JP2002358097A - Voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the precision by giving a low reliability to a word, which is an erroneously recognized word though acoustically resembling a correct answer word, when giving the reliability showing certainty of the recognition result to each of words constituting a word string which is the recognition result of continuous voice recognition. SOLUTION: A voice recognition device is provided with a linguistic reliability calculation means 13 which uses a forward statistical language model 14 for reliability calculation and a backward statistical language model 15 for reliability calculation to calculate the reliability showing whether a word has been correctly recognized or not to each of words constituting the word string, which is the result of continuous voice recognition, on the basis of a linguistic statistic.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus, and more particularly, to a speech recognition apparatus having a function of assigning a reliability indicating the certainty of a recognition result to each of words forming a word string as a result of continuous speech recognition. The present invention relates to a voice recognition device having the same.

[0002]

2. Description of the Related Art Current speech recognition technology does not always provide a recognition rate of 100%. In particular, an utterance to be recognized in continuous speech recognition often includes a plurality of words, and therefore, there is a greater chance of erroneous recognition occurring in any one of the words in one utterance.

As a countermeasure against this kind of erroneous recognition, a method of calculating the reliability indicating the certainty of the recognition for each word of the recognition result, and rejecting the word with low reliability or confirming it with the user can be considered. As a method of calculating the reliability of the recognition result, there is, for example, Japanese Patent Application Laid-Open No. 4-255900. Here, as a prior art, Japanese Patent Application Laid-Open No.
The technique disclosed in Japanese Patent No. 55900 will be described.

FIG. 3 is a block diagram showing the configuration of a conventional speech recognition apparatus disclosed in Japanese Patent Application Laid-Open No. 4-255900. In FIG. 3, reference numeral 1 denotes an input end of an audio signal, 2 denotes an input audio signal input from the input end 1, 3 denotes analysis means for performing an acoustic analysis of the input audio signal 2, and 4 denotes an input audio obtained by the analysis means 3. 5 is a continuous speech recognition means for performing continuous speech recognition using the feature vector time series 4, 6 is an acoustic model, 7 is a speech recognition language model, and 8 is a continuous speech recognition means 5. The reference likelihood calculating means 9 calculates the reference likelihood of the time series 4 of the feature vector of the input speech signal 2 using the acoustic model 6, and the reference 10 obtained by the reference likelihood calculating means 9 Likelihood, 11 is an acoustic reliability calculating means for calculating the reliability of each word of the word string included in the speech recognition result 8 using the reference likelihood 10, and 12 is a recognition result provided with the reliability. .

The acoustic model 6 is a continuous distribution type H
MM (Hidden Markov Model, Hidden Markov Model) is used. The acoustic model 6 models one word with a word unit, that is, one model. Therefore, the same number of acoustic models as the number of words to be recognized are prepared.
One model is composed of a plurality of states, and the topology of the model is a left-to-right type.

Further, a word bigram model, which is a statistical language model, is used as the language model 7 for speech recognition. The recognition target is, for example, a user utterance related to hotel reservation. The speech recognition language model 7 is learned in advance using text data in which a large amount of user utterances related to hotel reservations are transcribed.

Next, the operation of the conventional speech recognition apparatus will be described with reference to FIG. When an audio signal 2 is input from the input terminal 1 of the audio signal, the analysis means 3 divides the audio signal 2 into a plurality of short time intervals (hereinafter referred to as frames) on a time axis. For example,
Acoustic analysis is performed using the LPC (Linear Predictive Coding) method, and converted into a feature vector X. This feature vector X is, for example, an LPC cepstrum. The analysis means 3 performs the acoustic analysis on all frames, and obtains a time series 4 of feature vectors X ₁ , X ₂ , X ₃ ,. . . , And outputs the X _T. Here, the subscript indicates the frame number of each feature vector, and T indicates the total number of frames of the audio signal 2.

