JP2999726B2 - Continuous speech recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a continuous speech recognition apparatus for continuously recognizing speech based on a speech signal of an input speech sentence.

[0002]

2. Description of the Related Art Conventionally, the present applicant has been developing a continuous speech recognition system (hereinafter referred to as a first conventional example) for the purpose of speech recognition of spontaneous utterances (for example, see the prior art document). 1 “Nagai, Takami, Sagayama,“ The SSS-LR Conti
nuous Speech Recognition System: Integrating SSS-D
erivrd Allopohne Models and a Phoneme-Context-Depe
ndent LR Parser ", Proc. of ICSLP92, pp. 1511-1514, 1992
Year "and prior art document 2" Shimizu, Monzen, Singer, Mats
unaga, “Time-Synchronous Continuous Speech Recogni
zer Driven by a Context-Free Grammar ", Proc.of ICAS
SP95, pp. 584-587, 1995 ". ). In this first conventional example, a phoneme hidden Markov model (hereinafter, a hidden Markov model is referred to as H
It is called MM. ) And a word dictionary, while performing speech recognition while managing the history and grammatical state of words from the start of utterance.

On the other hand, a speech recognition method using a word graph (hereinafter referred to as a second conventional example) is disclosed in Prior Art Document 3.
“Ney, Aubert,“ A Word Graph Algorithm for Large Vo
cabulary, Continuous Speech Recognition ", Proc.of I
CSLP94, pp.1355-1358, 1994 "and prior art document 4" Wo
odland, Leggetter, Odell, Valtchev, Young, “The 1994 H
TKLarge Vocabulary Speech Recognition System ", Pro
c. of ICASSP95, pp.73-76, 1995 ”.

The main idea of the word graph of the second conventional example is to process word hypothesis candidates in a region of a speech signal where ambiguity in speech recognition is relatively high. The advantage is that pure speech recognition is separate from the language model application, and that complex language models can be applied to known steps following the word currently being recognized. The number of word hypothesis candidates needs to change according to the level of ambiguity in speech recognition. The difficulties in constructing a good word graph efficiently are: The start time of a word is
Generally, it depends on the preceding word. In the first approximation, a so-called word pair approximation method as described below is obtained by restricting this dependency to the immediately preceding preceding word. That is, given a word pair and its end time, the word boundary between two words is independent of another preceding word. This word-pair approximation method was originally introduced to efficiently calculate a plurality of sentences or n best sentences. This word graph is expected to be more efficient than the approach of obtaining the n bests (hereinafter referred to as the n best method). In the method using this word graph, it is necessary to generate a plurality of word hypotheses only locally, while in the n-best method, each local word hypothesis candidate is a list of n best sentences. Need a whole sentence to add to.

However, in the first conventional example,
Since it is necessary to manage the history and grammatical state of words from the start of utterance, if it is used for recognition of natural utterances in which interjections are inserted, stagnation, and rephrasing frequently, calculations required for merging or dividing word hypotheses There was a problem that the cost was extremely large. In other words, the amount of processing required for speech recognition is large, and a storage device having a relatively large storage capacity is required. On the other hand, there is a problem that the processing time is long because the amount of processing is large.

In the second conventional word pair approximation method, one hypothesis is represented for each preceding word.
The approximation effect is still relatively small. For this reason, the same problem as the first conventional example occurs.

[0007] In order to solve the above problems, the present applicant has filed a patent application of Japanese Patent Application No. 7-234043.
"By detecting the word hypothesis of the uttered speech sentence based on the speech signal of the input uttered speech sentence and calculating the likelihood,
In a continuous speech recognition apparatus provided with a speech recognition means for continuously recognizing a speech, the speech recognition means performs, for a word hypothesis of the same word having a same end time and a different start time, for each head phoneme environment of the word. Continuous speech recognition characterized by narrowing down word hypotheses so as to be represented by one word hypothesis having the highest likelihood among the calculated total likelihoods from the utterance start time to the end time of the word. apparatus. (Hereinafter referred to as a third conventional example).

