /, / (i) m (a) /, / (m) a (s) /,
Of the / (a) s (u) /, / (s) u (#) /, dialog parameters 15 of the m phoneme set model existing in the m phoneme set notation held in the dialogue m m phoneme set model memory 16 are used. The speech m phoneme set model and the dialogue speech m are read from the speech m phoneme set model memory 16 and stored in the corresponding locations of the serial connection model created in the recognition target vocabulary model creation procedure 1.
A parallel connection model for the selected vocabulary to be recognized is created by connecting the phoneme set models in parallel.

The connection in parallel means that the initial state and the final state of the m-phoneme set model and the m-phoneme set model to be connected are shared, and the transition from one initial state to either model is performed. It is possible to connect so that the transition to either model is completed in a common final state. For example, if the m phoneme set existing in the m phoneme set notation held by the dialogue m phoneme set model memory 16 is / (y) o (y) /, / (o) y (a) /, /
(A) k (u) /, / (u) o (n) /, / (n) e
(G) /, / (e) g (a) /, / (g) a (i) /,
/ (I) m (a) /, the dialogue voice m
The phoneme set model is connected in parallel to the series connection model of FIG. 18 as shown in FIG. 19 to create a parallel connection model.

(4) Recognition target vocabulary model creation procedure 4: Next, the recognition target vocabulary model creation unit 28 refers to the n phoneme set notation held in the dialogue speech n phoneme set model memory 21,
It is an n (n = 5 in this example) phoneme set that constitutes the above / yoyokuoneigashimasu // (##) y
(Oy) /, / (# y) o (ya) /, / (yo) y
(Ak) /, / (oy) a (ku) /, / (ya) k
(Uo) /, / (ak) u (on) /, / (ku) o
(Ne) /, / (uo) n (eg) /, / (on) e
(Ga) /, / (ne) g (ai) /, / (eg) a
(Is) /, / (ga) i (si) /, / (ai) s
(Im) /, / (is) i (ma) /, / (si) m
(As) /, / (im) a (su) /, / (ma) s
Of (u #) /, / (as) u (##) /, the parameter 20 of the model of the n phoneme set in the n phoneme set notation held in the dialogue n phoneme set model memory 21 is set to the dialogue sound n.
The recognition target vocabulary model for the selected recognition target vocabulary is created by reading from the phoneme set model memory 21 and further connecting in parallel to a corresponding place of the parallel connection model created in the recognition subject vocabulary model creation procedure 3.
For example, the n phoneme set present in the n phoneme set notation held in the dialogue phonetic n phoneme set model memory 21 is / (# y) o (y
a) /, / (yo) y (ak) /, / (ne) g (a
i) / If there are three, the recognition target vocabulary model creating unit 28 connects the dialogue speech n phoneme set models as shown in FIG. 20 to create the recognition target vocabulary model. The recognition target vocabulary model creating unit 28 sends the parameters of the recognition target vocabulary model that has been connected and the phonetic notation 29 thereof to the recognition target vocabulary model memory 31.

(5) Recognition-target vocabulary model creation procedure 5: Next, the recognition-target vocabulary model creation unit 28 refers to the recognition-target vocabulary memory 26 and recognizes all the recognition-target vocabularies existing in the recognition-target vocabulary memory 26. Until the creation of the target vocabulary model is completed, the next recognition target vocabulary to be model-created is selected in the order of the vocabulary number, and the above-described recognition target vocabulary model creation steps 2 to 4 are repeated. When the recognition target vocabulary model creation unit 28 creates the recognition target vocabulary models for all the recognition target vocabularies existing in the recognition target vocabulary memory 26 in this way, the recognition target vocabulary model creation procedure ends.

Next, the recognition operation of the speech recognition apparatus according to the fifth embodiment will be described. Before starting the recognition operation, the recognition unit 32 reads the parameters of all the recognition target vocabulary models stored in the recognition target vocabulary model memory 31 and the phonemic notation modeled by each recognition target vocabulary model. For example, if the recognition target vocabulary is as shown in FIG. 17, the recognizing unit 32 reads from the recognition target vocabulary model memory 31 1000 recognition target vocabulary models and phonemic notations corresponding to these recognition target vocabulary models.