[0008] The continuous voice recognition means 5 outputs the output of the analysis means 3.
The time vector 4 of the feature vector
Pattern of time series 4 of Torr and acoustic model 6 in word units
Perform matching. As a method of pattern matching
For example, using the one-pass DP matching method,
Process in the forward direction of the time axis from
As the game progresses, the speech recognition language model 7
Connect acoustic model 6 according to word bigram probability
To perform pattern matching. Putter up to frame T
When the matching is completed, the continuous speech recognition means 5
Word string w as recognition result 8₁, W_Two, ..., w_i, ..., w
_N(W_iIs the ith word from the beginning in the word string of the recognition result,
N is the length of the word string) and each word w_n(N = 1 to N) start
Frame s_n(N = 1 to N), end frame e_n(N = 1
~ N) and the likelihood L of each word_n(N = 1 to N) is output
I do. Each word w_n(N = 1 to N) start frame
S_n(N = 1 to N), end frame e_n(N = 1 ~
N) is a pattern map by the one-pass DP matching method.
Ching is a time series 4 of feature vectors from frame 1
X ₁, X_Two, X_Three, ..., X_TUp to the end frame T
The time axis is reversed from the end frame T to frame 1.
Traceback the pattern matching result to
Therefore, it can be obtained.

On the other hand, the reference likelihood calculating means 9 comprises the analyzing means 3
LR _i (i = 1 to T), which is the reference likelihood 10 for each frame, is output according to the following equation (1) by using the time series 4 of the feature vector output as the input.

[0010]

(Equation 1)

[0011] b _k (X _i ) on the right side of the above equation (1)
Is the likelihood of the state k of the acoustic model 6 with respect to the feature vector X _i of the frame i. Therefore equation (1) means that the reference likelihood LR _i the maximum likelihood in states of all the acoustic model.

Next, the acoustic reliability calculation means 11 outputs
Word sequence w that is knowledge result 8₁, W_Two, ..., w_i, ..., w_NAnd each
Word w_n(N = 1 to N) start frame s_n(N = 1 ~
N), end frame e_n(N = 1 to N) and each unit
Word likelihood L_n(N = 1 to N) and L which is the reference likelihood 10
R_i(I = 1 to T) as input and make the following equation (2)
Thus, the reliability S of each word⁽¹⁾ _n(N = 1 to N)
You. Then, simply as the recognition result 12 to which the reliability is given,
Word string w₁, W_Two, ..., w_i, ..., w_NAnd reliability S⁽¹⁾ ₁, S
⁽¹⁾ _Two, ..., S⁽¹⁾ _i, ..., S⁽¹⁾ _NIs output.

[0013]

(Equation 2)

Word likelihood L_nDepends on the speakers
Move, but the reference likelihood LR_iAlso depends on the speakers
Fluctuate. Therefore, as shown in equation (2),
Taking the difference, the number of word frames (e_n-S_n+1)
By doing so, fluctuations due to differences in speakers are reduced,
Can be used as an index to indicate the reliability of recognition results
You. Note that the reference likelihood LR_iIs the total acoustic model for each frame
Ln <= Σ_iLR_iconnection of
Holds. Therefore, S calculated by equation (2) ⁽¹⁾ _n
Is S when the reliability is the highest⁽¹⁾ _n= 0 and reliability
Becomes a large negative value as the value decreases.

[0015]

In the conventional speech recognition apparatus, since it is configured as described above, for example,
The utterance is “(pause) staying for two nights (pause)” and the continuous speech recognition result is “(pause) staying for almost two nights (pause)” (w ₁ = pause, w ₂ = “two nights”, w ₃ = “near”, w ₄ = “do”, w ₅ = pose)
w ₃ = “Nearby” is misrecognized, but acoustically “Accommodation”
Due to the similar, start frame s ₃ end frame e cumulative value sigma _i LR _i and likelihood L ₃ word reference likelihood at intervals of up to ₃ becomes a value close, close to the high reliability (0 ) Value. That is, when the incorrectly recognized word is acoustically similar to the correct word, there is a problem that the reliability increases.

The present invention has been made to solve such a problem. Even when a recognized word is acoustically similar to a correct word, if the word is a misrecognized word, the reliability is low. It is an object of the present invention to provide a speech recognition device that can provide the following.

[0017]

According to the present invention, there is provided a continuous voice recognition means for performing continuous voice recognition of an input voice and outputting a word string corresponding to the input voice as a recognition result; Means for storing a statistical language model for calculating reliability, which stores at least one statistical language model for calculating reliability, which gives the probability of occurrence of each word that can be linguistically followed by
For each of the words constituting the word string of the recognition result, using the statistical language model for reliability calculation,
A linguistic reliability calculation means for calculating the reliability of whether or not each word is correctly recognized.

