JP2001242883A - Device and method for speech recognition, and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve speech recognition rate. SOLUTION: A word preliminary selection section 13 selects a word string, that is connected to the words for which scores are already computed and are made up with several words, from a dictionary database 18A. A matching section 14 computes a score for the leading word, that is the leading word of the word string, based on more than one word following the leading word of the word string. Then, a word connection information storage section 16 stores the connecting relationship between the leading word whose score is computed and the word that is connected with the leading word and the score is already computed. A control section 11 decides a speech recognition result, based on the connecting relationship and the score.

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, a speech recognition method, and a recording medium, and more particularly to, for example, a speech recognition apparatus and a speech recognition that can perform high-speed and high-accuracy speech recognition. The present invention relates to a method and a recording medium.

[0002]

2. Description of the Related Art FIG. 1 shows the configuration of an example of a conventional speech recognition apparatus.

[0003] A voice emitted by a user is input to a microphone (microphone) 1, which converts the input voice into a voice signal as an electric signal. This audio signal is supplied to an AD (Analog Digital) converter 2.
The AD converter 2 samples and quantizes an audio signal that is an analog signal from the microphone 1 and converts it into audio data that is a digital signal. This audio data is supplied to the feature extraction unit 3.

[0004] The feature extraction unit 3 performs audio processing on the audio data from the AD conversion unit 2 for each appropriate frame, thereby obtaining, for example, an MFCC (Mel Frequency Cep).
A feature amount such as strumCoefficient) is extracted and supplied to the matching unit 4. In addition, the feature extraction unit 3 further includes, for example, a spectrum, a linear prediction coefficient, a cepstrum coefficient,
It is possible to extract a feature amount such as a line spectrum pair.

The matching unit 4 uses the feature amount from the feature extracting unit 3 to refer to the acoustic model database 5, the dictionary database 6, and the grammar database 7 as necessary, and to input the voice ( Input voice)
Is recognized based on, for example, a continuous distribution HMM method or the like.

That is, the acoustic model database 5 stores acoustic models representing acoustic features such as individual phonemes and syllables in the language of the speech to be recognized. Here, since speech recognition is performed based on the continuous distribution HMM method,
As an acoustic model, for example, HMM (Hidden Markov M
odel) is used. The dictionary database 6 includes information on pronunciation (phonological information) of each word to be recognized.
Is stored in the word dictionary. The grammar database 7 stores grammar rules (language models) describing how words registered in the word dictionary of the dictionary database 6 are linked (connected). here,
As grammar rules, for example, context-free grammar (CFG)
Alternatively, a rule based on a statistical word chain probability (N-gram) or the like can be used.

[0007] The matching unit 4 refers to the word dictionary of the dictionary database 6 and connects the acoustic models stored in the acoustic model database 5 to form a word acoustic model (word model). Further, the matching unit 4 connects several word models by referring to the grammar rules stored in the grammar database 7, and uses the word models connected in this way to generate a continuous The voice input to the microphone 1 is recognized by the distributed HMM method. That is, the matching unit 4
The feature extraction unit 3 detects a sequence of a word model having the highest score (likelihood) at which a time-series feature amount is observed, and outputs a word sequence corresponding to the word model sequence as a speech recognition result. I do.

That is, the matching section 4 accumulates the appearance probabilities of the respective feature amounts for the word string corresponding to the connected word model, and uses the accumulated value as a score, and outputs the word string having the highest score as a voice. Output as recognition result.

Generally, the score calculation is performed by using an acoustic score (hereinafter, appropriately referred to as an acoustic score) given by an acoustic model stored in an acoustic model database 5 and a language given by a grammatical rule stored in a grammatical database 7. Score (hereinafter referred to as language score as appropriate)
It is performed by comprehensively evaluating

That is, for example, in the case of using the HMM method, the acoustic score is obtained from the acoustic model constituting the word model.
It is calculated for each word based on the probability of observing the sequence of feature values output by the feature extracting unit 3 (probability of appearance).
In the case of a bigram, for example, the language score is determined based on the probability that the word of interest and the word immediately before the word are linked (connected). Then, a speech recognition result is determined based on a final score (hereinafter, appropriately referred to as a final score) obtained by comprehensively evaluating an acoustic score and a language score for each word.

Specifically, the k-th word in a word string composed of a certain N words is w _k , the acoustic score of the word w _k is A (w _k ), and the language score is L (w _k ). , Respectively, the final score S of the word string is calculated according to, for example, the following equation.

S = Σ (A (w_k) + C_k× L (w_k)) (1) where Σ is the summation of changing k from 1 to N
It represents that. Also, C _kIs the word w_kLanguage score L
(W_k).

In the matching unit 4, for example, N which maximizes the final score shown in the equation (1), the word string w ₁ ,
w _2, · · ·, matching processing for obtaining the w _N is performed, the word sequence _{w 1, w 2, ···,} w N is output as a speech recognition result.

By performing the above processing,
In the speech recognition device of FIG. 1, for example, when the user utters “I want to go to New York”, an acoustic score is added to each word such as “New York”, “Nii”, “I want to go”, “I”. And the linguistic score, and when the final score obtained by comprehensively evaluating them is the largest, and when the word strings "New York", "to", "want to go",
"Is" is output as the speech recognition result.

In the above case, the word dictionaries of the dictionary database 6 include "New York", "Ni",
When 5 words of "I want to go", and "is" is registered, the arrangement of the 5 words that can be constructed using these five words, the presence of 5 ⁵ ways. Thus, simply, the matching section 4 evaluates the word sequence of the 5 ⁵ kinds, among them, must determine the best match to the utterance of the user (which maximized the final score). If the number of words to be registered in the word dictionary increases, the number of words arranged as the number of words becomes the same as the number of words to the number of words, so the number of word strings to be evaluated is huge. Becomes

Furthermore, since the number of words included in an utterance is generally unknown, not only a word string consisting of a sequence of five words but also a word string consisting of one word, two words,.
Need to be evaluated. Therefore, the number of word strings to be evaluated becomes even more enormous, and among such enormous word strings, the most probable speech recognition result is determined from the viewpoint of the amount of calculation and the memory capacity used. Efficient decisions are a very important issue.

As a method for improving the efficiency of the calculation amount and the memory capacity, for example, in the process of obtaining an acoustic score, if the acoustic score obtained in the middle of the process becomes equal to or less than a predetermined threshold, the score calculation is terminated. There are acoustic pruning techniques and linguistic pruning techniques for narrowing down words to be scored based on linguistic scores.

According to these pruning methods, the target of score calculation is a predetermined criterion (for example, an acoustic score in the middle of calculation or a language score given to a word as described above).
, The calculation amount can be reduced. However, on the other hand, if the refinement is strengthened, that is, if the criterion is made strict, even the correct speech recognition result is originally pruned, resulting in erroneous recognition. Therefore, in the case of the pruning method, it is necessary to perform narrowing down with a certain margin so that a correct speech recognition result is not pruned, and it is difficult to greatly reduce the calculation amount. .