The recognizing operation of the recognizing unit 32 is performed as follows.
When the audio signal 23 is input from the input terminal 22, the acoustic analysis unit 24 converts the audio signal 23 into a time series 25 of feature vectors. The time series 25 of this feature vector is, for example, LP
It is a time series of C cepstrum.

The recognizing unit 32 receives the time series 25 of the feature vector as input, calculates the likelihood with all the vocabulary models to be read in advance by, for example, the Viterbi algorithm, and obtains the vocabulary to be recognized having the highest likelihood. The phoneme notation modeled by the model is output as a recognition result 33.

Next, a specific description will be given of a speech recognition method used by the speech recognition apparatus according to the fifth embodiment. As described above, in the voice recognition method according to the fifth embodiment,
Before performing recognition, a vocabulary model to be recognized is created, and the created vocabulary model to be recognized is stored in the vocabulary model memory 31 to be recognized. First, a procedure for creating a vocabulary model to be recognized will be described.

FIG. 21 is a flowchart showing details of the procedure for creating a recognition target vocabulary model in the speech recognition method according to the fifth embodiment of the present invention. In the following, referring to FIG. Will be described.

First, the recognition target vocabulary model creation unit 28
In step ST1001, recognition target vocabulary memory 2
6, the vocabulary to be recognized having vocabulary number 1 to be model-created is selected, and the phoneme notation 27 of this vocabulary to be recognized is selected.
Is read from the recognition target vocabulary memory 26. For example, if the contents of the recognition target vocabulary memory 26 are as shown in FIG. 17, the recognition target vocabulary model creation unit 28 first has the vocabulary number 1 / y
Select oyakuonegaisimasu /.

Next, the recognition target vocabulary model creation unit 28
In step ST1002, in step ST10
01 or the parameters 13 of the read-aloud m-phoneme set model from the read-aloud m-phoneme set model memory 14 in accordance with the phoneme notation 27 of the vocabulary to be recognized selected in step ST1007, and the read-aloud m-phoneme set models are connected in series. Create a serial connection model for the vocabulary to be recognized. For example, phoneme notation is / yoyakuo
In the case of negaishimasu /, if m = 3, then / (#) y (o) /, / (y) o (y) /, /
(O) y (a) /, / (y) a (k) /, / (a) k
(U) /, / (k) u (o) /, / (u) o (n) /,
/ (O) n (e) /, / (n) e (g) /, / (e) g
(A) /, / (g) a (i) /, / (a) i (s) /,
/ (I) s (i) /, / (s) i (m) /, / (i) m
(A) /, / (m) a (s) /, / (a) s (u) /,
/ (S) u (#) /, a total of 18 m phoneme set models are connected. Here, / # / means a silent section before and after the utterance. As described above, in the fifth embodiment, each m-phoneme set model as shown in FIG. 26 has a five-state structure. Therefore, phonemic notation / yoyaku
FIG. 18 shows a series connection model for onegaijima /.

Next, the recognition target vocabulary model creation unit 28
In step ST1003, by referring to the m-phoneme set notation held in the dialogue m-phoneme set model memory 16, the m-phoneme set constituting the above-described phoneme notation / yoyakuonegai simasu // (#) y (o) /, / (y ) O
(Y) /, / (o) y (a) /, / (y) a (k) /,
/ (A) k (u) /, / (k) u (o) /, / (u) o
(N) /, / (o) n (e) /, / (n) e (g) /,
/ (E) g (a) /, / (g) a (i) /, / (a) i
(S) /, / (i) s (i) /, / (s) i (m) /,
/ (I) m (a) /, / (m) a (s) /, / (a) s
Of (u) /, / (s) u (#) /, the parameter 15 of the m phoneme set model present in the m phoneme set notation held in the dialogue m m phoneme set model memory 16 is used as the dialogue m m phoneme set model. The data is read from the memory 16 and is read from the step ST10
A parallel connection model for the selected vocabulary to be recognized is created by connecting the read-aloud m-phoneme set model and the dialogue m-phoneme set model in parallel to the corresponding location of the serial connection model created in step 02.

