JPH08171396A - Speech recognition device - Google Patents

Abstract

PURPOSE: To provide the speech recognition device which more surely corrects the error of phoneme recognition to obtain a higher speech recognition rate in comparison with a conventional example. CONSTITUTION: A collation part 5 uses phoneme HMM(hidden Markov model) to perform the phoneme collation of an input spoken speech and outputs a recognized phoneme sequence corresponding to the spoken speech and its speech section. At the time of learning, the recognized phoneme sequence and its speech section are compared with a phoneme sequence as a correct answer and its speech section; and if the recognized phoneme sequence is different from the phoneme sequence as the correct answer, a pair of the recognized erroneous phoneme sequence and the phoneme sequence as the correct answer are extracted as an error inclination table. If the recognized phoneme sequence is compared with erroneous phoneme sequences in the error inclination table based on the recognized phoneme sequence and its speech section to detect an erroneous phoneme sequence at the time of recognition, the recognized phoneme sequence is substituted with the phoneme sequence as the correct answer to correct the error. The phoneme sequence may be a state sequence.

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 for recognizing a voice that is input using a hidden Markov model (hereinafter referred to as HMM).

[0002]

2. Description of the Related Art In a conventional voice recognition apparatus, in order to reduce the uncertainty of acoustic matching between a standard model and an input voice, a large beam width is used for matching. , A method of narrowing down phoneme and word candidates from language restrictions, and multiple result candidates with high scores,
A method of narrowing down appropriate candidates by using context conditions is used. However, in reality, due to the limitation of the processing time and the processing capacity, there is a problem that a sufficient beam width and the number of recognition candidates cannot be secured, or even if they are secured, the number of candidates is too large to narrow down the correct answer.

Conventionally, a method of performing error correction by learning an error tendency of a symbol sequence of a recognized syllable with respect to a symbol sequence of an input syllable with the maximum length when the sequence length is 3 syllables (hereinafter referred to as the first conventional technique). Shinji Kamiya et al., “Correction of cut-out / recognition error based on appearance probability of syllable chain”, Proceedings of ASJ, 3-5-8, pp.
Proposed at 103-104, its effectiveness has been confirmed.

However, in the first conventional example, in the comparison on the symbol series, even if the actually matched vocal utterance sections are different, they are processed as corresponding portions on the symbol, so that the same error is corrected. There is no guarantee that you can. Further, limiting the maximum length of the sequence also has a problem that the sequence that can correct the same error may be cut off in the middle. Therefore, the correction efficiency during voice recognition is relatively low.

Further, a method of calculating the error rate and the error content of each phoneme at the time of learning in consideration of context conditions before and after (hereinafter referred to as a second conventional example) is Tanaka et al.
"Speeding up phrase detection in Japanese Dictation system", IEICE Technical Committee, SP90-70, p.
p. 17-24, 1990.
In the second conventional example, the following method is used. If there is a phoneme error, the conventional Confusion Matrix of the phoneme recognizer is used.
Since the error probability of each phoneme can be calculated from, the error probability can be calculated in advance for the set of three phonemes included in the dictionary, and the calculated error probability is used as the weighting coefficient of the recognition score. Correct the recognition result.

However, in the second conventional example, only the error within the three phonemes including the contexts before and after can be corrected. Therefore, also in the second conventional example, the correction efficiency at the time of voice recognition is relatively low.

An object of the present invention is to solve the above problems and to provide a speech recognition apparatus capable of more reliably correcting phonological recognition errors and obtaining a higher speech recognition rate as compared with the conventional example. To provide.

[0008]

According to a first aspect of the present invention, there is provided a speech recognition apparatus according to the present invention, wherein phonemes are collated by using a phoneme-hidden Markov model for the inputted uttered speech, and the corresponding uttered speech is recognized. Based on the collating means for outputting the phoneme sequence and its speech section information, based on the phoneme sequence and its speech section information which are the collation result by the above collating means for the correct phoneme sequence and the learning vocalization whose speech section is known. ,
The recognized phoneme sequence and its phoneme section information are compared with the correct phoneme series and its phoneme section, and in the phoneme sections corresponding to each other, when the recognized phoneme sequence is different from the correct phoneme series, An error extraction unit that extracts a pair of the recognized error phoneme sequence and the correct phoneme sequence, and a pair of the error phoneme sequence and the correct phoneme sequence extracted by the error extraction unit are stored as an error tendency table. Based on the phoneme sequence that is the matching result by the matching means and the voice section information for the input voicing speech to be recognized, and referring to the error tendency table stored by the storage means, When the recognized phoneme sequence is compared with the error phoneme sequence in the error tendency table, and the error phoneme sequence is detected in the recognized phoneme sequence. The recognized phoneme sequence is replaced with a correct phoneme sequence corresponding to the erroneous phoneme sequence to add a phoneme sequence that is error-corrected to the phoneme sequence that is the matching result by the matching means, and the phoneme recognition result is obtained. An error correction processing means for outputting as a candidate is provided.