Further, when calculating the acoustic score, if the calculation is performed independently for all the words for which the score is to be calculated, the amount of calculation becomes large. Therefore, a part of the calculation of the acoustic score for a plurality of words is performed. There has been proposed a method of sharing (sharing) the same. As a method of this commonization,
Of the words in the word dictionary, if the first phoneme is the same, use the acoustic model in common from the first phoneme to the same phoneme, and use the acoustic model for subsequent different phonemes. There is a method in which a single tree-structured network is configured as a whole by using each of them individually, and an acoustic score is obtained using the network. Specifically, for example, considering the words “Akita” and “Akebono”, the phoneme information of “Akita” is “akita” and the phoneme information of “Akebono” is “akebo”
If "no", the acoustic scores of the words "Akita" and "Akebono" are the same phonemes a, k
Is calculated for both uses. Then, the remaining phonemes k, i, t, a of the word “Akita” and the remaining phonemes e,
For b, o, n, and o, an acoustic score is calculated independently.

Therefore, according to this method, the calculation amount of the acoustic score can be greatly reduced.

However, according to this method, it is not possible to determine the word for which the acoustic score is to be calculated in the common portion (the portion where the acoustic score is also calculated). That is, in the case of the above words “Akita” and “Akebono”, if the acoustic score is calculated for the first two phonemes a and k,
It is not possible to identify whether the word whose acoustic score is calculated is “Akita” or “Akebono”.

In this case, for "Akita",
When the calculation of the acoustic score is started for the third phoneme i, it is possible to identify that the calculation target is “Akita”.
When the calculation of the acoustic score for e is started, it is possible to identify that the calculation target is “Akebono”.

Therefore, if a part of the calculation of the acoustic score is standardized, the word cannot be identified at the start of the calculation of the acoustic score of the word. Can not. as a result,
Before the start of the acoustic score of a word, it becomes difficult to use the linguistic pruning method as described above, and useless calculation may be performed.

Further, when a part of the calculation of the acoustic score is shared, the above-mentioned tree-structure network is formed for all the words in the word dictionary. Capacity is required.

As a method of increasing the calculation amount and the memory capacity, when calculating the acoustic score, not all the words in the word dictionary but the acoustic score are calculated. There is a method of preliminarily selecting a word (preliminary selection) and calculating an acoustic score only for the preselected word. Here, the preliminary selection is performed using, for example, a simple acoustic model or a grammar rule that is not so high in accuracy.

The method of preliminary selection is, for example, LR
Bahl, SV De Gennaro, PS Gopalakrishnan and
RL Mercer, "A Fast Approximate Acoustic Match f
or Large Vocabulary Speech Recognition ", IEEE Tran
s. Speech and Audio Proc., vol. 1, pp. 59-67, 1993 and the like.

By the way, the acoustic score of a word is calculated using a sequence of feature amounts of speech. If the starting point and the end point of the sequence of feature amounts used for the calculation are different, the obtained acoustic score also changes. Then, this change affects the final score of Expression (1) obtained by comprehensively evaluating the acoustic score and the language score.

The start point and end point of a sequence of feature amounts corresponding to a certain word, that is, boundaries between words (word boundaries) can be obtained by using, for example, dynamic programming (Dynamic Programming). That is, a score obtained by comprehensively evaluating an acoustic score and a linguistic score of a word in a word string that is a candidate for a speech recognition result is set as an arbitrary point in the series of feature amounts as a candidate for a word boundary (hereinafter, the word score is appropriately determined) ) Are sequentially accumulated. Then, at the time of accumulating the word scores, word boundary candidates giving the largest accumulated value are stored together with the accumulated value.

Thus, when the final accumulated value of the word scores is obtained, the best accumulated value, that is, the word boundary giving the largest final score is also obtained.

The method of finding a word boundary as described above is called Viterbi decoding or One pass decoding. For example, Seiichi Nakagawa, "Speech Recognition by Stochastic Model" , IEICE, pp.20-26, July 1, 1988
Details are described on the date.

In order to perform the above-mentioned preliminary selection effectively, it is very important to determine where the word boundary is located, that is, which point of the feature value sequence (feature value sequence) is to be the starting point.

Specifically, for example, as shown in FIG. 2A, in the feature amount sequence obtained for the voice “Today is fine weather”, the correct word boundary between “today” and “ha” Is the time t ₁ , when the preliminary selection is performed on the word “ha” following the word “today”, the preliminary selection is performed starting from the time t _1-1 preceding the correct time t _1. The preliminary selection is affected not only by the feature amount of the word "ha" but also by the feature amount of the last part of the word "today" immediately before. On the other hand, the time t _{1 + 1} to the trailing than the correct time t ₁
Is used as a starting point, the feature amount of the first part of the word "ha" is not used in the preliminary selection.

Therefore, in any case, if the starting point is wrong,
This will adversely affect the pre-selection and eventually the subsequent matching process.

In FIG. 2 (the same applies to FIGS. 5 and 6 to be described later), the direction from the left to the right represents the passage of time.
The voice section ends at time T.

In the above-described method using the dynamic programming, the calculation of the word score (the acoustic score and the language score) must be completed until the end of the feature amount sequence, that is, in FIG. Since the final word boundary cannot be determined, it is difficult to uniquely determine the word boundary that is the starting point of the preliminary selection at the stage of performing the preliminary selection.

Therefore, there is a method in which word boundary candidates are held until the calculation of a word score using a feature amount sequence of a speech section is completed.

In this method, for example, when the word score of the word “today” is calculated starting from the start time 0 of the voice section, the times t _1-1 , t ₁ , and t _{1 + 1 correspond} to the words “today”. , The three times t _1-1 , t ₁ , and t _{1 + 1} are held, and starting from each of them, preliminary selection of the next word is performed.

Now, the time t _{1-1 is obtained} by this preliminary selection.
Is the starting point, the two words “ha” and “good”
When the time point t ₁ is the starting point, one word “ha” is the starting point, and when the time point t _{1 +1} is the starting point, two words “ha” and “good”
It is assumed that words are obtained, respectively, and that the results shown in FIGS. 2B to 2G are obtained by calculating word scores for these words.

That is, FIG. 2B shows the time t_1-1With the starting point
Then, calculate the word score of the word "ha", the end point
At time t_TwoShows the state where was obtained,
FIG. 2C shows the time t_1-1Starts with the word "good"
Calculate the word score of
Time t_{2 + 1}Shows the state obtained. FIG.
(D) is the time t₁From the word "ha"
The core is calculated, and the time t_{2 + 1}
FIG. 2E shows the state at time t.₁
Starting from, calculate the word score for the word "ha",
As a candidate for the end point, time t_TwoShows the state where
ing. Further, FIG._{1 + 1}Starting from
To calculate the word score for the word "ha"
As a candidate, time t_TwoIs shown in the figure, and FIG.
2 (G) is the time t_{1 + 1}Starting from the word "good"
Calculate the word score, and select the time
t_{2 + 2}Shows the state obtained. In addition, in FIG.
T_1-1<T₁<T_{1 + 1}<T_Two<T _{2 + 1}<T_{2 + 2}Become
I have.