For example, conversational speech m phoneme group model memory 1
The m phoneme set present in the m phoneme set notation held by 6 is /
(Y) o (y) /, / (o) y (a) /, / (a) k
(U) /, / (u) o (n) /, / (n) e (g) /,
/ (E) g (a) /, / (g) a (i) /, / (i) m
Assuming that the number of (a) / is eight, the recognition target vocabulary model creation unit 28 converts these conversational speech m phoneme set models into
To create a parallel connection model.

Next, the recognition target vocabulary model creation unit 28
In step ST1004, by referring to the n phoneme set notation held in the dialogue phonetic n phoneme set model memory 21, n (n = 5 in the fifth embodiment) a phoneme set constituting the above phoneme notation / yoyakuoneigaishimasu // ( ##) y (oy) /, / (# y) o (ya) /, /
(Yo) y (ak) /, / (oy) a (ku) /, /
(Ya) k (uo) /, / (ak) u (on) /, /
(Ku) o (ne) /, / (uo) n (eg) /, /
(On) e (ga) /, / (ne) g (ai) /, /
(Eg) a (is) /, / (ga) i (si) /, /
(Ai) s (im) /, / (is) i (ma) /, /
(Si) m (as) /, / (im) a (su) /, /
Of the (ma) s (u #) / and / (as) u (##) /, the parameters 20 of the n-phoneme set model present in the n-phoneme set notation held in the dialogue n-phoneme set model memory 21
Is read from the dialogue n-phoneme set model memory 21 and connected in parallel to the corresponding location of the parallel connection model created in step ST1003 to create a recognition target vocabulary model for the selected recognition target vocabulary. For example, the n phoneme set existing in the n phoneme set notation held by the dialogue phonetic n phoneme set model memory 21 is / (# y) o.
(Ya) /, / (yo) y (ak) /, / (ne) g
Assuming that (ai) / is 3, the dialogue n-phoneme set models are connected as shown in FIG. 20 to create a vocabulary model to be recognized.

Next, the recognition target vocabulary model creation unit 28
In step ST1005, the parameters of the vocabulary model to be recognized that have been connected in parallel and the phoneme notation 29 thereof are sent to the vocabulary model memory 31 to be recognized. The recognition target vocabulary model memory 31 holds the parameters of the received recognition target vocabulary model and the phoneme notation 29.

Next, the recognition target vocabulary model creation unit 28
In step ST1006, the recognition target vocabulary memory 2
It is checked whether or not a recognition target vocabulary model has been created for all the recognition target vocabularies existing in the recognition target vocabulary memory 26 with reference to No. 6; if there is an uncreated recognition target vocabulary, the process proceeds to step ST1007; The next vocabulary to be recognized is selected from the vocabulary memory to be recognized 26, and step ST10 is performed.
Return to 02. On the other hand, the recognition target vocabulary model creation unit 28
If the recognition target vocabulary for which the recognition target vocabulary model has not been created does not exist in the recognition target vocabulary memory 26, the recognition target vocabulary model creation procedure ends.

Next, the speech recognition procedure of the speech recognition method according to the fifth embodiment will be specifically described. As described above, before starting the recognition operation, the recognizing unit 32 determines the parameters of all the recognition target vocabulary models stored in the recognition target vocabulary model memory 31 and the phonemic notation modeled by each recognition target vocabulary model. And read. For example, if the recognition target vocabulary is as shown in FIG. 17, the recognizing unit 32 reads from the recognition target vocabulary model memory 31 1000 recognition target vocabulary models and phonemic notations corresponding to these recognition target vocabulary models.