According to a second aspect of the present invention, a speech recognition apparatus performs phoneme matching on an input uttered voice using a phoneme hidden Markov model, and a recognized state sequence corresponding to the uttered voice and its recognized state sequence. Based on the matching means that outputs the voice section information, the correct state series and the state series that is the result of the matching by the matching means with respect to the learning vocalization whose voice section is known and the voice section information, the recognition is performed. The state sequence and its voice section information are compared with the correct state sequence and its voice section, and when the recognized state sequence is different from the correct state sequence in the corresponding voice sections, the corresponding state sequence is recognized. The error tendency means for extracting a pair of the error state series and the correct state series, and the pair of the error state series and the correct state series extracted by the error extracting means as an error tendency. The error tendency table stored by the storage means based on the storage means for storing as a table, the state series as the result of the comparison by the matching means for the input uttered speech to be recognized, and the voice section information. Then, the recognized state series,
Compare with the error state series in the error tendency table,
When the error state sequence is detected in the recognized state sequence, the recognized state sequence is replaced with the correct state sequence corresponding to the error state sequence to obtain the error-corrected state sequence. An error correction processing means for outputting as a phoneme recognition result candidate in addition to the state series which is the matching result by the matching means.

The speech recognition apparatus according to a third aspect is the speech recognition apparatus according to the first or second aspect, in which the input uttered voice is composed of one sentence and is output from the error correction processing means. It is characterized by further comprising morpheme analysis means for performing morpheme analysis on the recognition result candidates with reference to a predetermined morpheme dictionary and outputting an optimum speech recognition result as one sentence.

Further, the voice recognition apparatus according to claim 4 is
3. The voice recognition device according to claim 1, wherein the input uttered voice is composed of one word, and a predetermined word dictionary is referred to for a phoneme recognition result candidate output from the error correction processing means. It is characterized by further comprising word analysis means for performing word analysis and outputting an optimum speech recognition result as one word.

The speech recognition apparatus according to claim 5 is the speech recognition apparatus according to claim 1, 3 or 4, wherein the error tendency table is a correct phoneme sequence and a state number of the phoneme hidden Markov model. It is characterized by including a pair of an erroneous phoneme sequence and a state number of the phoneme hidden Markov model. Furthermore, the speech recognition apparatus according to claim 6 is the speech recognition apparatus according to any one of claims 1 to 5,
The phoneme hidden Markov model is characterized in that one phoneme is a hidden Markov model composed of a plurality of states.

A speech recognition apparatus according to a seventh aspect is the speech recognition apparatus according to any one of the first to sixth aspects, wherein the speech section information is a speech section of an input uttered speech of a predetermined length. It is characterized in that it is represented by a starting frame number and an ending frame number where the phoneme starts among a plurality of frames separated by. Further, the voice recognition device according to claim 8 is dependent on claim 2 or 5.
In the speech recognition apparatus according to claim 7, the phoneme sequence and the correct phoneme sequence as a result of the matching by the matching means are each one phoneme and one phoneme, respectively.
It is characterized in that it is expressed in the form of a set of a state number, a starting frame number, and an ending frame number.

[0014]

In the speech recognition apparatus according to claim 1 configured as described above, at the time of learning, the error extracting means is the matching means with respect to the correct phonological sequence and the uttered speech for learning whose speech section is known. Based on the phoneme sequence and its speech section information which is the matching result by the, the recognized phoneme sequence and its speech section information are compared with the correct phoneme sequence and its speech section, and in the speech sections corresponding to each other, When the recognized phoneme sequence is different from the correct phoneme sequence,
A pair of the recognized error phoneme sequence and the correct phoneme sequence is extracted and stored in the storage means as an error tendency table. Further, at the time of recognition, the error correcting means stores the error tendency stored by the storing means on the basis of the phoneme sequence which is the result of the matching by the matching means with respect to the inputted uttered speech to be recognized, and the voice section information. By referring to a table, the recognized phoneme sequence is compared with the error phoneme sequence in the error tendency table, and when the error phoneme sequence is detected in the recognized phoneme sequence, the recognition The phonological sequence that has been error-corrected by replacing the phonological sequence that is generated with the correct phonological sequence corresponding to the erroneous phonological sequence is added to the phonological sequence that is the collation result by the collating means, and as a phonological recognition result candidate. Output.

