JPH0756595A - Voice recognizer - Google Patents

Abstract

PURPOSE:To recognize voice while coping with the varying conditions of a user's uttering and the changes in the user by controlling a collating means or a discriminating means based on the detection results of the changes in the uttering conditions and the changes in the user. CONSTITUTION:Voices, that are inputted and digitized through an input means 1, are acoustically analyzed for every constant time interval by a first analysis means 2 and the result of the analysis is outputted in a form which is suitable to a collating means 4. The means 4 performs collation between time sequential patterns and a standard pattern and outputs the score against each standard pattern. The score outputted from the means 4 is inputted to a discrimination means 5 and a candidate corresponding to a best scored standard pattern or plural higher ranking candidates are outputted as recognition results. A second analysis means 3 analyzes voices inputted through the means 1, extracts the changes in uttering conditions and the changes in uttering speeds, outputs these information and controls the means 4 and 5 based on the outputs.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device,
In particular, the present invention relates to a device for satisfactorily recognizing a voice when the utterance mode of the same speaker varies and a voice when the speaker changes.

[0002]

2. Description of the Related Art In a conventional voice recognition device, for example, a word recognition device, it is premised that a unit for uttering a voice is a word. When a plurality of words are continuously uttered to this device, the plurality of continuously uttered words are regarded as one word, and a correct recognition result is often not obtained. In this way, the user is limited to only vocalization divided into words.

Further, when the voice recognition device makes a mistaken recognition and the user politely rephrases by dividing each sound by one note, each sound produced by dividing the sound is regarded as one word. It becomes more difficult to recognize.

[0004]

SUMMARY OF THE INVENTION An object of the present invention is to enable recognition of voice in response to changes in the way the user speaks, changes in the speaker, and the like.

[0005]

The object of the present invention is to provide a second analysis means for detecting changes in various modes of utterance and changes in the speaker, and based on the results of the second analysis means. This is achieved by controlling the matching means or the judging means.

[0006]

According to the present invention, the collating means or the judging means is controlled on the basis of the result of analyzing the manner of utterance or the change of the speaker, so that it is possible to respond to the change of various manners of utterance and the change of the speaker. Can recognize voice.

[0007]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of a voice recognition device of the present invention. The present invention is different from the conventional one in that the second analyzing means 3 is provided to control the collating means 4 or the judging means 5. The voice digitized and input through the input unit 1 is sent to the first analysis unit 2 where acoustic analysis is performed at regular time intervals. The result of the first analysis means 2 is output as a format desired by the matching means 4 (for example, a time series pattern of a feature vector or a time series pattern of vector quantized code). The collation unit 4 collates the time series pattern, which is the acoustic analysis result obtained from the first analysis unit 2, with the standard pattern prepared in the standard pattern storage unit 6 in advance as the collation reference, The score for each standard pattern is output. The score output from the matching unit 4 is input to the determination unit 5, and one or a plurality of candidates of the higher rank corresponding to the standard pattern having the best score are output as the recognition result. Input means 1 up to here,
The first analysis unit 2, the collation unit 4, and the determination unit 5 have the same configuration as the conventional voice recognition device. The second analyzing means 3 which is different from the conventional one in the present invention analyzes the voice input through the input means 1, extracts the change in the utterance mode and the change in the utterance speed, and outputs this information. Then, the collating means 4 or the judging means 5 is controlled by the output of the second analyzing means 3.

In this embodiment, the information extracted by the second analysis means is set to the utterance mode. The speaking mode means a speaking form and a speaking style. In the case of modes, we consider the existence of multiple modes, but here we consider the three modes of "speaking in syllable units", "speaking in word units", and "sentences in sentence units".
2) Word mode, 3) Sentence mode. In the case of 1), when trying to convey a new word, or when the other party misses listening, one voice is produced slowly or in a segmented manner. For example, “Koku Kubu” Say "Every tone" carefully. In the case of 2), as in the case of transmitting a command or relatively simple information, one word is uttered, or a plurality of words are divided into word units and uttered, for example, Say "Kokubunji". In the case of 3), it is a case where the user speaks normally in sentences, for example, "I want to go to Kokubunji".

Next, the second analyzing means 3 for detecting the vocalization mode will be described in detail.

FIG. 2 is a block diagram showing an embodiment of the second analyzing means for detecting the utterance mode. FIG. 3 shows some of the information to be input / output in the block diagram in FIG. 2, and (a) to (f) are associated with each other in FIG. 2 and FIG. A voice having an amplitude w (t) as shown in FIG. 3 (a) is input to the power calculating means 301, and power (short section power) as shown in FIG.