FIG. 22 is a flowchart showing details of the speech recognition procedure in the speech recognition method according to the fifth embodiment of the present invention. The speech recognition procedure will be described below with reference to FIG. First, the acoustic analysis unit 24
In step ST1201, the audio signal 23 input from the input terminal 22 is converted into a time series 25 of feature vectors. The time series 25 of the feature vector is a time series of the LPC cepstrum.

Then, the recognizing unit 32 determines in step ST120
2, the likelihood with all recognition target vocabulary models read in advance is calculated by, for example, a Viterbi algorithm, and the recognition target vocabulary model showing the highest likelihood is modeled. The phoneme notation to be output is output as the recognition result 33.

When the speech recognition method according to the fifth embodiment is to be implemented by software, a recognition target vocabulary model creating step for creating a speech pattern model (that is, a recognition target vocabulary model) for the recognition target vocabulary, Signal 2 input from the input terminal 22
3 is converted into a time series 25 of feature vectors, and the time series 25 of the feature vectors are input, and the likelihoods of all the vocabulary models to be recognized which are read in advance are calculated by, for example, a Viterbi algorithm. Outputting the phoneme notation modeled by the recognition-target vocabulary model showing the highest likelihood as the recognition result 33. The computer-readable recording device stores a speech recognition program for causing the computer to execute speech recognition. A recording medium is required.

As described above, the speech recognition apparatus according to the fifth embodiment, as shown in FIG. 20, has a read-out speech m phoneme group model trained by the speech pattern model learning apparatus according to the third or fourth embodiment. Using m-phoneme group model and m-phoneme group model for dialogue speech, vocabulary to be recognized for m-phoneme group and n-phoneme group with fast utterance and low recognition performance like dialogue voice A voice pattern model is created separately and its m
The acoustic features of the phoneme set and the n-phoneme set are modeled with high accuracy, and the vocabulary model to be recognized is created by connecting the read-aloud speech m-phoneme set model in parallel. Therefore, according to the fifth embodiment, it is possible to recognize a polite utterance such as a reading voice with high accuracy, and to improve the recognition accuracy even for an utterly fast and ambiguous voice such as a dialogue voice. To play. Although the fifth embodiment has been described assuming that m = 3 and n = 5, m and n may be selected from any set of integers such that m <n. In this case, the same effect is obtained. .

[0190]

As described above, according to the present invention, a read-out voice m learned using a voice read out of a text.
M phoneme set extraction means or m phoneme set extraction step for extracting a m phoneme set whose recognition rate is equal to or less than a predetermined threshold value from the dialogue speech learning data using a phoneme set model; Since it is configured to include a model learning means or a model learning step for learning a dialogue m m phoneme set model using learning data, all m
Without learning the m-phoneme group model for the dialogue speech for the phoneme group, a m-phoneme group model for dialogue speech that can recognize dialogue speech that was difficult to recognize with the m-phoneme group model read aloud speech was used. It has the effect of being able to learn efficiently.

According to the present invention, the m-phoneme-set extracting means or the m-phoneme-set extracting step determines that the number of data having the same m-phoneme-set notation is more than a predetermined number from the interactive speech learning data.
Since the phoneme set is selected, the m-phoneme set selected using the read-aloud m-phoneme set model is recognized, and if the recognition rate is equal to or less than a predetermined threshold, the selected m-phoneme set is extracted, In the m-phoneme set model of the spoken m-phoneme set having a low recognition rate, learning of a model having a data number less than a predetermined number and having a statistically low reliability is avoided, and only a model having a statistically high reliability is used. There is an effect that can be learned efficiently.