2B to 2G, FIG.
(B), FIG. 2 (E), and FIG. 2 (F), the word strings as candidates for the speech recognition result are the same as “today” and “ha”. also the end of the last "is" of the word string, because it is the same at the time t _2,
Those most appropriate among them, for example, select based on the accumulated value or the like of the word scores up to time t _2, the it is possible to discard the rest.

However, at this point, FIG.
(B), FIG. 2 (E), or FIG. 2 (F), plus the ones shown in FIG. 2 (C), FIG. 2 (D), and FIG. 2 (G). Since the correct one cannot be determined, it is necessary to keep these four cases. And for each of these four cases,
The preliminary selection will be performed again.

Accordingly, in this method, it is necessary to calculate the word score while holding many word boundary candidates until the calculation of the word score using the feature amount sequence of the voice section is completed. It is not preferable from the viewpoint of the efficiency of the calculation amount and the memory capacity.

In this case, as word boundary candidates,
If a truly correct word boundary is retained, in principle, the same correct word boundary as in the case of using the dynamic programming described above can be finally obtained, but on the other hand, the word boundary If you fail to maintain truly correct word boundaries,
Misrecognize words that start or end at that word boundary,
Further, due to this, the following word may be erroneously recognized.

On the other hand, in recent years, an acoustic model that depends on (considers) the context before and after has been used. The acoustic model depending on the context before and after is an acoustic model of the same syllable (or phoneme) that is modeled as being different depending on the syllable immediately before or immediately after it. Means that the syllable immediately before or after
For example, different acoustic models are modeled for “ka” and “sa”.

Acoustic models that depend on contexts before and after are roughly divided into acoustic models that depend on contexts within words and acoustic models that depend on contexts between words.

When an acoustic model that depends on the context in a word is used, when a word model of “today” (today) is created by connecting the acoustic models “kyo” and “u”, As the model "Kyo", a model that depends on the syllable "U" immediately after it (an acoustic model "Kyo" considering the influence of the syllable "U" immediately after) is used. The one that depends on the syllable "Kyo" immediately before it is used.

On the other hand, when an acoustic model that depends on a context extending between words is used, as described above,
When creating the word model “Today” by connecting the acoustic models “Kyo” and “U”, if the word that follows is “Ha”, it is assumed to be the acoustic model “U”, The one that depends on the first syllable "ha" of the word is used. This acoustic model that depends on the context spanning between words is called a crossword model.

When this crossword model is applied to speech recognition for making a preselection, the context between the preselected word and the immediately preceding word can be considered, but the immediately following word can be considered. Has not yet been determined, so the context between it and the immediately following word cannot be considered.

Therefore, a method of obtaining a word model which is likely to be located immediately after the word to be preliminarily selected in advance and creating a word model in consideration of the context between the word and the word is proposed. is there. That is, for example,
Immediately after the word "today",
If "ga" or "no" is likely to be located,
There is a method of creating a word model using an acoustic model “u” corresponding to the last syllable of the word model of the word “today”, which considers “ha”, “ga”, and “no”, respectively.

However, in this method, an extra context is always considered, which is not preferable from the viewpoint of the efficiency of the calculation amount and the memory capacity.

For the same reason, it is also difficult to calculate the language score of a preselected word in consideration of the word immediately after it.

On the other hand, as a speech recognition method taking into account not only the word preceding the word of interest (the word of interest) but also the following word, for example, R. Schwartz and Y.
L. Chow, "The N-Best Algorithm: An Efficient and E
xact Procedure for FindingThe Most Likely Sentence
Hypotheses ", Proc. ICASSP, pp.81-84, 1990, two-pass decoding.
ng).

That is, FIG. 3 shows a schematic configuration of an example of a conventional speech recognition apparatus for performing speech recognition by two-pass decoding.

[0054] In FIG. 3, the matching section 4 ₁ performs, for example, the same matching processing as the matching section 4 of FIG. 1, and outputs a word string obtained as a result. However, multiple matching section 4 _1, a plurality of word strings obtained as a result of the matching process, what is the final speech recognition result instead of outputting only one of those probable candidates for the speech recognition result Output.

[0055] The output of the matching section 4 ₁ is supplied to the matching section 4 _2, the matching unit 4 _2, the matching unit 4 ₁
Performs a matching process for re-evaluating the likelihood that the word string is the speech recognition result for each of the plurality of word strings output by. That is, in the word string as a speech recognition result matching section 4 ₁ outputs, as well as the words immediately preceding a word, because there is also a word after, the matching section 4 _2, not only the last word, The acoustic score and the language score are newly obtained by the crossword model in consideration of the immediately succeeding word. Then, the matching section 4 _2, a plurality of word strings from the matching section 4 _1,
Based on the new acoustic score and linguistic score of each word string, a likely speech recognition result is determined and output.

[0056] In the two-pass decoding, such as described above, generally, the matching section 4 ₁ performs first matching processing, not so high accuracy, simple acoustic models, a word dictionary, and a grammar rule are used, the matching section 4 ₂ for subsequent matching process, high accuracy of acoustic models, a word dictionary, grammar rules used. Thus, the speech recognition apparatus of FIG. 3, the amount of processing in the matching unit 4 ₁ and 4 _2, respectively, while being relaxed to each other, thereby making it possible to obtain high speech recognition result with high accuracy.

[0057] Figure 3, as described above, 2 is a voice recognition device of the path decoding, the subsequent stage of the subsequent matching section 4 ₂ of FIG. 3, the multi-pass decoding adding the same matching unit (multi Pass speech recognition systems have also been proposed.

However, in the two-pass decoding and the multi-pass decoding, the next matching process cannot be performed until the first matching process is completed. Therefore, the delay time from the input of the voice to the output of the final voice recognition result increases.

Therefore, for example, when the first matching process for some words is completed, for the some words, the subsequent matching process using the crossword model is sequentially repeated. The way
For example, M. Schuster, "Evaluation of a Stack Decode
r on a Japanese Newspaper Dictation Task ",
1-R-12, pp.141-142, 1997, etc.

[0060]

However, even in the method of sequentially repeating the matching process, in the first matching process, the word immediately after the preselected word has not been determined yet, so that the word immediately after the preselected word is not determined. It is difficult to consider the context between.

The present invention has been made in view of such a situation, and is intended to enable the acoustic or linguistic evaluation of a word to be performed in consideration of the immediately succeeding word. .

[0062]

According to the present invention, there is provided a speech recognition apparatus for selecting a word string including a plurality of words connected to a word whose score has been calculated from a group of words to be subjected to speech recognition. Selecting means, score calculating means for calculating a score for a head word which is the head word of the word string based on one or more words following the head word of the word string, head word for which the score has been calculated, A connection relation storage unit for storing a connection relation between a word to which the first word is connected and a word for which a score has been calculated, and a determination means for determining a speech recognition result of voice based on the connection relation and the score. It is characterized by the following.