Further, in the speech recognition apparatus according to the second aspect, at the time of learning, the error extraction means is a matching result by the matching means with respect to a training utterance whose correct state sequence and its voice section are known. Based on the state series and its voice section information, the recognized state series and its voice section information are compared with the correct state series and its voice section, and the recognized state series are shown in the corresponding voice sections. Is different from the correct answer state series, a pair of the recognized error state series and the correct answer state series is extracted and stored in the storage means as an error tendency table.
Further, at the time of recognition, the error correction means, based on the state series and its voice section information, which is the result of the matching by the matching means with respect to the input uttered voice to be recognized,
By referring to the error tendency table stored by the storage means, the recognized state series is compared with the error state series in the error tendency table, and the error state is included in the recognized state series. When the sequence is detected, the recognized state sequence is replaced with the correct state sequence corresponding to the error state sequence, and the error-corrected state sequence is converted into the state sequence which is the collation result by the collating means. In addition, it outputs as a phoneme recognition result candidate.

Further, in the speech recognition apparatus according to the third aspect, the morpheme analysis means includes the input uttered speech as one sentence and the phoneme recognition result candidate output from the error correction processing means is Then, the morpheme analysis is performed with reference to a predetermined morpheme dictionary, and the optimum speech recognition result is output as one sentence. Therefore, the voice recognition of the uttered voice sentence can be performed.

Further, in the speech recognition apparatus according to the fourth aspect, the word analysis means includes a single word in the input uttered voice, and the phoneme recognition result candidate output from the error correction processing means is supplied to the word recognition means. Then, the word analysis is performed with reference to a predetermined word dictionary, and the optimum speech recognition result is output as one word. Therefore, it is possible to voice-recognize the word of the uttered voice.

Further, in the speech recognition apparatus according to claim 5, the error tendency table is preferably such that the correct phoneme sequence and the state number of the phoneme hidden Markov model, the error phoneme sequence and the phoneme hidden Markov model thereof. Including a pair with the state number of. Further, in the speech recognition apparatus according to the sixth aspect, the phoneme hidden Markov model is preferably a hidden Markov model in which one phoneme is composed of a plurality of states.

Further, in the speech recognition apparatus according to the seventh aspect of the invention, the speech section information is preferably a phoneme of a plurality of frames obtained by dividing the input uttered speech into frame sections of a predetermined length. It is represented by the starting frame number and the ending frame number. Further, in the voice recognition device according to claim 8, the phoneme sequence and the correct phoneme sequence that are the matching results by the matching means are preferably, for one phoneme, the phoneme and the state number, respectively. It is expressed in the form of a set of a starting frame number and an ending frame number.

[0020]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a voice recognition device according to an embodiment of the present invention. This voice recognition device (a) feature extraction for extracting a predetermined acoustic feature parameter from a digital voice signal of a voiced voice which is A / D converted by an A / D converter 2 after being input to a microphone 1. Based on the feature parameters input from the unit 3 and (b) the feature extraction unit 3 via the buffer memory 4, phoneme matching is performed using the phoneme HMM stored in the phoneme HMM memory 10 to correspond to the vocalized voice. The collation result storage buffer memory (hereinafter,
The storage buffer memory. ) 11 stored in the acoustic parameter collation unit (hereinafter referred to as collation unit) 5, and (c) a phoneme sequence which is a collation result by the collation unit 5 with respect to a correct phoneme sequence and a learning vocalization whose voice section is known. And the phoneme section information thereof, the recognized phoneme sequence and the phoneme section information thereof are compared with the correct phoneme series and the phoneme section thereof, and the recognized phoneme series are correctly answered in the corresponding phoneme sections. Recognition error phoneme sequence error extraction unit (hereinafter referred to as error extraction unit) 6 for extracting a pair of the recognized error phoneme sequence and the correct phoneme sequence when the phoneme sequence is different from the phoneme sequence of ) An error tendency table memory 1 for storing a pair of an error phoneme sequence and a correct phoneme sequence extracted by the error extraction unit 6 as an error tendency table.
2 and (e) the error tendency table stored in the error tendency table memory 12 is referred to based on the phoneme sequence which is the result of the matching by the matching unit 5 with respect to the input uttered speech to be recognized and its voice section information. Then, the recognized phoneme sequence is compared with the error phoneme sequence in the error tendency table, and when the error phoneme sequence is detected in the recognized phoneme sequence, the recognized phoneme sequence is detected. A result of adding a phoneme sequence that has been error-corrected by replacing the sequence to a correct phoneme sequence corresponding to the error phoneme sequence to the phoneme sequence that is the matching result by the matching unit 5 and outputting as a phoneme recognition result candidate. A candidate error correction processing unit (hereinafter referred to as an error correction processing unit) 7 and (f) when the input uttered voice is composed of one sentence, is output from the error correction processing unit 7. The phoneme recognition result candidate is subjected to morphological analysis with reference to the morpheme dictionary in the morpheme dictionary memory 13 and the N-gram dictionary in the N-gram dictionary memory 14, and the optimum speech recognition result as one sentence. And a morphological analysis unit 8 for outputting