However, in the time-synchronous beam search in the continuous speech recognition apparatus as in the third conventional example, when a certain likelihood width is used as the threshold from the maximum likelihood candidates, the acoustic likelihood and the language likelihood are not considered. There was a problem with local fluctuations over time. In order to absorb local fluctuations with respect to time, it is necessary to widen the beam width or look ahead for likelihood. However, a wide beam width directly leads to an increase in the amount of calculation required for search, and there is a problem that the look-ahead of likelihood may complicate the algorithm or, in some cases, increase the amount of calculation in the look-ahead. . Hereinafter, this problem will be described in detail.

For example, the time synchronization beam search in the third conventional example has the following problems. (A) The likelihood of an unsearched portion is unknown: High likelihood up to a certain time t does not guarantee that the likelihood of the entire sentence is high. (B) In the intermediate state of a phoneme, the same pruning is performed as at the end of the phoneme: The training of the acoustic model is learned so that the likelihood of one phoneme or a phoneme sequence is maximized. Therefore, it is not guaranteed whether the likelihood is maximized in the intermediate state of phonemes. However, when a beam having a constant width is used, the intermediate state and the end of the phoneme are pruned under the same conditions. (C) Insertion of phonemes (words) with a short duration: phonemes with a short stay time and low likelihood contribute little to the accumulated likelihood. The cumulative likelihood is dominated by phonemes with long stay times and high likelihood. As a result, insertion of “phonemes with a short stay time and low likelihood”, which have little contribution to the accumulated likelihood, frequently occurs.

[0010] As means for solving these problems, a technique disclosed in prior art document 5 “Ney et al.,“ An overview of the phil
ips research system for large vocabulary continuou
s speech recognition ", International Journal of Patt
ern Recognition and Artificial Intelligence, Vol. 8,
No. 1, pp. 58-59, 1994 ", a method of pre-reading the likelihood (phoneme look-ahead) (hereinafter referred to as a fourth conventional example) is proposed. In the fourth conventional example, stricter pruning is performed at a phoneme boundary than at an intermediate state of phonemes in consideration of the likelihood of an unsearched portion for one phoneme. That is, the acoustic likelihood of one phoneme is separately calculated in advance, and when the end of the phoneme is reached, the second beam search is performed in consideration of the likelihood of the subsequent one phoneme.

However, in the fourth conventional example,
Limited look-ahead length and language constraints. That is, if the amount of calculation in the pre-reading is too large, the pruning effect based on the pre-reading is offset. For this reason, there is a problem that the time width of the pre-reading is limited to several frames, and the language constraint is limited to a very simple constraint.

An object of the present invention is to solve the above problems, to narrow down word hypotheses with a narrower beam width than in the conventional example, and to perform continuous speech recognition of spontaneous utterance at a smaller calculation cost. It is to provide a continuous speech recognition device which can perform the following.

[0013]

According to a first aspect of the present invention, there is provided a continuous speech recognition apparatus for detecting a word hypothesis of an uttered speech sentence based on an input speech signal of the uttered speech sentence and calculating an acoustic likelihood. By calculating, in the continuous speech recognition device provided with a speech recognition means for continuously recognizing speech, the speech recognition means sets the acoustic likelihood of each phoneme of the word in the time direction at a higher value than that of the center. The sound likelihood of the word hypothesis is corrected by delaying to move to the delayed time.

According to a second aspect of the present invention, there is provided the continuous speech recognition apparatus according to the first aspect, wherein the speech recognition means is adapted to detect a word hypothesis of the same word having the same end time and different start time. For each head phoneme environment of the word, one word hypothesis having the highest total likelihood among the total likelihoods including the acoustic likelihood calculated from the utterance start time to the end time of the word is represented. Thus, the word hypothesis is narrowed down.

[0015]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows a block diagram of a continuous speech recognition apparatus according to an embodiment of the present invention. The continuous speech recognition apparatus according to the present embodiment detects the word hypothesis of the uttered speech sentence based on the feature parameter of the speech signal of the input uttered speech sentence using a known one-pass Viterbi decoding method, and In the continuous speech recognition device including the word matching unit 4 that calculates and outputs the likelihood, the time direction of each phoneme of the word with respect to the word hypothesis output from the word matching unit 4 via the buffer memory 5 And a likelihood correction unit 7 that delays the peak of the acoustic likelihood in the central part of the word hypothesis so as to move to a time delayed from the central part, and corrects the acoustic likelihood of the word hypothesis. 7 based on the word hypothesis having the total likelihood including the acoustic likelihood output from the utterance start time from the utterance start time to the end time of the word for each head phoneme environment of the word. One word with the highest likelihood Characterized by comprising a word hypotheses narrowing-down unit 6 that performs narrowing of the word hypotheses to be represented by the theory.