In the score calculation means, for the first word,
An acoustic or linguistic score can be calculated based on one or more words following the first word of the word sequence.

The score calculating means can calculate a linguistic score with reference to a probabilistic language model.

The speech recognition apparatus according to the present invention may further include storage means for storing a word group to be subjected to speech recognition.

Further, the speech recognition apparatus of the present invention can further comprise an extracting means for extracting the characteristic amount from the voice, and in this case, the score calculating means calculates the score using the characteristic amount. Can be done.

The speech recognition method according to the present invention comprises: a word string selection step of selecting a word string consisting of a plurality of words connected to a word whose score has been calculated from a group of words to be subjected to speech recognition; A score calculation step of calculating a score based on one or more words following the first word of the word string for the first word that is the first word of the word string, and connecting the first word for which the score has been calculated and the first word; It is characterized by comprising a connection relation storing step of storing a connection relation between words whose scores have already been calculated, and a determination step of determining a speech recognition result of the voice based on the connection relation and the score.

The recording medium of the present invention includes a word string selecting step of selecting a word string composed of a plurality of words connected to a word whose score has been calculated from a group of words to be subjected to speech recognition; For the first word, the first word,
A score calculation step of calculating a score based on one or more words following the first word of the word string; and a first word for which the score has been calculated and a word for which the first word is connected and for which a score has been calculated. Based on the connection relationship and the score,
And a determining step of determining the voice recognition result of the voice.

In the speech recognition apparatus, the speech recognition method, and the recording medium of the present invention, a word string composed of a plurality of words connected to the words whose scores have already been calculated is selected from a group of words to be subjected to speech recognition. Then, a score is calculated for the first word, which is the first word of the word string, based on one or more words following the first word of the word string. Then, the connection relation between the head word for which the score has been calculated and the word to which the head word is connected and for which the score has already been calculated is stored. Based on the connection relation and the score, the speech recognition result is obtained. Is determined.

[0070]

FIG. 4 shows a configuration example of an embodiment of a speech recognition apparatus to which the present invention is applied. In the figure, parts corresponding to those in FIG. 1 are denoted by the same reference numerals, and a description thereof will be omitted as appropriate below.

The sequence of the feature amount of the voice uttered by the user, which is output by the feature amount extraction unit 3, is divided into frames by the control unit 11
The control unit 11 supplies the feature value from the feature value extraction unit 3 to the feature value storage unit 12.

The control unit 11 refers to the word connection information stored in the word connection information storage unit 16 and controls the matching unit 14 and the reevaluation unit 15. Further, the control unit 11
Generates word connection information based on an acoustic score, a linguistic score, or the like as a matching processing result obtained by the matching unit 14 performing the same matching processing as the matching unit 4 in FIG. The storage content of the word connection information storage unit 16 is updated with the connection information. Further, the control unit 11 corrects the storage content of the word connection information storage unit 16 based on the output of the reevaluation unit 15. Further, the control unit 11 determines and outputs the final speech recognition result based on the word connection information stored in the word connection information storage unit 16.

The feature amount storage unit 12 stores a series of feature amounts supplied from the control unit 11 until, for example, a recognition result of a user's voice is obtained. The control unit 11 sets the time at which the feature amount output by the feature extraction unit 3 is obtained (hereinafter, appropriately referred to as an extraction time) with the start time of the voice section as a reference (for example, 0) together with the feature amount. , Feature amount storage unit 1
2 and the feature amount storage unit 12
The feature amount is stored together with the extraction time.

The word preliminary selecting section 13 includes a matching section 14
In response to a request from the user, the matching unit 14 selects a word to be subjected to a matching process while referring to the word connection information storage unit 16, the acoustic model database 17A, the dictionary database 18A, and the grammar database 19A as necessary. The preliminary selection processing is performed using the feature amount stored in the feature amount storage unit 12. However, in the present embodiment, the word preliminary selection unit 13 also selects, in addition to the word to be subjected to the matching processing, a word that is likely to be the word immediately after the word in the utterance, and selects the two words. Is supplied to the matching unit 14 as a result of the word preliminary selection processing.

Here, as described above, the word preliminary selection unit 13
Among the two words forming the word string output as a result of the word preliminary selection process, the first word is referred to as a first word, and the second word is referred to as a second word.

Under the control of the control unit 11, the matching unit 14 refers to the word connection information storage unit 16, the acoustic model database 17B, the dictionary database 18B, and the grammar database 19B as necessary, and selects a preliminary word selection unit. 13 is performed using the feature amount stored in the feature amount storage unit 12 for the first word of the word string obtained as a result of the word preliminary selection process from the feature amount storage unit 12, and the result of the matching process is transmitted to the control unit. 11

The re-evaluation unit 15 stores the data in the word connection information storage unit 16 while referring to the acoustic model database 17C, the dictionary database 18C, and the grammar database 19C as needed based on the control from the control unit 11. The reevaluation of the word connection information is performed using the feature amount stored in the feature amount storage unit 12, and the reevaluation result is transmitted to the control unit 11.
To supply.

The word connection information storage unit 16 stores the word connection information supplied from the control unit 11 until a recognition result of the user's voice is obtained.

Here, the word connection information indicates a connection (chain or concatenation) relationship between words constituting a word string that is a candidate for a final speech recognition result, and includes an acoustic score and a language score of each word, It also includes the start time and end time of the utterance corresponding to each word.

That is, FIG. 5 shows the word connection information stored in the word connection information storage unit 16 using a graph structure.

In the embodiment shown in FIG. 5, the graph structure as the word connection information includes an arc representing a word (in FIG. 5, a portion indicated by a line connecting circles) and a node representing a boundary between words ( (Indicated by a circle in FIG. 5).

Each node has time information, and the time information indicates the time of extraction of the feature value corresponding to the node. As described above, the extraction time is the time at which the feature amount output by the feature extraction unit 3 is obtained, with the start time of the voice section being 0, so in FIG. 5, the start of the voice section, that is, the first time, The time information of the node Node ₁ corresponding to the beginning of the word is 0. The nodes are the start and end of the arc, and the time information of the start node (start node) or the end node (end node) is the start time or end of the utterance of the word corresponding to that node, respectively. It is time.

In FIG. 5, the time from the left to the right represents the passage of time. Therefore, among the nodes on the left and right of a certain arc, the left node is the start node, and the right node is the end node. Become.

The arc has an acoustic score and a linguistic score of a word corresponding to the arc, and the arc is sequentially connected with the node serving as a terminal node as a start node, thereby enabling speech recognition. A sequence of candidate words for the result is formed.

That is, in the control section 11, first,
Is a node Node that represents the start of a voice section ₁Against the voice
Arcs corresponding to probable words are connected as recognition results
Is done. In the embodiment shown in FIG.
Arc₁Arc for "good"₆, And "Heaven
Arc for Ki₁₁Is connected. In addition, voice recognition
The matching section determines whether a word is likely to be
The acoustic and linguistic scores determined in 14
It is determined based on.