In addition, for each of the words constituting the word string of the recognition result, the reliability of whether or not each word is correctly recognized is calculated based on the acoustic likelihood. Reliability calculating means, and reliability integrating means for calculating integrated reliability using both values of the reliability calculated by the linguistic reliability calculating means and the reliability calculated by the acoustic reliability calculating means. And further comprising:

Further, a word n-gram model is used as the reliability calculation statistical language model.

A word class n-gram model in which words are classified into several classes and summarized is used as the reliability calculation statistical language model.

As one or more kinds of statistical language models for calculating reliability, a forward statistical language model for calculating reliability, which is a conditional probability model of the word and a predetermined number of preceding words, A backward statistical language model for reliability calculation, which is a conditional probability model with a predetermined number of words, wherein the linguistic reliability calculation means is configured to generate a word sequence of a recognition result of the continuous speech recognition. Using the reliability calculation forward statistical language model to calculate a first reliability of whether or not each word is correctly recognized, using the reliability calculation backward statistical language model A second reliability of whether the word is correctly recognized is calculated.

Further, the linguistic reliability calculating means outputs a larger one of the first reliability and the second reliability as the reliability of the word.

Further, the linguistic reliability calculating means outputs a weighted sum of the first reliability and the second reliability as the reliability of the word.

[0024]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiment 1 FIG. 1 is a block diagram showing a configuration of the speech recognition device according to the first embodiment of the present invention. 1, reference numeral 1 denotes an input end of an audio signal, 2 denotes an input audio signal input from the input end 1, 3 denotes an analysis unit for performing an acoustic analysis of the input audio signal 2, and 4 denotes an input audio obtained by the analysis unit 3 A time series of feature vectors of the signal 2, a continuous speech recognition means 5 for receiving a time series 4 of the feature vectors of the input speech signal 2, performing continuous speech recognition of the input speech, and outputting a word string as a recognition result, and 6 a sound Model, 7 is a language model for speech recognition, 8 is a speech recognition result output from the continuous speech recognition means 5, and 14 is a probability of occurrence of a word that can follow a predetermined word (that is, next to the word w _n−1 ). One or more types of forward-looking statistical language models for calculating reliability, indicating the conditional probability that a word w _n is connected, 15 is a probability of appearance of a word that can follow a given word (ie, a word before a word w _{n + 1} ). conditional probability that the word w _n is connected to the ), One or more types of backward statistical language models for calculating reliability, 13 are words that are the continuous speech recognition results 8 using the forward statistical language model 14 for calculating reliability and the backward statistical language model 15 for calculating reliability. A linguistic reliability calculating means for calculating the reliability of each word constituting the column as to whether or not each word is correctly recognized.

As in the prior art, in this embodiment, a continuous distribution type HMM is used as the acoustic model 6. The acoustic model 6 models one word with one word, that is, one model. Therefore, the same number of acoustic models as the number of words to be recognized are prepared. One model is composed of a plurality of states, and the topology of the model is left-to-to.
-A right type.

The speech recognition language model 7 uses a word bigram model, which is a statistical language model, as in the prior art. The recognition target is, for example, a user utterance related to hotel reservation. The speech recognition language model 7 is learned in advance using text data in which a large amount of user utterances related to hotel reservations are transcribed.

In the present embodiment, the same model as the language model 7 for speech recognition is used as the forward statistical language model 14 for reliability calculation newly added in the present invention.

In the backward statistical language model 15 for calculating reliability, the occurrence probability of the word in the ordinary n-gram model is the conditional probability with the preceding several words, whereas in the present embodiment, The occurrence probability of the word is set as a conditional probability with the following several words. For example, when a backward bigram model is used as the backward statistical language model 15 for calculating reliability, the occurrence probability of “two nights” in the word string “tomorrow” + “two nights” + “stay” is P (two nights | stay). Is calculated as The backward statistical language model 15 for reliability calculation is also learned in advance using text data in which a large amount of user utterances related to hotel reservations are transcribed.

Next, the operation of the speech recognition apparatus according to the present embodiment will be described. When a voice signal 2 is input from a voice signal input terminal 1, the analysis means 3 and the continuous voice means 5 perform the same operation as the conventional voice recognition apparatus, and the continuous voice means 5 outputs the word string w ₁ as the voice recognition result 8. , W ₂ , ...,
w _i ,..., w _N (where w _i is i
The second word, N, is the length of the word string.