The voice recognition apparatus of this embodiment has a learning mode for extracting an error tendency table and a recognition mode for recognizing voiced speech input from a microphone. Of the processing units 1 to 8, the extraction unit 6 operates only in the learning mode, and the error correction processing unit 7 and the morphological analysis unit 8 operate only in the recognition mode.

In FIG. 1, the voice uttered by the speaker is input to the microphone 1 and converted into a voice signal, and then input to the A / D converter 2, and the frame period is, for example, 10 milliseconds. To digital audio signal from A / D
To be converted. At this time, a frame number, which is a serial number from the start of input, is given to each frame divided into frame sections of a predetermined length of the digital voice signal, and the digital voice signal is input to the feature extraction unit 3. The feature extraction unit 3 performs, for example, LPC analysis on the input digital audio signal, and extracts 34-dimensional feature parameters including logarithmic power, 16th-order cepstrum coefficient, Δlogarithmic power, and 16th-order Δcepstrum coefficient. . The time series of the extracted characteristic parameters is input to the matching unit 5 via the buffer memory 4.

The HMM in the HMM memory 10 connected to the collation unit 5 is composed of a plurality of states and arcs indicating transitions between the states. Each arc has a transition probability between states and an output probability for an input code. have. In the present embodiment, the phoneme HMM is a HMM in which the features of each phoneme are expressed in three states per phoneme, and is a continuous HMM having three continuous states and four loops. In the present embodiment, the HMM expressed in three states per phoneme is used, but the present invention is not limited to this, and an HMM expressed in a plurality of states per phoneme may be used.

The collation unit 4 calculates the distance from each phoneme HMM to the input characteristic parameter data in time series, and obtains the distance less than a predetermined distance as the matching result. Phoneme matching, that is, DP (dynami
C programming) matching (so-called DP matching) is performed, and the phoneme sequence corresponding to the uttered speech and its speech section information are output to the storage buffer memory 11 and stored. In the present embodiment, the collation unit 4 uses, for example, the well-known frame synchronization type One Pass DP method for Vi.
By performing matching with the terbi algorithm, phonological sequence candidates from the most probable phonological sequence up to the n-th candidate in each frame along the time series are output, and when the processing reaches the final frame, all The phonological recognition result candidates of the input uttered speech of are output.

In the learning buffer mode, the storage buffer memory 11 compares the phoneme sequence which is the recognition result obtained by the collation with the H sequence obtained by the collation.
The state number of the MM and the matching section of each recognized phoneme, that is, the corresponding speech section information are stored and stored. On the other hand, at the time of recognition, a phoneme sequence that is the recognition result obtained by collation,
Stores and stores the HMM status number.

For example, when recognition is performed using an HMM in which one phoneme is composed of three states and "koNbaNwa (good evening)" is recognized as "kaNbaNna", the storage buffer memory 11 stores the following As shown in Table 1, a phoneme sequence as a result of phoneme recognition, an HMM state number, and a start frame number and an end frame number that are corresponding voice section information are stored. Here, as the state number of the HMM, three independent numbers are given to each phoneme at the time when the phoneme HMM is created.
Three state numbers corresponding to one phoneme are stored. In the table below, the starting frame number is referred to as the starting FN, and the ending frame number is referred to as the ending FN.

[0027]

[Table 1] ─────────────────────────── Recognized phoneme sequence State number Start FN End FN ────────── ───────────────── k [1,2,3] 1 20 a [13,14,15] 21 35 N [7,8,9] 36 45 b [10 , 11, 12] 46 50 a [13, 14, 15] 51 65 N [7, 8, 9] 66 80 n [19, 20, 21] 81 88 a [13, 14, 15] 89 100 ── ────────────────────────

In the present embodiment, the voice section information is represented by the start frame number and the end frame number at which the phoneme starts among a plurality of frames obtained by dividing the input uttered voice into frame sections of a predetermined length. Has been done. In addition, the recognized phoneme sequence and the correct phoneme sequence described below are each a phoneme, a state number, and
It is expressed in the form of a set of a starting frame number and an ending frame number.