Then, similarly, the terminal node Node ₂ which is the terminal of the arc Arc ₁ corresponding to “today” and the terminal node Node which is the terminal of the arc Arc ₆ corresponding to “good”
₇ , End node No. which is the end of Arc ₁₁ corresponding to "weather"
Similarly, for each of the de ₁₂ , arcs corresponding to probable words are connected.

By connecting arcs as described above, one or more paths composed of arcs and nodes are formed from left to right starting from the start of the voice section. , When all of the paths reach the end of the voice section (time T in the embodiment of FIG. 5),
The control unit 11 accumulates, for each path formed from the start to the end of the voice section, the acoustic score and the language score of the arc constituting the path, and obtains the final score. Then, for example, a word string corresponding to an arc constituting a path having the highest final score is determined and output as a speech recognition result.

More specifically, for example, in FIG. 5, an arc Arc ₁ corresponding to “today” and a node
Node ₂ , arc Arc ₂ corresponding to “ha”, node Node ₃ ,
Arc Arc ₃ and Node ₄ corresponding to "good", Arc Arc ₄ and Node ₅ corresponding to "weather"
If the highest final score is obtained for the path composed of arc Arc ₅ corresponding to and the node Node ₆ , the word strings “today”, “ha”, “good”, “weather”,
"Issue" is output as a speech recognition result.

In the above case, arcs are always connected to nodes in the voice section to form a path extending from the start of the voice section to the end.
In the process of constructing such a path, for a path that is apparently inappropriate as a result of speech recognition based on the scores of the paths constructed up to that point, the path construction is discontinued at that point (after that, The arc is not connected to it).

Further, according to the above-described path configuration rules, the end of one arc is the start node of one or more arcs to be connected next. Is formed, except that the end of one arc coincides with the end of another arc, that is, the end node of one arc and the end node of another arc are common to the same node. In some cases.

That is, when bigrams are used as grammatical rules, if two arcs extending from different nodes correspond to the same word, and the ending time of the utterance of the word is also the same, , The ends of the two arcs coincide.

In FIG. 5, the arc Arc ₇ extending from the node Node ₇ as a starting point and the arc Arc ₁₃ extending from the node Node ₁₃ as a starting point both correspond to “weather” and have the same ending time of the utterance. Therefore, the terminal node is shared by the same node Node ₈ .

Although it is possible not to share the nodes, it is preferable to do so from the viewpoint of increasing the memory capacity.

In FIG. 5, bigrams are used as grammatical rules, but other nodes such as trigrams can be used in common.

Returning to FIG. 4, the acoustic model database 17
A, 17B, and 17C basically store the acoustic model as described in the acoustic model database 5 of FIG.

However, the acoustic model database 17B is
The acoustic model database 17C stores a high-accuracy acoustic model capable of performing processing with higher accuracy than the acoustic model database 17A.
A higher-accuracy acoustic model capable of higher-accuracy processing is stored. That is, the acoustic model database 17A
Assuming that, for example, only one pattern acoustic model independent of the context before and after is stored for each phoneme or syllable, the acoustic model database 17B
Stores, for each phoneme or syllable, for example, an acoustic model that does not depend on contexts before and after, as well as an acoustic model that depends on contexts extending between words, that is, a crossword model. The acoustic model database 17C stores, for each phoneme or syllable, for example, an acoustic model and a crossword model that do not depend on the context before and after, as well as an acoustic model that depends on the context within a word.

The dictionary databases 18A, 18B and 18C basically store word dictionaries as described in the dictionary database 6 of FIG.

That is, the dictionary databases 18A to 18C
The same set of words is registered in the word dictionary.
However, the word dictionary of the dictionary database 18B stores high-accuracy phonemic information that can be processed with higher accuracy than the word dictionary of the dictionary database 18A, and the word dictionary of the dictionary database 18C is the word dictionary of the dictionary database 18B. It stores high-accuracy phonemic information that enables more accurate processing than before. That is, for example, assuming that only one type of phonological information (reading) is registered for each word in the word dictionary of the dictionary database 18A, for example, for each word, hand,
A plurality of types of phoneme information are registered. In the word dictionary of the dictionary database 18C, for example, more phonemic information is registered for each word.

Specifically, for example, for the word “Ohashi”, only one type of phonemic information “Ohayo” is stored in the word dictionary of the dictionary database 18A.
In the word dictionary of B, in addition to "Ohayo", "Ohayo" and "Ohayo" are included, and in the word dictionary of the dictionary database 18C, "Ohayo", "Ohayo", and "Ohayo" , "Hayo" and "hayo" are each registered as phoneme information.

The grammar databases 19A, 19B and 19C basically store grammar rules as described in the grammar database 7 of FIG.

However, the grammar database 19B stores high-precision grammar rules that enable processing with higher precision than the grammar database 19A, and the grammar database 19C
Stores higher-precision grammar rules that allow more precise processing than the grammar database 19B. That is, if the grammar database 19A stores, for example, grammar rules based on unigrams (probabilities of occurrence of words), the grammar database 19B stores, for example, bigrams (probabilities of occurrence of words in consideration of the relationship with the immediately preceding word). ) Is remembered. The grammar database 19C stores grammar rules based on, for example, trigrams (occurrence probabilities of words in consideration of the relation between the immediately preceding word and the immediately preceding word) and context-free grammar.

As described above, the acoustic model database 1
7A shows an acoustic model of one pattern for each phoneme or syllable, an acoustic model database 17B shows an acoustic model of a plurality of patterns for each phoneme or syllable, and an acoustic model database 17C shows an acoustic model of each phoneme or syllable. Further, acoustic models of more patterns are stored. The dictionary database 18A stores one type of phonological information for each word.
, A plurality of types of phoneme information are stored for each word, and the dictionary database 18C stores more types of phoneme information for each word. And
The grammar database 19A stores simple grammar rules, the grammar database 19B stores high-precision grammar rules, and the grammar database 19C stores higher-precision grammar rules.

As a result, the acoustic model database 17
A, the word preliminary selection unit 13 referring to the dictionary database 18A and the grammar database 19A can obtain the acoustic score and the linguistic score quickly for many words, although the accuracy is not so high. ing. Also, an acoustic model database 17B, a dictionary database 18B, and a grammar database 19B
In the matching unit 14 that refers to a certain number of words, a high-accuracy acoustic score and a high-accuracy language score can be quickly obtained for a certain number of words. Then, the re-evaluation unit 15 referring to the acoustic model database 17C, the dictionary database 18C, and the grammar database 19C quickly targets a small number of words,
More accurate acoustic scores and language scores can be obtained.

Here, the accuracy of the acoustic model stored in each of the acoustic model databases 17A to 17C is set to be superior, but the same acoustic model is stored in each of the acoustic model databases 17A to 17C. In this case, the acoustic model databases 17A to 17C can be shared by one acoustic model database. Similarly, the storage contents of the respective word dictionaries of the dictionary databases 18A to 18C and the grammar rules of the grammar databases 19A to 19C can be the same.