Next, linguistic reliability calculation means 13, a word string w _1, w ₂ is a speech recognition result 8, ..., w _i, ..., as an input w _N, forward statistical language model for reliability calculation 14 and the backward statistical language model 15 for reliability calculation, the reliability S ⁽²⁾ _n (n =
1 to N). P (w _n | w _n-1 ) in the equation (3) is
This is the conditional probability that the word w _n is connected next to the word w _{n −1} held by the reliability calculation forward statistical language model 13. That is, using the probability value of statistical language models as the reliability of the word w _n. Then, the linguistic reliability calculating means 13 outputs the word string w ₁ ,
w ₂ ,..., w _i ,..., w _N and the reliability S ⁽²⁾ ₁ , S ^{(2 )} ₂ ,.
S ⁽²⁾ _i , ..., S ⁽²⁾ _N are output.

[0031]

(Equation 3)

[0032] In the normal voice recognition statistical language model is a word string _{w 1, w 2, ...,} w i, ..., used to determine the language likelihood for the entire w _N. And the word sequence language likelihood for the entire word string by acoustic likelihood and weighted sum _{w 1, w 2, ...,} w i, ..., and calculates a speech recognition score w _N. Therefore, for example, the utterance is “(pause) staying for two nights (pause)”, and “(pause) staying for nearly two nights (pause)” and “(pause) staying for two nights (pause) as voice recognition result candidates )), It can be expected that the latter has a higher language likelihood, but the acoustic likelihood is higher in the former, and the speech recognition score obtained by weighted sum of the language likelihood and the acoustic likelihood is also higher in the former. Can be higher. In this case, it is difficult in the above-described related art to lower the reliability of the word “near” which is an erroneously recognized word.

On the other hand, in this embodiment, a word string
The language likelihood for each individual word
The feature is that the language likelihood is used as the reliability of the word.
Thus, the field of recognition errors due to acoustically similar words
Linguistically preceding word w _n-1When the connection probability with is low
If the word w_nCan be less reliable
It works. For example, the utterance is "(pause) I will stay for two nights
(Pause) "and the continuous speech recognition result is" (pause)
I'll be close (pause) "(w₁= Pause, w_Two= "Two
Night ", w_Three= "Near", w_Four= "I do", w_Five= Poe
Z), w_Three= "Nearby" is incorrect and incorrect
It is acoustically similar to "Accommodation", but in normal Japanese
The word "Night Night" + "Near" is rare, so the word
Likelihood P (near | two nights) of the Japanese language model decreases, w_Three=
The reliability for “near” can be reduced.

The reliability can be calculated by the following equation (4) or (5). P (w _n | w _{n + 1} ) in the equations (4) and (5) indicates that the word w _n is connected before the word w _{n + 1} held by the reliability calculation backward statistical language model 15. Is the conditional probability of MAX (,) in the expression (4) is an operator for selecting the larger one of the two values. Further, α in the equation (5) is a constant set in advance, for example, α = 0.5. In the expressions (4) and (5), the first reliability of whether each word is correctly recognized and the backward statistics for reliability calculation are calculated by using the reliability calculation forward statistical language model 14. Using the language model 15, a second reliability indicating whether each word is correctly recognized is obtained. In the equation (4), the larger value of the first and second reliability is determined. Calculated as word reliability,
In the equation (5), the weighted sum of the first and second reliability is calculated as the reliability of the word.

[0035]

(Equation 4)

[0036]

(Equation 5)

As described above, the reliability can be calculated with higher accuracy by considering the conditional probability of the word connection from both the front and the rear. For example, the utterance is “(pause) staying for two nights (pause)”, and the speech recognition result is “(pause) staying near two nights (pause)” (w ₁ = pause, w ₂ = “two nights”, w ₃ = “near”, w ₄ = “do”, w ₅ = pause), w ₄ = “do” is correct recognition, but w ₃ = “near” is erroneous recognition.
In the calculation of the reliability of the word w ₄ by the equation (3), P
(Near | near) value will be used. Since the "near" because of erroneous recognition "near" + word chain of "you" is rare in normal Japanese, w ₄ is a positive recognition reliability becomes low. On the other hand, in the formulas (4) and (5), in addition to the value of P (do | near), the value of P _bw (do | pause) is considered, and "do" + "pause" Word chains often occur in Japanese, so P _bw (I |
Pause) becomes a high value, reliability of w ₄ is a correct recognition has the effect of suppressing be lower.