In the learning mode, the error extraction unit 6 is based on the phoneme sequence which is the matching result of the matching unit 5 with respect to the correct phoneme sequence and the learning vocalization whose voice segment is known, and its voice segment information. Then, the recognized phoneme sequence and its speech segment information are compared with the correct phoneme sequence and its speech segment, and when the recognized phoneme sequence is different from the correct phoneme sequence in mutually corresponding speech segments. Then, a pair of the recognized error phoneme sequence and the correct phoneme sequence is extracted. Then, the pair of the extracted error phoneme sequence and correct phoneme sequence is stored in the error tendency table memory 12 as an error tendency table together with the state number of the HMM.

The error tendency table stores the correct phoneme sequence and the error phoneme sequence extracted by the above-mentioned processing of the error extracting unit 6 together with the HMM state number. Here, for example, the correct phoneme sequence is “koNbaNwa”, and the data stored in advance in the storage buffer memory 11 before the learning mode processing is, as shown in Table 2, the correct phoneme sequence, Consider a case where the state number of the HMM and the voice section information are stored in a format.

[0031]

[Table 2] ─────────────────────────── Correct answer phonological sequence State number Start FN End FN ─────────── ───────────────── k [1,2,3] 1 20 o [4,5,6] 21 35 N [7,8,9] 36 45 b [10 , 11, 12] 46 50 a [13, 14, 15] 51 65 N [7, 8, 9] 66 80 w [16, 17, 18] 81 88 a [13, 14, 15] 89 100 ─── ────────────────────────

Further, in the recognition mode, when the operator utters "koNbaNwa" into the microphone 1, the collation unit 5 produces a collation result in the same format as the previous example, as shown in Table 3 below. It is assumed that the data is stored and stored in the storage buffer memory 11.

[0033]

[Table 3] ─────────────────────────── Recognition phoneme sequence State number Start FN End FN ─────────── ───────────────── k [1,2,3] 1 20 a [13,14,15] 21 35 N [7,8,9] 36 45 b [10 , 11, 12] 46 50 a [13, 14, 15] 51 65 N [7, 8, 9] 66 80 n [19, 20, 21] 81 88 a [13, 14, 15] 89 100 ── ────────────────────────

These correct phoneme sequences in Table 2 and Table 3
When the extraction unit 6 compares the recognized phoneme sequence of No. 1 with the recognized phoneme sequence, the extracted phoneme sequence is extracted as an error phoneme sequence and stored in the error tendency table memory 12. Here, the error tendency table includes the correct phoneme sequence, the state number of the phoneme HMM, the error phoneme sequence, and the phoneme HMM.
Including a pair with the state number of.

[0035]

[Table 4] ──────────────────────────────── Correct phoneme sequence and its state number → Error phoneme sequence and its state Number ──────────────────────────────── o, N, [4,5,6,7,8,9] → a, N, [13,14,15,7,8,9] w, [9,10,11] → n, [19,20,21] ────────────── ────────────────────

Further, the error correction processing section 7 refers to the error tendency table on the basis of the phoneme sequence and its speech section information which are the matching results by the matching section 5 with respect to the inputted uttered speech to be recognized. The recognized phoneme sequence is compared with the error phoneme sequence in the error tendency table, and when the error phoneme sequence is detected in the recognized phoneme sequence, the recognized phoneme sequence is The phoneme sequence that has been error-corrected by replacing it with the correct phoneme sequence corresponding to the error phoneme sequence is added to the phoneme sequence that is the matching result by the matching means, and is output to the morphological analysis unit 8 as a phoneme recognition result candidate. To do.

For example, when the error tendency table is created as shown in Table 4 during learning, and the phoneme sequence of the matching result by the matching unit 5 is "kaNnichina", the error phoneme sequence in the error tendency table is obtained. Since the phoneme sequence “an” and the phoneme “n”, which are the same phoneme sequences as the above, are included in the collation result, the error correction processing unit 7 refers to the error tendency table and refers to these phoneme sequences. Alternatively, the candidates shown in Table 5 below in which the phonemes are replaced with correct phoneme sequences are also added to the matching result, that is, the phoneme recognition result.