According to the present invention, an m-phoneme set whose recognition rate is equal to or less than a first predetermined threshold is extracted from conversational speech learning data by using a m-phoneme set model of a read-out voice trained using a text-reading voice. M phoneme set extraction means or m phoneme set extraction step, and for each extracted m phoneme set, a dialogue speech m phoneme set model learning means or dialogue speech for learning a dialogue speech m phoneme set model using the dialogue speech learning data. Using an m-phoneme group model learning step, and extracting the n-phoneme group whose recognition rate is equal to or less than a second predetermined threshold from the dialogue speech learning data, using the read-out speech m-phoneme group model and the dialogue m-phoneme group model. The n phoneme set extracting means or the n phoneme set extraction step to perform, and for each of the extracted n phoneme sets, a dialogue speech n phoneme set model is learned using the dialogue speech learning data. It is configured to include the dialogue speech n-phoneme set model learning means or the dialogue speech n-phoneme set model learning step, so that the speech rate is as fast and vocal as the dialogue speech. With the phoneme set model, there is an effect that the conversational speech n phoneme set model can be efficiently learned for each n phoneme set for which sufficient recognition performance cannot be obtained.

According to the present invention, the n-phoneme-set extracting means or the n-phoneme-set extracting step determines that the number of data having the same n-phoneme-set notation is more than a predetermined number from the interactive speech learning data.
Selecting a phoneme set, recognizing the selected n phoneme set using the read-aloud speech m-phoneme set model and the dialogue speech m-phoneme set model, and selecting the n-phoneme set if the recognition rate is equal to or less than a second predetermined threshold value; Since n phoneme sets are extracted, learning of a model with less than a predetermined number of data and statistically low reliability among n phoneme sets of conversational speech with a low recognition rate in the n-phoneme set model of the read-out voice is avoided. However, there is an effect that only a model having a statistically high reliability can be efficiently learned.

According to the present invention, a read-aloud m m-phoneme set model, a dialogue m-phoneme set model, and a dialogue n-phoneme set model learned by the voice pattern model learning apparatus or the voice pattern model learning method are connected in parallel. A recognition target vocabulary model generating means or a recognition target vocabulary model generating step of generating a voice pattern model for the recognition target vocabulary by using the voice pattern model for the recognition target vocabulary generated by the recognition target vocabulary model generating means. Since it is configured to have a recognition means or a recognition step for performing recognition, it is possible to recognize polite utterances such as read-out voices with high accuracy, and to improve recognition accuracy even for vocal voices whose utterance speed is fast like dialogue voices There is an effect that can be.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a speech pattern model learning device according to Embodiment 1 of the present invention.

FIG. 2 is a diagram showing an example of the contents of an extracted m phoneme set notation memory of the speech pattern model learning device according to the first embodiment of the present invention.

FIG. 3 is a flowchart showing a procedure of a voice pattern model learning method according to the first embodiment of the present invention.

FIG. 4 is a flowchart showing a learning procedure of a read-aloud m-phoneme set model according to Embodiment 1 of the present invention;

FIG. 5 is a flowchart showing an extraction procedure for extracting m phoneme sets having a low recognition rate according to Embodiment 1 of the present invention.

FIG. 6 is a dialogue voice m according to the first embodiment of the present invention.
It is a flowchart which shows the learning procedure of a phoneme set model.

FIG. 7 is a flowchart showing an extraction procedure for extracting a m-phoneme set having a low recognition rate and having a predetermined number of tokens or more in the voice pattern model learning device according to the second embodiment of the present invention.

FIG. 8 is a block diagram showing a configuration of a speech pattern model learning device according to Embodiment 3 of the present invention.

FIG. 9 is a diagram showing an example of a pentaphone set table held in a conversational speech learning data memory of the speech pattern model learning device according to the third embodiment of the present invention.

FIG. 10 is a diagram showing an example of a five-phoneme set notation provided together with a three-phoneme set notation held in a dialogue speech learning data memory of the speech pattern model learning device according to the third embodiment of the present invention;

FIG. 11 is a diagram showing an example of the contents of an extracted n phoneme set notation memory of the speech pattern model learning device according to the third embodiment of the present invention.

FIG. 12 is a flowchart showing a procedure of a voice pattern model learning method according to Embodiment 3 of the present invention.

FIG. 13 is a flowchart showing an extraction procedure for extracting n phoneme sets having a low recognition rate according to Embodiment 3 of the present invention.

FIG. 14 is a flowchart showing a learning procedure of a conversational speech n-phoneme set model according to Embodiment 3 of the present invention.