Next, referring to the flowchart of FIG.
The speech recognition processing by the speech recognition device in FIG. 4 will be described.

When the user speaks, the speech as the speech is converted into digital voice data via the microphone 1 and the AD converter 2 and supplied to the feature extractor 3. The feature extracting unit 3 sequentially extracts a feature amount of a voice from the voice data supplied thereto for each frame,
It is supplied to the control unit 11.

The control unit 11 recognizes the voice section by some method. In the voice section, the control unit 11 associates a series of feature amounts supplied from the feature extraction unit 3 with the extraction time of each feature amount. Then, it is supplied to and stored in the feature amount storage unit 12.

Further, after the start of the voice section, the control unit 11 generates a node (hereinafter, appropriately referred to as an initial node) representing the start of the voice section and supplies it to the word connection information storage unit 16 in step S1. Remember. That is, in step S1, the control unit 11 sets the node No. in FIG.
de ₁ is stored in the word connection information storage unit 16.

Then, the process proceeds to a step S2, wherein the control unit 11
Determines whether there is an intermediate node by referring to the word connection information in the word connection information storage unit 16.

That is, as described above, in the word connection information shown in FIG. 5, a path extending from the start to the end of the voice section is formed by connecting the arc to the terminal node. In step S2, among the end nodes, those whose arcs have not yet been connected and which have not reached the end of the voice section are intermediate nodes (for example, nodes Node ₈ and Node ₁₀ in FIG. 5). Node
₁₁ ), and it is determined whether such an intermediate node exists.

As described above, the voice section is recognized by some method, and the time corresponding to the terminal node can be recognized by referring to the time information of the terminal node. Whether or not the unconnected end node is an intermediate node that has not reached the end of the voice section can be determined by comparing the last time of the voice section with the time information of the end node. .

If it is determined in step S2 that an intermediate node exists, the process proceeds to step S3, where the control unit 1
1 selects one of the intermediate nodes existing in the information connection information as a node that determines a word as an arc connected to the node (hereinafter, appropriately referred to as a target node).

That is, when there is only one intermediate node in the information connection information, the control unit 11 selects that intermediate node as the target node. The control unit 11
When there are a plurality of intermediate nodes in the information connection information, selects one of the plurality of intermediate nodes as a target node. Specifically, for example, the control unit 11 refers to the time information of each of the plurality of intermediate nodes, and indicates that the time indicated by the time information is the oldest (the start side of the voice section) or the newest ( Is selected as the node of interest. Alternatively, the control unit 11 may, for example,
Accumulate the acoustic score and the linguistic score of the arc constituting the path to each of the plurality of intermediate nodes,
The node at the end of the path where the accumulated value (hereinafter, appropriately referred to as a partial accumulated score) is the largest or the path where the accumulated value is the smallest is selected as the node of interest.

After that, the control unit 11 outputs to the matching unit 14 and the re-evaluation unit 15 a command to perform matching processing using the time information of the node of interest as a start time (hereinafter, appropriately referred to as a matching processing command). .

Upon receiving the matching processing command from the control unit 11, the reevaluation unit 15 proceeds to step S4, and refers to the word connection information storage unit 16 to determine the path from the initial node to the target node (hereinafter, referred to as the node). A word string (hereinafter, appropriately referred to as a partial word string) represented by an arc constituting a partial path is appropriately recognized, and the partial word string is reevaluated. That is, as described later, the partial word string is a word string that is a candidate for a speech recognition result obtained by performing a matching process on the word preselected by the word preselection unit 13 by the matching unit 14. Although it is an intermediate result, the re-evaluation unit 15 determines that the intermediate result is
Be evaluated.

More specifically, the re-evaluation unit 15 reads out a series of feature amounts corresponding to the partial word strings from the feature amount storage unit 12 in order to recalculate the language score and the acoustic score for the partial word strings. In other words, the re-evaluation unit 15 determines a sequence of feature amounts (feature amounts) associated from the time represented by the time information of the initial node, which is the first node of the partial path, to the time represented by the time information of the node of interest. series)
Is read from the feature amount storage unit 12. Further, the reevaluation unit 15 refers to the acoustic model database 17C, the dictionary database 18C, and the grammar database 19C,
Using the feature amount sequence read from the feature amount storage unit 12,
The language score and the acoustic score are recalculated for the partial word string. This recalculation is performed without fixing the word boundaries of each word constituting the partial word string. Therefore, the reevaluation unit 15 recalculates the linguistic score and the acoustic score of the partial word string, and determines the word boundaries of each word constituting the partial word string based on the dynamic programming. Will be

As described above, the re-evaluation unit 15 obtains a new language score and acoustic score of each word in the partial word string and a new word boundary, and uses the new language score and acoustic score to obtain word connection information. The language score and the acoustic score of the arc constituting the partial path corresponding to the partial word string in the storage unit 16 are corrected, and the partial path corresponding to the partial word string in the word connection information storage unit 16 is changed by the new word boundary. Correct the time information of the constituent nodes. In the present embodiment, the correction of the word connection information by the reevaluation unit 15 is performed via the control unit 11.

That is, for example, when the node Node ₅ shown in FIG. 7 is set as the target node, for example, of the partial paths from the initial node Node ₁ to the target node Node ₅ ,
Node Node _3, the arc Arc ₃ corresponding to the word "good", the node Node _4, word word string "good" represented by part of Arc _4, and the node Node ₅ corresponding to "weather", attention to "weather" Then, the reevaluation unit 15 refers to the acoustic model database 17C and the dictionary database 18C using the feature amount sequence from the time corresponding to the node Node ₃ to the time corresponding to the node Node ₅ , and the word “good”, The word model of each "weather" is constructed, and the acoustic score is calculated. Further, the re-evaluation unit 15 checks the
By referring to C, the language score of each of the words “good” and “weather” is calculated. Specifically, for example, when a grammar rule based on a trigram is stored in the grammar database 19C, the reevaluation unit 15 determines that the word “good” is a word “ha” immediately before the word “good”, and further, Using the word "Today" before it, the words "Today", "Ha",
The probability of being linked to "good" is obtained, and the language score is calculated based on the probability. In addition, the re-evaluation unit 15 uses the word “good” immediately before the word “weather” and the word “ha” immediately before the word “weather”, and the words “ha”, “good”,
A probability linked to “weather” is obtained, and a language score is calculated based on the probability.

The re-evaluation unit 15 accumulates the acoustic score and the linguistic score obtained as described above, and determines the word boundary between the words “good” and “weather” so that the accumulated value becomes the largest. Then, the re-evaluation unit 15 uses the acoustic score and the language score thus obtained to
Arc Arc ₃ corresponding to the word "good" and Arc ₄ corresponding to the word "weather" have their acoustic and linguistic scores corrected, and the determined word boundaries determine
Node No. corresponding to the word boundary between the words "good" and "weather"
The time information of de ₄ is corrected.