In the above description, the forward statistical language model 14 for calculating reliability and the backward statistical language model 15 for calculating reliability are described.
Has been described using a bigram in word units as an example, but a word class in which parts of speech and words are classified into several classes and put together may be used as a unit of the language model. Similar effects can be obtained by using a trigram or another statistical language model.

For example, when a word class bigram (word class n-gram model) is used, the reliability may be calculated by the following expression (6) or (7) instead of expression (3).

[0040]

(Equation 6)

[0041]

(Equation 7)

In the equations (6) and (7), c _n is a class to which the word w _n belongs, and c _n-1 is a class to which the word w _n-1 belongs. P (c _n | c _n-1 ) is a conditional probability that the class to be connected next is c _n when the preceding class is c _n-1 . The P (w _n | c _n) the word w _n in the class c _n
Is the appearance probability of. Note that the difference between equation (6) and equation (7) is that P (w _n | c _n ) is not multiplied. This is because if the number of words belonging to word class c _n is large, P (w _{n n}
| C _n ) is reduced, and even if the bigram probability P (c _n | c _n-1 ) between word classes is large, there is an effect of preventing the reliability from being lowered. For example, as shown in this example, in a statistical language model relating to hotel reservation, a hotel name is set as one class, and the appearance probability P (w _n | c _n ) of each hotel name in the class is set as an equal probability. To N
If the number is P, the value of P (w _n | c _n ), which is the appearance probability of the hotel name from the class of the hotel name, is 1 / N. Therefore, as N increases, the reliability calculated by equation (6) decreases, but in equation (7), the value of reliability does not depend on the value of N because P (w _n | c _n ) is not multiplied. effective.

When the word class bigram is used, instead of the expression (4), the reliability may be calculated by the expression (8) or the expression (9). In Equations (8) and (9), P _bw (c _n | c _{n + 1} ) is a conditional probability that the word class c _{n + 1} precedes the word class c _n .

[0044]

(Equation 8)

[0045]

(Equation 9)

Similarly, instead of equation (5), the reliability may be calculated by equation (10) or (11).

[0047]

(Equation 10)

[0048]

[Equation 11]

As described above, in the speech recognition apparatus according to the present embodiment, the forward statistical language for calculating the reliability is defined as the occurrence probability of the word to be recognized as the conditional probability with the n words preceding the word. A word sequence which is a result of continuous speech recognition, comprising a model 14, and a backward statistical language model 15 for reliability calculation in which the occurrence probability of the word to be recognized is a conditional probability of n words following the word. The reliability of whether each word is correctly recognized is calculated based on the linguistic statistic about the likelihood of the context of the word for each of the words constituting Even when the word is acoustically similar to the correct word, if the word is a misrecognized word, low reliability can be given, and the accuracy of speech recognition can be increased.

Embodiment 2 FIG. 2 is a block diagram showing another configuration example of the speech recognition device according to the present embodiment.
In FIG. 2, the same parts as those in the above-described conventional example or the first embodiment are denoted by the same reference numerals, and description thereof will be omitted. In the present embodiment, the newly added portions are the reliability 16 calculated by the linguistic reliability calculation means 13 and the reliability 1 calculated by the acoustic reliability calculation means 11.
The reliability integration means 17 calculates the sum of the weights with 2 as the final reliability 18.

In the present embodiment, a continuous distribution type HMM is used as the acoustic model 6 as in the first embodiment. It is assumed that the acoustic model 6 models one word with a word unit, that is, one model. Therefore, the same number of acoustic models as the number of words to be recognized are prepared. One model is composed of a plurality of states, and the topology of the model is left-t
o-right type.

The speech recognition language model 7 is also used in the first embodiment.
The word bigram model, which is a statistical language model, is used in the same manner as described above. The recognition target is, for example, a user utterance related to hotel reservation. The speech recognition language model 7 is learned in advance using text data in which a large amount of user utterances related to hotel reservations are transcribed.

As the forward-looking statistical language model for reliability calculation newly added in the present invention, the same model as the speech recognition language model 7 is used in this embodiment.

The backward statistical language model 15 for calculating the reliability uses a backward bigram model as in the first embodiment, and learns in advance using text data in which a large amount of user utterances related to hotel reservation have been transcribed. Shall be.