FIG. 15 is a flowchart showing an extraction procedure for extracting an n-phoneme set having a predetermined number of tokens or more and a low recognition rate in the speech pattern model learning device according to the fourth embodiment of the present invention.

FIG. 16 is a block diagram showing a configuration of a speech recognition device according to Embodiment 5 of the present invention.

FIG. 17 is a diagram showing an example of the contents of a recognition target vocabulary memory of the speech recognition device according to the fifth embodiment of the present invention.

FIG. 18: Recognition target vocabulary / yoyakuonegai
It is a figure which shows the series connection model with respect to simasu /.

FIG. 19 is a diagram showing a parallel connection model created by the voice recognition device according to the fifth embodiment of the present invention and in which the dialogue m-phoneme set model is connected in parallel to the serial connection model of FIG. 18;

FIG. 20 is a diagram illustrating a recognition target vocabulary model created by the speech recognition apparatus according to the fifth embodiment of the present invention, in which the dialogue n phoneme group model and the dialogue n phoneme group model are connected in parallel to the serial connection model of FIG. 18; FIG.

FIG. 21 is a flowchart showing a procedure for creating a vocabulary model to be recognized in the speech recognition method according to the fifth embodiment of the present invention.

FIG. 22 is a flowchart showing details of a voice recognition procedure in a voice recognition method according to Embodiment 5 of the present invention.

FIG. 23 is a block diagram showing a configuration of an example of a conventional voice pattern model learning device.

FIG. 24 is a diagram showing an example of a three-phoneme set table held in a learning data memory of a conventional speech pattern model learning device.

FIG. 25 is a diagram showing an example of a three-phoneme set notation of a token stored in a learning data memory of a conventional voice pattern model learning device.

FIG. 26 is a diagram illustrating a 5-state left-to-right model that is an example of the structure of a three-phoneme set model.

[Explanation of symbols]

3 model learning unit (model learning means), 6 reading voice learning data memory, 8,80 conversational voice learning data memory, 10 m phoneme group extraction unit (m phoneme group extraction means), 1
2 extracted m phoneme set notation memory, 14 read-aloud speech m phoneme set model memory, 16 conversational speech m phoneme set model memory, 17 n phoneme set extraction unit (n phoneme set extraction means), 19
Extracted n phoneme set notation memory, 21 dialogue speech n phoneme set model memory, 24 acoustic analysis unit, 26 recognition target vocabulary memory, 28 recognition target vocabulary model creation unit (recognition target vocabulary model creation means), 30 model learning unit (interaction speech) m phoneme set model learning means, dialogue speech n phoneme set model learning means), 31 recognition target vocabulary model memory, 32 recognition unit (recognition means).

 ──────────────────────────────────────────────────続 き Continued on the front page (54) [Title of the Invention] A voice pattern model learning device, a voice pattern model learning method, a computer-readable recording medium storing a voice pattern model learning program, a voice recognition device, and voice. Recognition method and computer-readable recording medium recording voice recognition program

Claims (15)