Therefore, the re-evaluation unit 15 determines the word boundaries between the words in the partial word string based on the dynamic programming, and replaces the word connection information stored in the word connection information 16 with:
It will be corrected sequentially. Then, since the processing is performed in the word preliminary selection unit 13 and the matching unit 14 with reference to the corrected word connection information, the accuracy and reliability of the processing in each of them can be improved.

Furthermore, since the word boundary of the word connection information is corrected in the reevaluation unit 15, the number of word boundary candidates stored as word connection information can be significantly reduced.

[0122] That is, conventionally, for example, as described in Figure 2 above, as candidates for word boundaries of words between "kyou" and "wa", the time _{t 1-1, t 1, t 1} + 1 Must be held, and if the time t ₁ , which is a correct word boundary, is not maintained, the subsequent matching process is adversely affected. In contrast, in the case of sequentially correcting the word boundary in the re-evaluation section 15, for example, even if not only hold only one time t _1-1 is incorrect word boundaries, the re-evaluation section 15 , the erroneous time t _1-1 is a word boundary, since is modified at time t ₁ is the correct word boundary, it does not adversely affect the subsequent matching processing.

The re-evaluation unit 15 calculates an acoustic score using a crossword model that takes into account the words connected before and after the words constituting the partial word string except for the first and last words. Further, the calculation of the language score can be performed in consideration of the words connected before and after the language score, respectively, and therefore, high-precision processing can be performed. Moreover, since the processing of the re-evaluation unit 15 is performed sequentially, a large delay unlike the above-described two-pass decoding does not occur.

Returning to FIG. 6, when the re-evaluation unit 15 finishes correcting the word connection information in the word connection information storage unit 16 as described above, the re-evaluation unit To supply.

As described above, after receiving the matching processing command from the control unit 11, the matching unit 14 receives from the reevaluation unit 15 via the control unit 11 that the word connection information has been corrected. , The noticed node, and the time information of the noticed node are supplied to the word preliminary selection unit 13 to request a word preliminary selection process.

The word preliminary selecting section 13
Receives a request for word preliminary selection processing from step S5, in step S5, a word string consisting of a first word and a second word as candidates for an arc word to be connected to the node of interest and subsequent word candidates, respectively. In other words, the first word that is directly targeted for the matching process, and the second word that follows
A word pre-selection process of selecting a word string composed of words is performed on words registered in the word dictionary of the dictionary database 18A.

That is, the word preliminary selection unit 13 recognizes the start time of a series of feature amounts used for calculating the language score and the acoustic score from the time information of the node of interest,
A required feature amount series after the start time is read from the feature amount storage unit 12. Further, the word preliminary selection unit 13
The word model of each word registered in the word dictionary of the dictionary database 18A is configured by connecting the acoustic models stored in the acoustic model database 17A, and further, for example, by connecting any two word models, A word string model (word string model) composed of the first word and the second word is configured. Then, the word preliminary selection unit 13 calculates an acoustic score for each of the first word and the second word of the word string corresponding to each word string model, using the feature amount sequence read from the feature amount storage unit 12. .

Further, the word preliminary selection section 13 calculates the language score of the word string corresponding to each word string model based on the grammar rules stored in the grammar database 19A. That is, the word preliminary selection unit 13 obtains a linguistic score for the first word of the word string based on, for example, a unigram, and for the second word, for example, the first word and the second word are linked. Based on the bigram that defines the probability,
Find the language score.

In the word preliminary selecting section 13, the calculation of the acoustic score of the first word or the second word of the word string can be performed by using the second word or a crossword model depending on the first word, respectively. It is. Further, the word preliminary selection unit 13 calculates the acoustic score of the first word by referring to the word connection information, and calculates the acoustic score of the word immediately before the first word (the word corresponding to the arc ending at the focused node). ) Can be performed using a crossword model that depends on the above.

Further, the word preliminary selection unit 13 refers to the word connection information to calculate the language score of the first word and determine the probability that the word immediately before the first word is linked to the first word. Or a trigram defining the probability of the first word being chained, the word immediately before the first word, the word immediately before the first word, and the probability that the first word is linked. Similarly, the calculation of the linguistic score of the second word also includes the word immediately before the second word (the word immediately before the first word), the first word that is the word immediately before the second word, and the second word. This can be performed based on a trigram that defines the probability of chaining.

When the word preliminary selection unit 13 obtains an acoustic score and a language score for each word string including the first word and the second word as described above, the word preliminary selecting unit 13 calculates a score obtained by comprehensively evaluating the acoustic score and the language score. , Below, as appropriate,
(Referred to as a word string score), and the upper L words are supplied to the matching unit 14 as a word string used in the matching process.

The matching section 14 includes the word preliminary selecting section 13
When the L word strings (hereinafter, appropriately referred to as selected word strings) used for the matching process are received from step S6, in step S6, the first word of each selected word string is subjected to the cross depending on the preceding and succeeding words. Perform a matching process using a word model.

That is, the matching unit 14 recognizes, from the time information of the node of interest, the start time of a series of feature amounts used for calculating the language score and the acoustic score, and calculates the necessary feature amount after the start time. Are read from the feature amount storage unit 12. Further, the matching unit 14 recognizes the phoneme information of the first word of the selected word string by referring to the dictionary database 18B, reads out the acoustic model corresponding to the phoneme information from the acoustic model database 17B, and connects it. Form a word model of the first word.

Here, the matching unit 14 uses, as an acoustic model corresponding to each phoneme (or syllable) of the first word, a model that depends on the phonemes (or syllables) before and after it.
Construct a word model of the first word. Therefore, the matching unit 14 refers to the word connection information for the first phoneme of the first word, and uses the crossword model that depends on the last phoneme of the word immediately before the first word to obtain the first word. Is constructed. Further, in the matching unit 14, for the last phoneme of the first word, a word model of the first word is configured using a crossword model that depends on the first phoneme of the second word that is the word immediately after the first phoneme. .

As described above, for example, when the first word is “good morning” and the second word is “is”, the last phoneme / u / of the first word is replaced by the phoneme / o immediately before it. / (The vowel part of "yo") and the phoneme / g /
(The consonant part of “go”),
A word model of the first word "good morning" is constructed using an acoustic model that represents the last phoneme / u / of the first word with higher accuracy.

[0136] The matching unit 14 is based on the word model constructed as described above,
The acoustic score of the first word is calculated by using the feature amount sequence read from.

The matching unit 14 calculates the language score of the first word by referring to the grammar database 19B in consideration of words before and after the first word. That is, the matching unit 14 recognizes the word immediately before the first word and the word before that by referring to the word connection information, for example, and recognizes the word two words before the first word and the word before the first word. The probability based on the immediately preceding word and the trigram in which the first word is chained, and the probability based on the bigram in which the first word and the second word are chained are determined. Further, the matching unit 14
Calculates the language score of the first word from the probabilities based on these trigrams and the probabilities based on bigrams.