[0038]

[Table 5] ────────────── “koNnichina” “kaNnichiwa” “koNnichiwa” ───────────────

Further, the morpheme analysis unit 8 applies the morpheme in the morpheme dictionary memory 13 to the phoneme recognition result candidate output from the error correction processing unit 7 when the input uttered voice consists of one sentence. Morphological analysis is performed with reference to the dictionary and the N-gram dictionary in the N-gram dictionary memory 14, and an optimum speech recognition result is output as one sentence. That is, the morpheme analysis unit 8 first converts the plurality of recognition result candidates represented by the phoneme sequence into the Hiragana notation.
Next, in order to determine whether or not the result written in Hiragana is valid as a sentence, each recognition result candidate is referred to by referring to the morpheme dictionary that stores the word to be recognized written in Hiragana and its part-of-speech name. On the other hand, morphological analysis is performed. Multiple morphological analysis results are possible for one sentence input,
At that time, the probabilities of occurrence of adjacent words or parts of speech are checked in advance, and these are stored in the N-gram dictionary memory 14 as an N-gram dictionary. The morphological analysis unit 8 uses the N-
The gram dictionary is referred to and the sentence having the highest possibility of the word or part of speech that is most adjacent among the plurality of morphological analysis results is output as the most probable morphological analysis result. On the other hand, when an incorrect sentence is input to the morpheme analysis unit 8, when there is no corresponding word in the morpheme dictionary, or when there is a word or part of speech that has no possibility of being adjacent, the analysis fails in the middle. .

For example, consider the case where the following four recognition result candidates are input to the morphological analysis unit 8 as the recognition results output from the error correction processing unit 7. (A1) "kaNinichina" (A2) "koNnichina" (A3) "kaNnichiwa" (A4) "koNnichiwa" At this time, recognition result candidates (A1), (A2), and (A)
In 4), the morpheme analysis unit 8 fails in the analysis, and the morpheme analysis unit 8 outputs the analysis result that only the recognition result candidate (A4) is "Konichiwa (verb)".

Further, the result of the simulation performed by the present inventor using the speech recognition apparatus of FIG. 1 will be described below. Inputs to the morphological analysis unit 8 in the speech recognition apparatus including the error correction processing unit 7 of the present invention are a plurality of recognition result candidates including a result that seems to be correct. In order to evaluate this error correction method, how efficiently a correct phoneme sequence is included in a plurality of recognition result candidates input to the morphological analysis unit 8 was examined. As a result, 1.33 times as many phonological sequence candidates were included in the same number of candidates as in the case where the top n candidates of the recognition result were input as morphological analysis without error correction. In addition, compared to the first conventional example,
For the same number of candidates, 1.24 times the correct phoneme sequence was included. From the above, it can be seen that the error correction method of the present invention has better correction efficiency than the conventional method. This makes it possible to more reliably correct phonological recognition errors,
It is possible to obtain a higher voice recognition rate as compared with the conventional example.

As described above, according to this embodiment, the error tendency can be extracted without depending on the unit of HMM, and as a result, the correction efficiency is higher than that of the conventional example.
In addition, since the recognition result can be corrected using the error tendency table, it is possible to correct a word or a sentence different from the error learning with high accuracy. Therefore, an error in phoneme recognition can be corrected more reliably, and a higher speech recognition rate can be obtained as compared with the conventional example.

In the above embodiment, the morphological analysis unit 8
However, the present invention is not limited to this, and a word analysis unit may be provided instead of this. When the input uttered voice consists of one word, the word analysis unit applies to the phoneme recognition result candidates output from the error correction processing unit 7,
The word analysis is performed with reference to a predetermined word dictionary, and the optimum speech recognition result is output as one word.

In the above embodiment, the case where one phoneme corresponds to one set of three state numbers is shown, but the present invention is not limited to this, and the recognition result sequence, The correct sequence and the error sequence may be represented by a state sequence represented by HMM state numbers or state symbols. That is, in order to perform the speech recognition more reliably, there are cases where the same phoneme can give different states depending on the phoneme environment before and after each phoneme. For example, the phoneme "N" may be represented by a sequence of state numbers [7, 8, 9] or a sequence of state numbers [7, 32, 33] different from this. In order to deal with this, the series of recognition results, the series of correct answers, and the series of errors are represented by a state series represented by an HMM state number or state symbol. In this modified example, since the phoneme is expressed in a more detailed format, the correction efficiency can be improved and the speech recognition can be performed more reliably as compared with the device according to claim 1 which uses the phoneme sequence. it can.
In the case of this modified example represented by a sequence of HMM state numbers, there is no phoneme sequence in Tables 1 to 4, but only a sequence of state numbers. Even in this case, the phoneme HMM is
One phoneme may be represented by a plurality of states other than three.