[0012]

[Equation 1]

Is output. However, T is the section width of the short section analysis. The short section power pw (t) is input to the power threshold value judging means 302, converted into 0 (no power) / 1 (with power), and a speech section sp (t) as shown in FIG. 3C is output. . The short-term power pw (t) is also input to the power change amount calculation means 304, and according to the following equation,

[0014]

## EQU00002 ## dpw (t + 1) = | pw (t + 1) -p
w (t) | Power change amount dpw (t) as shown in FIG. 3 (d) is calculated. The power change amount dpw (t) is calculated by the change amount threshold determination means 3
It is input to 05, according to the following formula,

[0015]

[Equation 3] if dpw (t) ≦ DPW _TH then f
ix (t) = 1 else fix (t) = 0 It is determined whether it is a stationary part, and 0 (unsteady) / 1 (steady)
As shown in FIG. 3E, the steady section fix (t) is output. However, DPW _TH is a constant determined for each system. Next, the vowel steady section determining unit 306 receives the output sp (t) from the power threshold determining unit 302 and the output fix (t) from the variation threshold determining unit 305 as inputs,

[0016]

## EQU00004 ## spfix (t) = sp (t) & fix
(t) (& is a logical product) is used to determine the stationary segment due to vowels (vowel stationary segment) spfi
x (t) is output at 0/1 as shown in FIG. Subsequently, the stationary section length calculating unit 307 is configured to include the vowel stationary section determining unit 3
The stationary section length (fixsz) is obtained by the number of consecutive 1's in the 0/1 sequence of spfix (t) output from 06. The steady section evaluating means 308 is the steady section calculating means 3
Every time the steady section length is obtained from 07,

[0017]

[Equation 5] if fixsz ≧ SZ1 _TH the
The number n _A of long steady intervals or the number n _B of short steady intervals is obtained by n n _A = n _A +1 else if fixzz ≧ SZ2 _TH then n _B = n _B +1. However, the initial values of n _A and n _B are both 0. Further, SZ1 _TH and SZ2 _TH are constants determined for each system, and SZ1 _TH > SZ2 _TH . Finally, when the voice section detecting means 303 detects the end of the voice, the mode determining means 309 is activated. The mode determination means 309 receives n _A and n _B from the steady section evaluation means 308, and determines the mode as follows. Here, n represents the total number of syllables, and n = n _A + n _B.

[0018]

[Equation 6] if n _A / n> N1 _TH

[0019]

[Equation 7] or n <N2 _TH then mode =
Syllable mode else if n <N3 _TH then mode = word mode else mode = text mode However, N1 _TH , N2 _TH , and N3 _TH are constants determined for each system. First, the mode determination unit 309 determines whether the input voice is a syllable mode in which the input voice is uttered slowly and carefully depending on whether or not the ratio of the number n _A of long steady intervals to the total number n of syllables exceeds a certain threshold. Further, the mode is determined by the size of the total number of syllables n. The collating means 4 or the judging means 5 is controlled by the output of the mode judging means 309.

In order to obtain the vowel stationary section, only the power change is used here. However, it is possible to obtain only the spectrum change or a combination of the power change and the spectrum change. Needless to say.

Next, the collating means 4 used in this embodiment will be described with reference to FIG.

FIG. 4 is a block diagram showing the structure of the collating means 4 when the mode is switched using the output of the second analyzing means 3. This is provided with a selecting means 44 in front of a plurality of collating means. The selecting means 44 outputs one or more of a plurality of collating means (in this example, the syllable collating means 41, the word collating means 42, and the sentence collating means 43) according to the output of the second analyzing means 3. The information from the first analysis means 2 is sent to the selected matching means. When a plurality of selections are made, the determination means 5 outputs one or a plurality of candidates as a recognition result based on the score. The HMM 61 stores a model for each syllable that is statistically learned in advance. The syllable matching means 41 performs syllable matching according to this model, and outputs one or a plurality of syllable candidates together with a score as a matching result. The word dictionary 62 stores information about words (for example, information about what syllable sequence is composed of). The word matching means 42
The word-by-word model is combined using the syllable-based model stored in the HMM 61 according to the information in the word dictionary 62, and one or a plurality of word candidates are output together with the score as a matching result. To do. Grammar 63
Stores the grammar. The sentence matching means 43 is the HMM 6
1. Matching is performed based on the word dictionary 62 and the grammar 63, and one or a plurality of sentences or clause candidates are output as a matching result together with the score.

The syllable matching means 41 and the word matching means 4
2. Various methods are conceivable for realizing the sentence matching unit 43, but here, an HMM (Hidden Mark) is used.
Consider a method using ovModel). For the method of realizing a speech recognition system using HMM, see "Seiji Nakagawa, Development of time series pattern matching algorithm in speech recognition,
AI Society, Vol.3, No.4, pp414-423, 1988. ”or“ Kai-Fu Lee, Automatic speech recognition: th
e development of the SPHINX system, Kluwer Academi
c Publisher, 1989. ”.

Next, another embodiment of the matching means 4 will be described with reference to FIG.