The likelihood correction of the likelihood correction unit 7 used in the embodiment according to the present invention is performed by a delayed decision.
Beam search. The beam search for this delay determination is performed by delaying the evaluation of the likelihood of the path that has already been searched, without relying on likelihood look-ahead and the likelihood mapping by a nonlinear function as in the fourth conventional example. Address local variations in likelihood. In the following calculation, the likelihood indicates the log likelihood. In the present embodiment, each code is defined as follows. (A) t: time; (b) S: path of beam search; (c) q _A (S, t): path S, acoustic likelihood at time t; (d) Q _A (S, t): path S, cumulative acoustic likelihood from beginning of a sentence at time _{t; (e) Q L (} S, t): path S, the cumulative language likelihood from beginning of a sentence at time t.

Here, the acoustic likelihood is the likelihood calculated by the word matching unit 4 with reference to the phoneme HMM in the phoneme HMM memory 11, and the language likelihood is calculated by the statistical language model in the word matching unit 4. The likelihood is calculated with reference to the language model in the memory 13. When defined as described above, generally, the cumulative acoustic likelihood is obtained by adding the acoustic likelihood for each frame by the following equation.

[0018]

## EQU1 ## Q _A (S, t) = Q _A (S, t-1) + q
_A (S, t)

The cumulative total likelihood Q _all (S, t) from the head of the sentence used for beam search is the acoustic likelihood Q _A (S,
t) and the language likelihood Q _L (S, t) are calculated by the following equation.

[0020]

## EQU2 ## Q _all (S, t) = Q _A (S, t) + α · Q
_L (S, t)

Here, the constant α is a weight coefficient for the sound likelihood of the language likelihood, and in a preferred embodiment, α
= 4.5. In the beam search for delay determination in the present embodiment, as shown in the following equation,
Using the _A (S, t) likelihood Q _A '(S, t) obtained by subtracting the Q _A (S, t) delayed from the acoustic likelihood Q _Ad (S, t) instead of. That is, at time t-1, FIG.
As shown _{in, Q A (S, t-} 1) Q A (S instead of,
The likelihood Q _A ′ (S, t−1) obtained by subtracting the delayed acoustic likelihood Q _Ad (S, t−1) from t−1) is used.

[0022]

## EQU3 ## Q _A '(S, t) = Q _A (S, t) -Q _Ad (S,
t)

Here, the likelihood Q of the second term on the right side of the above equation (3)
_Ad (S, t) is calculated by the following equation.

[0024]

D = Q _Ad (S, t-1) + q _A (S, t)

## _EQU5 ## Q _Ad (S, t) = F (D) .D

When the above equation (3) is rewritten, when the above equation (1) is rewritten, the following equation is obtained.

[0026]

## EQU6 ## Q _A '(S, t) = Q _A ' (S, t-1) +
q _A '(S, t)

Here, the likelihood q _A ′ (S, t) is determined by the following equation.

[0028]

Equation 7] _{q A '(S, t)} = f (x) = f (q A (S, t) + Q A (S, t-1) -Q A'
(S, t-1)

Here, ΔQ _A (S, t−
1) -Q _A '(S, t-1)} is Q _Ad (S, t-1), which is an underestimate one time before compared to the third conventional example, and this data is Are sequentially stored in the underestimated likelihood memory 14 connected to the likelihood correction unit 7 and are used to correct the acoustic likelihood at the next time t to calculate the total likelihood. Therefore, in the present embodiment, the likelihood correction unit 7
At time (t-1), for each word hypothesis, ΔQ
Calculate the _{A (S, t-1)} -Q A '(S, t-1)}, and stored in underestimated likelihood memory 14, then, at time t, and the number 6 and the number 7 To calculate the acoustic likelihood Q _A ′ (S, t) corrected to be underestimated, and then using the following equation 8 obtained by rewriting equation 2 to obtain the total likelihood that is the cumulative likelihood. The degree Q ′ _all (S, t) is calculated, and the word hypothesis having the calculated total likelihood Q ′ _all (S, t) is output to the word hypothesis narrowing unit 6 via the buffer memory 5.