[Claims] 1. A m-phoneme set model, which is a phoneme taking into account the difference between (m-1) / 2 phonemes before and after the m-phoneme set, which is learned using a voice read out from a text for the m-phoneme set. M phoneme set extraction means for recognizing each m phoneme set included in dialogue speech learning data obtained by acoustic analysis of a person-to-person dialogue speech and extracting m phoneme sets whose recognition rate is equal to or less than a predetermined threshold value; A speech pattern model learning apparatus comprising: a model learning means for learning a dialogue m m phoneme set model using the dialogue speech learning data for each m phoneme set extracted by the m phoneme set extraction means. 2. An m-phoneme set extraction unit selects m-phoneme sets in which the number of data having the same m-phoneme set notation is equal to or more than a predetermined number from the interactive speech learning data, and selects the m-phoneme set using a read-out m-phoneme set model. 2. The speech pattern model learning apparatus according to claim 1, wherein the m phoneme sets are recognized, and the selected m phoneme sets are extracted if the recognition rate is equal to or less than a predetermined threshold. 3. A m-phoneme set model, which is a phoneme that takes into account the difference between (m-1) / 2 phonemes before and after the m-phoneme set and is trained using a speech read out text for the m-phoneme set. M phoneme set extraction means for recognizing each m phoneme set included in dialogue speech learning data obtained by acoustic analysis of a person-to-person dialogue speech and extracting a m phoneme set whose recognition rate is equal to or less than a first predetermined threshold value A dialogue m-phoneme set model learning means for learning a dialogue m-phoneme set model using the dialogue speech learning data for each m-phoneme set extracted by the m-phoneme set extraction means; Each n-phoneme set included in the dialogue speech learning data is a phoneme that takes into account differences in phonemes in a longer range than the m-phoneme set where n> m, using the set model and the m-phoneme set model for dialogue speech. Recognize N phoneme group extraction means for extracting an n phoneme group whose recognition rate is equal to or less than a second predetermined threshold value; and for each n phoneme group extracted by the n phoneme group extraction means, a dialogue speech using the above dialogue speech learning data. A speech pattern model learning device comprising: a dialogue speech n phoneme group model learning means for learning an n phoneme group model. 4. An n-phoneme set extracting means selects n-phoneme sets whose number of data having the same n-phoneme set notation is equal to or more than a predetermined number from the dialogue speech learning data, and reads out a read-out speech m-phoneme set model and a dialogue speech m. 4. The method according to claim 3, wherein the selected n phoneme sets are recognized using a phoneme set model, and the selected n phoneme sets are extracted if the recognition rate is equal to or less than a second predetermined threshold. Voice pattern model learning device. 5. A reading voice m learned by the voice pattern model learning device according to claim 3.
A recognition target vocabulary model creating unit for creating a speech pattern model for the recognition target vocabulary by connecting a phoneme group model, a dialogue m m phoneme group model, and a dialogue speech n phoneme group model in parallel; A speech recognition device comprising: a recognition unit configured to recognize an input speech by using a created speech pattern model for a recognition target vocabulary. 6. A m-phoneme set model, which is a phoneme that takes into account the difference between (m-1) / 2 phonemes before and after the m-phoneme set and is trained using a voice read out from a text for the m-phoneme set, Recognize each m-phoneme set included in dialogue speech learning data obtained by acoustic analysis of human-to-person dialogue speech, extract m-phoneme sets whose recognition rate is equal to or less than a predetermined threshold, and for each of the extracted m-phoneme sets. And a speech pattern model learning method for learning a m-phoneme set model of conversational speech using the conversational speech learning data. 7. When extracting m phoneme sets, m phoneme sets whose number of data having the same m phoneme set notation is equal to or more than a predetermined number are selected from conversational speech learning data, and a m-phoneme set model for reading out speech is used. 7. The speech pattern model learning method according to claim 6, wherein the m phoneme set selected by the above is recognized, and if the recognition rate is equal to or less than a predetermined threshold, the selected m phoneme set is extracted. 8. A m-phoneme set model, which is a phoneme taking into account the difference between (m-1) / 2 phonemes before and after the m-phoneme set, which is trained using a speech obtained by reading a text for the m-phoneme set, Recognize each m-phoneme set included in dialogue-speech learning data obtained by acoustic analysis of a person-to-person dialogue voice, and extract m-phoneme sets whose recognition rate is equal to or less than a first predetermined threshold,