Next, the operation of the speech recognition apparatus according to the present embodiment will be described. When the voice signal 2 is input from the voice signal input terminal 1, the analyzing means 3, the continuous voice means 5, and the linguistic reliability calculating means 13 perform the same operations as those of the voice recognition apparatus of the first embodiment, and The calculating means 13 outputs the word strings w ₁ , w ₂ ,
…, W _i ,…, w _N and reliability S ⁽²⁾ ₁ , S ⁽²⁾ ₂ ,…,
S ⁽²⁾ _i , ..., S ⁽²⁾ _N are output.

The time series 4 of the feature vector and the acoustic model 6 which are the outputs of the analysis means 3 are input to the reference likelihood calculation means 9, and the reference likelihood calculation means 9 and the acoustic reliability calculation means 1
1 operate in the same manner as in the prior art, and as a result, the acoustic reliability calculation means 11 outputs the reliability S ⁽¹⁾ ₁ , S ⁽¹⁾ ₂ ,.
⁽¹⁾ _i , ..., S ⁽¹⁾ Output _N.

Next, the reliability integration means 17 outputs the word strings w ₁ , w ₂ ,..., W _i ,..., W _N which are the recognition results 16 output from the linguistic reliability calculation means 13 and the reliability S ⁽²⁾ ₁ ,
S ⁽²⁾ ₂ ,..., S ⁽²⁾ _i ,..., S ⁽²⁾ _N and the reliability S ⁽¹⁾ ₁ , S ⁽¹⁾ ₂ ,. S
⁽¹⁾ _i, ..., S ⁽¹⁾ where _{N is} the input, the following (12) integrated reliability according formula _{^{S (10) 1, S (}} 10) 2, ..., S (10) i,
…, S ⁽¹⁰⁾ _N is calculated. Β in the equation (12) is a constant set in advance, for example, β = 0.5.

[0058]

(Equation 12)

The reliability integration means 17 gives the reliability.
Word string w as the recognized recognition result 18 ₁, W_Two, ..., w_i,
…, W_NAnd integrated reliability S^(Ten) ₁, S^(Ten) _Two, ..., S^(Ten) _i,
…, S^(Ten) _NIs output.

As described above, in the present embodiment,
The same effects as in the first embodiment can be obtained, and
Furthermore, since the reliability integration means 17 is provided to integrate the reliability 12 based on the acoustic likelihood and the linguistic reliability 16, it is possible to consider the reliability from both sound and language. Therefore, a higher degree of reliability can be obtained.

In this example, the linguistic reliability is (3)
S ⁽²⁾ ₁ , S ⁽²⁾ ₂ ,..., S ⁽²⁾ _i ,.
⁽²⁾ _N was used, but S ⁽³⁾ ₁ , S ⁽³⁾ ₂ ,..., S ⁽³⁾ _i , calculated by equation (4) or (8) or (9)
, S ⁽³⁾ _N or S ⁽⁴⁾ ₁ , S ⁽⁴⁾ ₂ ,..., S ⁽⁴⁾ _i , calculated by equation (5), equation (10), or equation (11)
.., S ^{(4 )} _N may be used. As the reliability based on the acoustic likelihood, a value obtained by the same method as that of the related art is used, but a value obtained by another method may be used.

[0062]

According to the present invention, there is provided a continuous voice recognition means for performing continuous voice recognition of an input voice and outputting a word string corresponding to the input voice as a recognition result, and linguistically before and after a predetermined word. Means for storing a statistical language model for reliability calculation that stores one or more statistical language models for reliability calculation that give the probability of appearance of each possible word; For each of the above, the speech recognition device includes linguistic reliability calculation means for calculating the reliability of whether each of the words is correctly recognized using the statistical language model for reliability calculation. Even if the word is acoustically similar to the correct word, it can be given low reliability if the word is an erroneously recognized word, and the accuracy of voice recognition can be improved.

Further, for each of the words constituting the word string of the recognition result, the reliability of whether or not each word is correctly recognized is calculated based on the acoustic likelihood. Reliability calculating means, and reliability integrating means for calculating integrated reliability using both values of the reliability calculated by the linguistic reliability calculating means and the reliability calculated by the acoustic reliability calculating means. It is possible to consider the reliability from both the sound and the language,
Higher accuracy reliability can be obtained.