[0030]

Q ′ _all (S, t) = Q _A ′ (S, t) + α · Q
_L (S, t)

In the above equation (7), the function f (x)
Is a first function for calculating the delay ratio with respect to the likelihood x, an example of which is shown in FIG. As is evident from FIG. 5, the function x is approximately the function f as x increases.
This is a function that changes so as to reduce the slope of (x). Further, the function F (D) in the above equation 5 is the first
The second function for calculating the delay ratio with respect to the likelihood D in relation to the function
FIG. 6 shows an example of the function.

When a phoneme HMM is used as an acoustic model, as shown in FIG. 4, generally, the likelihood tends to decrease at the phoneme boundary and increase at the phoneme center. Therefore,
By using the functions of FIGS. 5 and 6, the acoustic likelihood is corrected so that the delay is large at the center of the phoneme and almost eliminated at the phoneme boundary as shown in FIG. In other words, the sound likelihood of the word hypothesis is delayed by delaying the peak of the sound likelihood of the central part of each phoneme of the word in the time direction (group delay) so as to move to a time delayed from the central part. to correct. As a result, all or a part of the acoustic likelihood at the phoneme center is evaluated at a time close to the phoneme boundary, and the same effect as in the fourth conventional example can be expected.

Next, the configuration and operation of the continuous speech recognition apparatus of FIG. 1 will be described. In FIG. 1, a phoneme HMM memory 11 is connected to the word matching unit 4 and stores a phoneme HMM in advance, and the phoneme HMM is represented including each state,
Each state has the following information. (A) State number (b) Acceptable context class (c) List of preceding state and succeeding state (d) Parameter of output probability density distribution (e) Self transition probability and transition probability to succeeding state The phoneme HMM used in the example is created by converting a predetermined speaker mixed HMM because it is necessary to specify which speaker each distribution originates from. Here, the output probability density function is a Gaussian mixture distribution having a 34-dimensional diagonal covariance matrix.

The word dictionary memory 12 is connected to the word collating unit 4 and stores the word dictionary in advance. The word dictionary stores a reading of each word of the phoneme HMM in the phoneme HMM memory 11 with a symbol. The symbol string shown is stored. Further, the statistical language model memory 13 is connected to the word matching unit 4 and stores a predetermined statistical language model in advance. Here, the statistical language model is described in, for example, the related art document 6
"Hirokazu Masataki et al.," Variable Length Statistical Language Model for Continuous Speech Recognition ", IEICE Technical Report, SP95
-73, November 1995 ".
A language model called variable length N-gram whose length in the time direction is variable can be used. The statistical language model includes a variable length N-gram of a part of speech class and a word.
And expressed as a bigram between the following three types of classes. (A) part-of-speech class, (b) class of word separated from part-of-speech class, and (c) class formed by connecting connected words.

In the continuous speech recognition apparatus of FIG. 1, the feature extracting unit 2, the word matching unit 4, the likelihood correcting unit 7, and the word hypothesis narrowing unit 6 are constituted by, for example, a digital computer having a CPU. Is done. The buffer memories 3 and 5 and the phoneme H
The MM memory 11, the word dictionary memory 12, the statistical language model memory 13, and the underestimated likelihood memory 14 are composed of, for example, a hard disk memory.

In FIG. 1, a uttered voice of a speaker is input to a microphone 1 and converted into a voice signal, and then input to a feature extracting unit 2. After performing A / D conversion on the input audio signal, the feature extraction unit 2 performs, for example, LPC analysis, and performs 34-dimensional feature parameters including logarithmic power, 16th-order cepstrum coefficient, Δlogarithmic power, and 16th-order Δcepstrum coefficient. Is extracted. The time series of the extracted feature parameters is input to the word matching unit 4 via the buffer memory 3.

The word collating unit 4 uses the one-pass Viterbi decoding method to store the phoneme HMM memory 1 based on feature parameter data input via the buffer memory 3.
1, a word hypothesis is detected using the word dictionary in the word dictionary memory 12, and the statistical language model in the statistical language model memory 13, and the acoustic likelihood based on the phoneme HMM, And calculates the linguistic likelihood based on the statistical language model, and outputs the linguistic likelihood together with the word hypothesis to the likelihood correcting unit 7. Here, the word matching unit 4 determines, for each state of each HMM at each time,
The likelihood within a word and the acoustic likelihood from the start of utterance are calculated. The likelihood including the acoustic likelihood and the linguistic likelihood is individually provided for each word identification number, word start time, and preceding word difference. Also,
In order to reduce the amount of calculation processing, the grid hypothesis of a low total likelihood among the total likelihoods calculated based on the phoneme HMM, the word dictionary, and the statistical language model is reduced. The word matching unit 4 outputs the resulting word hypothesis and information on the overall likelihood to the likelihood correction unit 7 together with time information (specifically, for example, a frame number) from the utterance start time.