[0045]

As described above in detail, according to the speech recognition apparatus of the first aspect of the present invention, the input uttered voice is phoneme-matched by using the phoneme hidden Markov model to correspond to the uttered voice. Collating means for outputting the recognized phoneme sequence and its speech section information, and a phoneme sequence and its speech section information which are the collation result by the collating means for the correct phoneme sequence and the learning vocalization whose speech section is known. Based on the above, the recognized phoneme sequence and its speech section information are compared with the correct phoneme sequence and its phoneme section, and the recognized phoneme sequence is different from the correct phoneme series in mutually corresponding phoneme sections. Error extracting means for extracting a pair of the recognized error phoneme sequence and the correct phoneme sequence, and the error phoneme sequence and correct phoneme system extracted by the error extracting means. And a pair of and are stored as an error tendency table, and are stored by the storage means based on the phoneme sequence as a result of the matching by the matching means with respect to the input uttered speech to be recognized, and its speech section information. When the detected phoneme sequence is detected in the recognized phoneme sequence by comparing the recognized phoneme sequence with the error phoneme sequence in the error tendency table by referring to the recognized error tendency table Further, the recognized phoneme sequence is replaced with a correct phoneme sequence corresponding to the error phoneme sequence to add a phoneme sequence that is error-corrected to the phoneme sequence that is the matching result by the matching means, Error correction processing means for outputting as a recognition result candidate.

Therefore, the error tendency can be extracted without depending on the unit of the HMM, and as a result, the correction efficiency is higher than that of the conventional example and the accuracy is high even in the words and sentences different from the error learning. Correction is possible. Therefore, an error in phoneme recognition can be corrected more reliably, and a higher speech recognition rate can be obtained as compared with the conventional example.

According to the second aspect of the speech recognition apparatus of the present invention, the input uttered voice is phoneme-matched using the phoneme hidden Markov model, and the recognized state sequence corresponding to the uttered voice is obtained. And a matching means for outputting the voice section information thereof, based on the state series and its voice section information which are the result of the matching by the above-mentioned matching means with respect to the correct state series and the learning vocalization of which the voice section is known. The recognized state series and its voice section information are compared with the correct state series and its voice section, and when the recognized state series is different from the correct state series in the corresponding voice sections, An error extraction unit that extracts a pair of the recognized error state sequence and the correct state sequence, and a pair of the error state sequence and the correct state sequence extracted by the error extraction unit are A storage unit for storing the tendency table, and an error tendency table stored by the storage unit based on the state series and the voice section information which are the collation result by the collating unit for the input uttered voice to be recognized. With reference, the recognized state sequence is compared with the error state sequence in the error tendency table, and when the error state sequence is detected in the recognized state sequence, the recognized state sequence is recognized. The state sequence corrected for error by replacing the state sequence with the correct state sequence corresponding to the error state sequence is added to the state sequence that is the collation result by the collating means, and is output as a phoneme recognition result candidate. Error correction processing means.

Therefore, the error tendency can be extracted without depending on the unit of HMM, and as a result, the correction efficiency is higher than that of the conventional example, and the accuracy is high even in the words and sentences different from the error learning. Correction is possible. Therefore, an error in phoneme recognition can be corrected more reliably, and a higher speech recognition rate can be obtained as compared with the conventional example. Further, in this case, since the phoneme is expressed in a more detailed format, the correction efficiency can be improved and the voice recognition can be performed more reliably as compared with the voice recognition device according to claim 1 which uses the phoneme sequence. be able to.

According to the speech recognition apparatus of the third aspect, the inputted uttered speech is composed of one sentence, and the morpheme recognition result candidate output from the error correction processing means has a predetermined morpheme. The apparatus further includes a morphological analysis unit that performs morphological analysis with reference to the dictionary and outputs an optimal speech recognition result as one sentence. As a result, the voice recognition of the uttered voice sentence can be executed with a higher voice recognition rate than the conventional example.

Further, according to the voice recognition device of the fourth aspect, the inputted uttered voice consists of one word,
The phoneme recognition result candidate output from the error correction processing means is further provided with word analysis means for performing word analysis by referring to a predetermined word dictionary and outputting an optimum speech recognition result as one word. As a result, it is possible to perform voice recognition of a voiced voice word with a higher voice recognition rate than in the conventional example.

[Brief description of drawings]

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

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... A / D converter, 3 ... Feature extraction part, 4 ... Buffer memory, 5 ... Acoustic parameter collation part, 6 ... Recognition error phoneme sequence extraction part, 7 ... Result candidate error correction processing part, 8 ... Morphological analysis unit, 10 ... Phoneme HMM memory 11 ... Collation result storage buffer memory, 12 ... Error tendency table memory, 13 ... Morphological dictionary memory, 14 ... N-gram dictionary memory.