Since the word n-gram model is used as the statistical language model for calculating the reliability,
Based on the conditional probability of the word preceding or succeeding the word, the reliability of each word constituting the word string of the recognition result can be obtained.

Further, since the word class n-gram model in which words are classified into several classes and put together is used as the statistical language model for calculating the reliability, the condition for the word preceding or following the word is used. Based on the attached probability,
The reliability of each word constituting the word string of the recognition result can be obtained.

As one or more kinds of statistical language models for calculating reliability, a forward statistical language model for calculating reliability, which is a conditional probability model of the word and a predetermined number of preceding words, A backward statistical language model for reliability calculation, which is a conditional probability model with a predetermined number of words, wherein the linguistic reliability calculation means is configured to generate a word sequence of a recognition result of the continuous speech recognition. Using the reliability calculation forward statistical language model to calculate a first reliability of whether or not each word is correctly recognized, using the reliability calculation backward statistical language model Since the second reliability of whether the word is correctly recognized is calculated, the reliability can be calculated with higher accuracy.

Further, the linguistic reliability calculating means outputs the larger value of the first reliability and the second reliability as the reliability of the word, so that the linguistic reliability can be further increased. Accuracy can calculate reliability.

Further, the linguistic reliability calculating means outputs the weighted sum of the first reliability and the second reliability as the reliability of the word. Degrees can be calculated.

[Brief description of the drawings]

FIG. 1 is a configuration diagram showing a configuration of a speech recognition device according to Embodiment 1 of the present invention.

FIG. 2 is a configuration diagram showing a configuration of a speech recognition device according to a second embodiment of the present invention.

FIG. 3 is a configuration diagram showing a configuration of a conventional voice recognition device.

[Explanation of symbols]

Reference Signs List 1 input terminal of voice signal, 2 input voice signal, 3 analysis means, 4 time series of feature vector of input voice signal, 5 continuous voice recognition means, 6 acoustic model, 7 language model for voice recognition, 8 voice recognition result, 9 Reference likelihood calculating means, 1
0 reference likelihood, 11 acoustic reliability calculation means, 12 recognition result, 13 linguistic reliability calculation means, 14 forward statistical language model for reliability calculation, 15 backward statistical language model for reliability calculation, 16, 18 reliability , 17 Reliability integration means.

Claims (7)

[Claims] 1. Continuous speech recognition of input speech is performed,
Continuous speech recognition means for outputting a word sequence corresponding to the input speech as a recognition result; and one or more statistical language models for calculating reliability, which give the appearance probabilities of words that can be linguistically continued before and after a predetermined word Using a statistical language model for reliability calculation, for each of the words forming the word string of the recognition result,
A speech recognition device comprising: a linguistic reliability calculation unit that calculates a reliability of whether each of the words is correctly recognized. 2. An acoustic reliability for each word constituting a word sequence of the recognition result, wherein the reliability of whether or not each word is correctly recognized is calculated based on the acoustic likelihood. Reliability calculation means; and reliability integration means for calculating integrated reliability using both values of the reliability calculated by the linguistic reliability calculation means and the reliability calculated by the acoustic reliability calculation means. The speech recognition device according to claim 1, further comprising: 3. The speech recognition device according to claim 1, wherein a word n-gram model is used as the reliability calculation statistical language model. 4. A word class n in which words are classified into several classes and compiled as the statistical language model for calculating reliability.
The speech recognition device according to claim 1 or 2, wherein a -gram model is used. 5. One or more types of statistical language models for calculating reliability, a forward statistical language model for calculating reliability, which is a conditional probability model of the word and a predetermined number of preceding words, A backward statistical language model for reliability calculation, which is a conditional probability model with a predetermined number of words, wherein the linguistic reliability calculation means comprises: Using the reliability calculation forward statistical language model to calculate a first reliability of whether or not each word is correctly recognized, using the reliability calculation backward statistical language model 5. The speech recognition device according to claim 1, wherein a second reliability of whether or not the word is correctly recognized is calculated. 6. The linguistic reliability calculating means outputs a larger value of the first reliability and the second reliability as the reliability of the word. 6. The voice recognition device according to 5. 7. The linguistic reliability calculating means outputs a weighted sum of the first reliability and the second reliability as the reliability of the word.
A speech recognition device according to claim 1.