In response to this, at time (t-1), likelihood correction section 7 calculates the ｛Q _A (S , T
-1) -Q _A ′ (S, t−1)} is stored in the underestimated likelihood memory 14, and at time t, the underestimated is calculated using the above equations 6 and 7. The sound likelihood Q _A ′ (S, t) corrected so as to calculate the total likelihood Q ′ _all (S, t) by using the above equation (8),
The word hypothesis having the calculated overall likelihood Q ′ _all (S, t) is output to the word hypothesis narrowing unit 6 via the buffer memory 5.

Based on the word hypotheses having the total likelihood output from the likelihood correction unit 7 via the buffer memory 5, the word hypothesis narrowing unit 6 determines the words of the same word having the same end time and different start time. With respect to the hypothesis, the word is represented by one word hypothesis having the highest likelihood among the calculated total likelihoods from the utterance start time to the end time of the word for each head phoneme environment of the word. After narrowing down hypotheses, of the word strings of all the narrowed word hypotheses,
A word string of a hypothesis having the maximum overall likelihood is output as a recognition result. In the present embodiment, preferably, the first phoneme environment of the word to be processed is three phonemes including the last phoneme of the word hypothesis preceding the word and the first two phonemes of the word hypothesis of the word. I mean a line.

[0040] For example, as shown in FIG. 2, the (i-1) th word W _i-1 of the following, a phoneme string a _1, a _2, ..., comes i-th word W _i, which consists of a _n Sometimes, six hypotheses Wa, Wb, Wc, Wd, We, and Wf are assumed as the word hypotheses of the word Wi _-1.
Exists. Here, the former three word hypotheses Wa, W
It is assumed that the final phonemes of b and Wc are / x /, and the final phonemes of the latter three word hypotheses Wd, We and Wf are / y /. Of the hypotheses in which the end time t _e is equal to the head phoneme environment (in FIG. 2, the top three word hypotheses whose head phoneme environment is “x / a ₁ / a ₂ ”), the hypothesis with the highest overall likelihood (for example, FIG. 2 except for the top hypothesis). Note that the fourth hypothesis from the top has a different phoneme environment, that is,
Since the last phoneme of the preceding word hypothesis is y instead of x, the fourth hypothesis from the top is not deleted. That is, only one hypothesis is left for each final phoneme of the preceding word hypothesis. FIG.
In the example, leave one hypothesis for the final phoneme / x /
Leave one hypothesis for the final phoneme / y /.

In the above embodiment, the head phoneme environment of the word is defined as a sequence of three phonemes including the last phoneme of the word hypothesis preceding the word and the first two phonemes of the word hypothesis of the word. Although defined, the present invention is not limited to this. The phoneme sequence of the preceding word hypothesis including the final phoneme of the preceding word hypothesis, and at least one phoneme of the preceding word hypothesis that is continuous with the final phoneme, And a phoneme sequence that includes a phoneme sequence that includes the first phoneme of the word hypothesis.

[0042]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present inventor conducted a word graph generation experiment using a natural utterance database to confirm the effectiveness of the continuous speech recognition device of FIG. A spoken language database owned by the applicant with the task of “travel planning” (for example, see Prior Art Document 7, “Morimoto et al.,
“A Speech and Language Database for Speech Transl
ation Research ", Proc. of ICSLP94, pp. 1791-1794, 1994
See year. 7) “Hotel reservation” dialogue (voices from the applicant of the dialogue, 3 males and 4 females, 100 voices, 9 voices)
83 words). The acoustic analysis was performed at a sampling frequency of 12 kHz.
z, frame interval 10 msec, hamming window 20 msec
c, where the feature parameters are 1 to
A 16th-order LPC cepstrum, 1st to 16th-order ΔLPC cepstrum, log power, and Δlog power were used. 40
An acoustic model that is HMNet mixed in one state with 5
The model trained using 150 reading utterances was speaker-adapted with about 20 natural utterances. The language model created a class bigram of 713 classes from 828 conversations including a total of 330,513 words. The word perplexity of the test set is 49.6. 6,635 vocabulary words
It is assumed that all words in the evaluation data are included and there are no unknown words. Further, the weight α of the language likelihood with respect to the acoustic likelihood in Equation 2 is set to 4.5.