For each of the extracted m-phoneme sets, a dialogue m-phoneme set model is learned using the dialogue speech learning data, and n> m is obtained using the read-out speech m-phoneme set model and the dialogue m-phoneme set model. Recognize each of the n phoneme sets included in the conversational speech learning data, which are phonemes taking into account differences in phonemes in a range longer than the m phoneme sets, and generate n phoneme sets whose recognition rate is equal to or less than a second predetermined threshold. A speech pattern model learning method for learning a conversational speech n-phoneme set model using the conversational speech learning data for each extracted and extracted n-phoneme set. 9. When extracting n phoneme sets, an n phoneme set whose number of data having the same n phoneme set notation is equal to or greater than a predetermined number is selected from the dialogue learning speech data, and a dialogue with the m-phoneme set model of the spoken speech is performed. 9. The method according to claim 8, further comprising: recognizing the selected n phoneme set by using a speech m phoneme set model; and extracting the selected n phoneme set if the recognition rate is equal to or less than a second predetermined threshold. The described speech pattern model learning method. 10. A method of connecting a speech m-phoneme set model, a dialogue m-phoneme set model, and an interactive speech n-phoneme set model learned by the speech pattern model learning method according to claim 8 in parallel. A speech recognition method for creating a speech pattern model for a recognition target vocabulary and recognizing an input speech using the created speech pattern model for the recognition target vocabulary. 11. A read-aloud speech m-phoneme set model, which is trained using a speech obtained by reading a text about a m-phoneme set, which is a phoneme in consideration of a difference between each of the preceding and succeeding (m-1) / 2 phonemes, Recognize each m phoneme set included in the conversational speech learning data obtained by acoustically analyzing the conversational speech between people,
An m-phoneme set extracting step of extracting an m-phoneme set whose recognition rate is equal to or less than a predetermined threshold value; and a m-phoneme set extracted by the m-phoneme set extracting step, the m-phoneme set of the dialogue speech using the dialogue speech learning data. Dialogue m to learn the model
A computer-readable recording medium storing a speech pattern model learning program, comprising: a phoneme group model learning step. 12. The m-phoneme set extraction step selects m-phoneme sets whose number of data having the same m-phoneme set notation is equal to or more than a predetermined number from the dialogue speech learning data, and selects the m-phoneme set using a read-out m-phoneme set model. 12. The recording medium according to claim 11, further comprising the step of recognizing the m phoneme sets and extracting the selected m phoneme sets if the recognition rate is equal to or less than a predetermined threshold. 13. A read-aloud speech m-phoneme set model, which is learned using a text-to-speech voice for a m-phoneme set, which is a phoneme taking into account the difference between each of the preceding and following (m-1) / 2 phonemes, Recognize each m phoneme set included in the conversational speech learning data obtained by acoustically analyzing the conversational speech between people,
An m-phoneme set extraction step of extracting m-phoneme sets whose recognition rate is equal to or less than a first predetermined threshold value, and a dialogue using the dialogue speech learning data for each m-phoneme set extracted in the m-phoneme set extraction step. A dialog speech m phoneme set model learning step for learning a speech m phoneme set model, and a range longer than the m phoneme set satisfying n> m, using the read-out speech m phoneme set model and the dialogue speech m phoneme set model. An n-phoneme set extraction step of recognizing each of the n-phoneme sets included in the conversational speech learning data, and extracting an n-phoneme set whose recognition rate is equal to or less than a second predetermined threshold value. A conversational speech n phoneme group model learning step of learning a conversational speech n phoneme group model using the conversational speech training data for each n phoneme group extracted in the n phoneme group extraction step. That a computer-readable recording medium recording a speech pattern model training program. 14. An n-phoneme group extracting step selects n-phoneme groups in which the number of data having the same n-phoneme group notation is equal to or greater than a predetermined number from the dialogue speech learning data, and reads out a read-out speech m-phoneme set model and a dialogue speech m. Recognizing the selected n phoneme set using a phoneme set model, and extracting the selected n phoneme set if the recognition rate is equal to or less than a second predetermined threshold value. Item 14. The recording medium according to Item 13. 15. A speech m-phoneme set model, a dialogue m-phoneme set model, and a dialogue n-phoneme set model, which are learned by the speech pattern model learning method according to claim 8 or 9, are connected in parallel. It has a recognition target vocabulary model creating step of creating a voice pattern model for the recognition target vocabulary, and a recognition step of recognizing the input voice using the voice pattern model for the recognition target vocabulary created in the recognition target vocabulary model creating step. And a computer-readable recording medium on which a voice recognition program is recorded.