FIG. 5 is a block diagram showing the configuration of the collating means 4 when the mode is switched using the output of the second analyzing means 3. The selecting means 44 is provided after the plurality of collating means. That is, a plurality of matching means (syllabic matching means 41, word matching means 42, sentence matching means 4).
3) operates in parallel, and the selecting means 44 selects one or more of the matching results from the matching means by the second analyzing means 3.
Select based on the results of. The configurations of the syllable matching unit 41, the word matching unit 42, and the sentence matching unit 43 may be the same as those in FIG.

Next, another embodiment of the matching means 4 will be described with reference to FIG.

FIG. 6 is a block diagram showing the structure of the collating means 4 when the judging means 5 is controlled by the output of the second analyzing means 3. The configuration is the same as that of FIG. 4 and the case except that there is no selection means and all the plurality of matching results are sent to the determination means 5.

Next, the judging means 5 used in this embodiment will be described.

The judging means 5 receives the output of the collating means 4 as an input. The determination means 5 outputs one candidate having the highest score or a plurality of candidates having the highest score as a recognition result. When a plurality of collating means are selected by the selecting means 44 in the collating means 4, the collating results are put together, and the judging means 5 selects one candidate having the best score or a higher rank based on the score. A plurality of candidates will be output as the recognition result. Further, the judging means 5 may receive the output of the second analyzing means 3 as an input in addition to the output of the collating means 4. The determination unit 5 corrects the score (for example, weights) the candidates sent from the matching unit 4 based on the information (in this case, the utterance mode) received from the second analysis unit 3. After that, one candidate with the highest score or a plurality of candidates with the highest score is output as a recognition result.

It goes without saying that both the collating means 4 and the judging means 5 can be controlled.

In the present embodiment, the second analysis means detects the voicing mode, but the second analysis means has a discrete value (for example, male or female voice or adult) related to speaker characteristics. It is possible to deal with changes in the speaker by performing an analysis for extracting (children).

Further, if the second analysis means performs an analysis for extracting a continuous value (for example, the number of syllables per unit time in the speech) related to the speech rate of the input speech, the speech is produced. Can respond to changes in speed.

[0033]

According to the present invention, the collating means or the judging means is controlled on the basis of the detection result of the change of the utterance mode and the change of the speaker, so that the various utterance changes of the user and the speaker are controlled. The voice can be recognized in accordance with the change of.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a voice recognition device of the present invention.

FIG. 2 is a block diagram showing an embodiment of a second analysis means.

FIG. 3 is an explanatory diagram showing a data flow in the second analysis means of the present embodiment.

FIG. 4 is a block diagram showing a configuration of a matching unit.

FIG. 5 is a block diagram showing another configuration of the matching unit.

FIG. 6 is a block diagram showing still another configuration of the matching unit.

[Explanation of Codes] 1 ... Input means, 2 ... First analysis means, 3 ... Second analysis means, 4 ... Collation means, 5 ... Judgment means, 6 ... Standard pattern storage means.

Claims (9)

[Claims] 1. A voice input means for inputting a voice, a first analyzing means for analyzing the voice input by the voice input means and outputting a time series pattern of a feature vector, which are prepared in advance as a recognition reference. By comparing the standard pattern storage means for storing the standard pattern with the time series pattern of the feature vector obtained from the standard pattern and the first analysis means,
In the voice recognition device, which comprises a collating means for obtaining a score for each standard pattern and a judging means for outputting one or a plurality of recognition candidates based on the score for each standard pattern, the voice input means inputs the voice recognition means. A voice recognition device characterized in that a second analysis means is provided for voice and the collation means and / or the determination means are controlled by using the output of the second analysis means. 2. The voice according to claim 1, wherein the second analysis means outputs a discrete value, and the collation means and / or the determination means is controlled by using the discrete value. Recognition device. 3. The matching means according to claim 2, wherein the matching means is provided with a plurality of matching means corresponding to the discrete values, and one or a plurality of the matching means are selected from the plurality of matching means based on the discrete values. A voice recognition device that properly selects and uses individual items. 4. The collating means according to claim 2, wherein the collating means is provided with a plurality of collating means corresponding to the discrete values, and all or a part of the plurality of collating means can be operated in parallel. Recognition device for selecting one or more of the results of the plurality of matching means based on a specific value. 5. The collating means according to claim 2, wherein the collating means is provided with a plurality of collating means corresponding to the discrete values, and all or a part of the plurality of collating means can be operated in parallel. A means for recognizing a plurality of matching results obtained from the plurality of matching means based on the discrete values. 6. The method according to claim 3, 4 or 5,
The analyzing means is a voice recognition device that outputs whether the utterance unit of the input voice is a syllable, a word, or a sentence. 7. The method according to claim 3, 4 or 5,
The speech recognition device is configured to extract discrete values related to speaker characteristics. 8. The output according to claim 1, wherein the output obtained from the second analysis means is a continuously varying amount, and the collating means and / or the judging means is used by using the continuously varying amount. A voice recognition device that is controlled. 9. The voice recognition device according to claim 8, wherein the second analysis means outputs a continuously changing amount related to a speaking rate.