The recognition performance of the apparatus was evaluated based on the word accuracy with respect to the beam width and the CPU time (hour). When a search is performed with a beam having a constant likelihood width, the word recognition rate increases as the beam width is increased, but when the beam width is increased to a certain degree or more, the word recognition rate decreases. This phenomenon can be explained as the fact that the expansion of the beam width for searching for the word hypothesis worked not to lower the lower limit of the beam but to raise the upper limit. Therefore, in this embodiment, the comparison is performed near the peak of the word recognition rate.

FIG. 7 is a graph showing the experimental results of the continuous speech recognition apparatus of FIG. 1 and showing the word recognition rate with respect to the beam width. FIG. 8 is an experimental result of the continuous speech recognition apparatus of FIG. 7 is a graph showing CPU calculation time (time) with respect to a beam width.

In FIG. 7, comparing the word recognition rates when the beam width is from 60 to 70, it can be seen that the peak reaches a narrower beam width in the case of the likelihood correction than in the case of no likelihood correction. . Further, when the characteristic curve without likelihood correction is shifted by about −2 in the beam width direction, the two characteristic curves almost overlap, and thus the likelihood correction has the same effect as reducing the beam width by about 2. As can be seen from FIG. 8, if the beam width is the same, the recognition time is not affected by the presence / absence of likelihood correction. Less is needed, where the reduction in computation time is about 10%.

As described above, according to the present embodiment, for the word hypothesis, the peak of the acoustic likelihood at the central part in the time direction of each phoneme of the word is set at the time delayed from the central part. , And the acoustic likelihood of the word hypothesis is corrected, so that the word hypothesis can be narrowed down with a narrower beam width compared to the third conventional example.
Continuous speech recognition of spontaneous utterance can be performed with a smaller calculation cost and a higher recognition rate.

Further, according to the present embodiment, for a word hypothesis of the same word having the same end time and different start time,
For each head phoneme environment of the word, narrow down the word hypothesis so that it is represented by one word hypothesis having the highest overall likelihood among the calculated overall likelihoods from the utterance start time to the end time of the word. Do. That is, one for each preceding word
Compared to the word pair approximation method of the second conventional example, which is represented by two word hypotheses, the number of word hypotheses is larger in order to collectively treat words having the same leading phoneme of the first phoneme (that is, the last phoneme of the preceding word). Can be reduced, and the approximation effect is large. In particular, when the number of words increases, the reduction effect is large. Therefore, even when the continuous speech recognition device is used for recognizing natural utterances in which interjections are inserted, stagnant, and rephrased frequently, the calculation cost required for merging or dividing word hypotheses is lower than in the conventional example. It will be smaller than that. That is, the processing amount required for speech recognition is reduced, and therefore, the working memory (not shown) of the word matching unit 4, the buffer memory 5, and the working memory (not shown) of the word hypothesis narrowing unit 6 are provided. )), The storage capacity required for a storage device for voice recognition is reduced, while the processing amount is reduced, so that the processing time for voice recognition can be reduced.

[0048]

As described above in detail, according to the continuous speech recognition apparatus of the first aspect of the present invention, the word hypothesis of the uttered voice sentence is detected based on the voice signal of the input uttered voice sentence. In a continuous speech recognition device provided with a speech recognition means for continuously recognizing speech by calculating an acoustic likelihood, the speech recognition means calculates the acoustic likelihood of a central part of each phoneme of a word in the time direction. The sound likelihood of the word hypothesis is corrected by delaying to move to a time delayed from the center. Therefore, word hypotheses can be narrowed down with a narrower beam width than in the third conventional example, and a smaller calculation cost, that is, a processing time for speech recognition is reduced, and a higher recognition rate is obtained. Can perform continuous speech recognition of natural speech.