Front page continued (72) Inventor Harald Singer, 5 Seiraya, Seika-cho, Seika-cho, Soraku-gun, Kyoto Pref. A / R Co., Ltd., Speech Translation Laboratory Machidaiji, Inui-ya, small, Mihiratani No. 5 Address, T.R. Co., Ltd.

Claims (8)

[Claims] 1. A matching means for performing phoneme matching on an input uttered voice using a phoneme hidden Markov model, and outputting a recognized phoneme sequence corresponding to the uttered voice and its speech section information, and a correct phoneme sequence. Based on the phoneme sequence that is the result of the matching by the matching means with respect to the learning vocalization whose voice segment is known and its voice segment information, the recognized phoneme sequence and its voice segment information are converted into the correct phoneme. When the recognized phoneme sequence is different from the correct phoneme sequence in the corresponding voice sections, the pair of the recognized error phoneme sequence and the correct phoneme sequence is compared. Error extracting means for extracting; a memory means for storing a pair of the error phoneme sequence and the correct phoneme sequence extracted by the error extracting means as an error tendency table; Based on the phoneme sequence that is the result of matching by the matching means with respect to the inputted voice and the voice section information thereof, the error tendency table stored by the storage means is referred to, and the recognized phoneme sequence, When the error phoneme sequence in the error tendency table is compared to detect the error phoneme sequence in the recognized phoneme sequence, the recognized phoneme sequence is set to the correct answer corresponding to the error phoneme sequence. Error correction processing means for adding a phoneme sequence that has been error-corrected by replacing it to the phoneme sequence that is the matching result by the matching means and outputting it as a phoneme recognition result candidate. Voice recognition device. 2. A collating means for performing phoneme matching on the input uttered speech using a phonological hidden Markov model, and outputting a recognized state series corresponding to the uttered speech and its speech section information, and a correct state series. Based on the state series and its voice section information, which is the result of the matching by the above-mentioned matching means for the voiced speech for learning whose voice section is already known, the recognized state series and its voice section information are set to the correct state. When the recognized state sequence is different from the correct state sequence in the corresponding speech sections, the pair of the recognized error state sequence and the correct state sequence is compared. Error extracting means for extracting; a memory means for storing a pair of the error state sequence and the correct state sequence extracted by the error extracting means as an error tendency table; Based on the state sequence and its voice section information which is the result of the collation by the collating means with respect to the input uttered voice, the error tendency table stored by the storage means is referred to, and the recognized state sequence, When the error state sequence in the recognized state sequence is detected by comparing with the error state sequence in the error tendency table, the recognized state sequence is set to the correct answer corresponding to the error state sequence. Error correction processing means for adding a state series that has been error-corrected by replacing the state series as a result of the verification to the state series that is the verification result by the verification means, and outputting the result as a phoneme recognition result candidate. Voice recognition device. 3. The input uttered voice is composed of one sentence, and a morphological analysis is performed on a phoneme recognition result candidate output from the error correction processing means by referring to a predetermined morpheme dictionary, and 1 The speech recognition apparatus according to claim 1 or 2, further comprising a morphological analysis unit that outputs an optimal speech recognition result as one sentence. 4. The input uttered voice consists of one word, and the phoneme recognition result candidates output from the error correction processing means are subjected to word analysis by referring to a predetermined word dictionary, and 3. The voice recognition device according to claim 1, further comprising a word analysis unit that outputs an optimum voice recognition result as one word. 5. The error tendency table includes a pair of a correct phoneme sequence and a state number of the phoneme hidden Markov model, and a pair of an error phoneme sequence and a state number of the phoneme hidden Markov model. The voice recognition device according to Item 1, 3 or 4. 6. The speech recognition apparatus according to claim 1, wherein the phoneme hidden Markov model is a hidden Markov model in which one phoneme is composed of a plurality of states. 7. The voice section information is represented by a start frame number and a end frame number at which the phoneme starts among a plurality of frames obtained by dividing the input uttered voice into frame sections of a predetermined length. The voice recognition device according to claim 1, wherein the voice recognition device is a voice recognition device. 8. The phoneme sequence and the correct phoneme sequence as a result of the matching by the matching means are a set of a phoneme, a state number, a starting frame number, and an ending frame number for one phoneme, respectively. Claim 7 dependent on Claim 2 or 5 characterized by being represented in the form of
The voice recognition device according to claim 7.