Further, in the continuous speech recognition apparatus according to the second aspect, in the continuous speech recognition apparatus according to the first aspect,
The speech recognition means calculates, for each word phoneme environment of the same word having the same end time and different start time, a calculated acoustic likelihood from the utterance start time to the end time of the word for each head phoneme environment of the word. The word hypotheses are narrowed down so as to be represented by one word hypothesis having the highest overall likelihood among the total likelihoods including the degrees. Therefore, even when the continuous speech recognition device is used for recognizing natural utterances in which interjections are inserted, stagnant, and rephrased frequently, the calculation cost required for merging or dividing word hypotheses is lower than in the conventional example. It will be smaller than that. That is, the amount of processing required for speech recognition is reduced, and therefore, the storage capacity required for the storage device for speech recognition is reduced, while the amount of processing is reduced, so that the processing time for speech recognition is reduced. can do.

[Brief description of the drawings]

FIG. 1 is a block diagram of a continuous speech recognition apparatus according to an embodiment of the present invention.

FIG. 2 is a timing chart showing a process of a word hypothesis narrowing section 6 in the continuous speech recognition device of FIG.

FIG. 3 is a diagram illustrating a relationship between acoustic likelihoods in a third conventional example of a continuous speech recognition apparatus and the continuous speech recognition apparatus of the present embodiment in FIG. 1;

FIG. 4 shows a phoneme / a in the continuous speech recognition apparatus of FIG.
6 is a graph showing an example of the relationship between the acoustic likelihood before and after correction by the likelihood correction unit 7 with respect to /, and is a graph showing a temporal change in the acoustic likelihood.

FIG. 5 is a graph showing a first function f (x) for calculating a delay ratio with respect to likelihood, which is used in the likelihood correction unit 7 of FIG. 1;

FIG. 6 is a graph showing a second function F (D) for calculating a delay ratio with respect to likelihood, which is used in the likelihood correction unit 7 of FIG. 1;

7 is a graph showing experimental results of the continuous speech recognition apparatus of FIG. 1 and showing a word recognition rate with respect to a beam width.

FIG. 8 is a graph showing experimental results of the continuous speech recognition apparatus of FIG. 1, showing CPU calculation time (time) with respect to a beam width.

[Explanation of Codes] 1 ... microphone, 2 ... feature extraction unit, 3, 5 ... buffer memory, 4 ... word collation unit, 6 ... word hypothesis narrowing unit, 7 ... likelihood correction unit, 11 ... phoneme HMM memory, 12 ... word dictionary memory, 13 ... statistical language model, 14 ... underestimated likelihood memory.

────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Yoshinori Sakasaka 5th Sanraya, Inaya, Seika-cho, Soraku-gun, Kyoto Pref. 241094 (JP, A) JP-A-5-341797 (JP, A) JP-A-8-6588 (JP, A) JP-A 8-123472 (JP, A) Patent 2731133 (JP, B2) Proceedings of the Fall Meeting of the 8th Annual Meeting ▲ I ▲ 3-3-6 “Examination of Delayed Decision Beam Search” p. 97-98 (Published September 25, 1996) Proceedings of the Acoustical Society of Japan Fall Meeting 2007 (I) 2-2-12 "Continuous Speech Recognition Method Using Word Graph" p. 61-62 (accepted by the National Diet Library on September 28, 1995) Transactions of the Institute of Electronics, Information and Communication Engineers, Vol. J 79-D- ▲ IIＮｏ No. 12, Decmber 1996, “Reducing the Number of Word Hypotheses for Large Vocabulary Continuous Speech Recognition,” p. 2117-2124, (Issued December 25, 1996) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 3/00 561 G10L 3/00 531 G10L 3/00 537 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A continuous speech having speech recognition means for continuously recognizing speech by detecting a word hypothesis of the speech speech sentence based on a speech signal of the speech speech input and calculating an acoustic likelihood. In the recognition device, the voice recognition means may delay the acoustic likelihood of a central part of each phoneme of the word in the time direction so as to move to a time delayed from the central part, and A speech recognition device characterized by correcting the following. 2. The method according to claim 1, wherein the speech recognition unit calculates a word hypothesis of the same word having a same end time and a different start time, for each head phone environment of the word, from a utterance start time to an end time of the word. 2. The continuous speech recognition apparatus according to claim 1, wherein the word hypothesis is narrowed down so as to be represented by one word hypothesis having the highest total likelihood among the total likelihoods including the acoustic likelihood. .