As described above, the matching unit 14
For all of the first words in the L selected word strings from the word preliminary selection unit 13, their acoustic scores and language scores are obtained, and the process proceeds to step S7. In step S7, for each of the first words in the L selected word strings, a word score obtained by comprehensively evaluating the acoustic score and the linguistic score is obtained, and based on the word score, the word connection information storage unit 1
The word connection information stored in No. 6 is updated.

That is, in step S7, the matching unit 1
4 obtains a word score for the first word of each selected word string, and compares the word score with a predetermined threshold to determine a word as an arc connected to the node of interest from among the first words. Refine. Then, the matching unit 14 supplies the first word remaining as a result of the narrowing down to the control unit 11 together with the acoustic score, the language score, and the end time of the word.

Note that the end time of a word is recognized from the extraction time of the feature amount used for calculating the acoustic score. When a plurality of extraction times having a high probability as the end time are obtained for a certain word, for the word, a set of each end time and the corresponding acoustic score and language score is stored in the control unit 11. Supplied to

Upon receiving the acoustic score, language score, and end time of the first word supplied from the matching unit 14 as described above, the control unit 11
From the target node in the word connection information (FIG. 5) stored in the word connection information storage unit 16 for each first word, the arc is extended, and the arc is set to the terminal node corresponding to the position of the end time. Connect to Further, the control unit 11 determines, for each arc, a corresponding word,
In addition to providing the acoustic score and the language score, a corresponding end time is given as time information to the terminal node of each arc. Then, the process returns to step S2, and thereafter, the same processing is repeated.

As described above, the word connection information is successively updated based on the processing result of the matching unit 14, and is successively corrected by the reevaluation unit 15, so that the word preliminary selection unit 13 and the matching unit 14 The processing can be always performed using the word connection information.

When updating the word connection information, the control unit 11, if possible, shares the terminal nodes as described above.

On the other hand, if it is determined in step S2 that there is no intermediate node, the process proceeds to step S8,
The control unit 11 obtains a final score by referring to the word connection information and accumulates a word score for each path configured as the word connection information. For example, a path having the highest final score is formed. A word string corresponding to the arc to be output is output as a speech recognition result for the utterance of the user, and the process ends.

As described above, the word preliminary selection section 13 selects not only the first word which is a candidate for the word following the intermediate node but also the second word which is likely to follow thereafter. For words, based on the second word,
A highly accurate acoustic score or language score can be calculated, and as a result, speech recognition accuracy can be improved.

Next, the above-described series of processing can be performed by hardware or software. When a series of processing is performed by software, a program constituting the software is
Installed on a general-purpose computer.

FIG. 8 shows an example of the configuration of an embodiment of a computer in which a program for executing the above-described series of processing is installed.

The program is stored in a hard disk 105 or a ROM 1 as a recording medium built in the computer.
03 can be recorded in advance.

Alternatively, the program may be a floppy disk, CD-ROM (Compact Disc Read Only Memory), M
O (Magneto optical) disc, DVD (Digital Versatile)
Disc), a magnetic disk, a semiconductor memory, or another such removable storage medium 111, which can be temporarily or permanently stored (recorded). Such a removable recording medium 111 can be provided as so-called package software.

The program can be installed in the computer from the removable recording medium 111 as described above, can be wirelessly transferred from a download site to the computer via a digital satellite broadcasting artificial satellite, or can be transmitted to a LAN (Local Area). Network), the Internet, and the like, and can be transferred to a computer by wire. In the computer, the transferred program can be received by the communication unit 108 and installed on the built-in hard disk 105.

The computer has a CPU (Central Processing).
Unit) 102. The CPU 102 has a bus 1
01, the input / output interface 110 is connected. The CPU 102 operates the input / output unit 107 including a keyboard, a mouse, a microphone, and the like by the user via the input / output interface 110. When a command is input, the ROM (Read O
nly Memory) 103 is executed. Alternatively, the CPU 102 may execute a program stored in the hard disk 105, a program transferred from a satellite or a network, received by the communication unit 108 and installed in the hard disk 105, or a removable recording medium 111 mounted in the drive 109. The program read and installed on the hard disk 105 is stored in a RAM (Random Access Memory).
y) Load to 104 and execute. As a result, the CPU 10
2 performs processing according to the above-described flowchart or processing performed by the configuration of the above-described block diagram. Then, the CPU 102 transmits the processing result as necessary, for example, via the input / output interface 110.
An output is made from an output unit 106 including an LCD (Liquid CryStal Display), a speaker, or the like, or transmitted from a communication unit 108, and further recorded on the hard disk 105.

Here, in this specification, processing steps for describing a program for causing a computer to perform various processes do not necessarily have to be processed in chronological order in the order described in the flowchart, and may be performed in parallel. Alternatively, it also includes processing executed individually (for example, parallel processing or processing by an object).

The program may be processed by one computer, or may be processed in a distributed manner by a plurality of computers. Further, the program may be transferred to a remote computer and executed.

Since the word (first word) to be scored by the matching unit 14 has been selected in advance by the word preliminary selection unit 13, the matching unit 14
Can be independently performed for each word without configuring a tree-structured network that shares a part of the calculation of the acoustic score as described above.
In this case, the memory capacity reserved for the matching unit 14 to calculate a score for each word can be reduced. Furthermore, in this case, when the score calculation of a word is started, it is possible to identify which word the word is. Therefore, as described above, the useless calculation cannot be performed because the word cannot be identified. Can be prevented.

The score calculation by the matching unit 14 and the re-evaluation unit 15 can be performed independently for each word in terms of time. In this case, by using the memory capacity required for the score calculation, Can be kept small.

The voice recognition apparatus shown in FIG. 4 is used, for example, to search a database by voice, to operate various devices, to input data to each device, to use a voice interactive system, and the like. Applicable to More specifically, for example, in response to an inquiry about a place name by voice,
A database search device that displays the corresponding map information,
The present invention can be applied to an industrial robot that sorts luggage in response to a voice command, a dictation system that creates text by voice input instead of a keyboard, and a dialog system of a robot that talks with a user.

Further, in the present embodiment, the word preliminary selection unit 13 selects a word string composed of two words (first word and second word).
In 3, it is also possible to select a word string consisting of three or more words. In this case, the matching unit 14 can calculate the language score of the first word of the word string using a grammar rule (language model) with higher accuracy based on two or more words that follow. .

Further, the number of words constituting the word string selected by the word preliminary selection section 13 does not need to be fixed, but can be variable.

[0159]

According to the speech recognition apparatus, the speech recognition method, and the recording medium of the present invention, a word consisting of a plurality of words connected to a word whose score has been calculated from a group of words to be subjected to speech recognition. A row is selected, and a score is calculated for the first word, which is the first word of the word string, based on one or more words following the first word of the word string. Then, the connection relation between the head word for which the score has been calculated and the word to which the head word is connected and for which the score has already been calculated is stored. Based on the connection relation and the score, the speech recognition result is obtained. Is determined. Therefore, for example, an acoustic score or a linguistic score for a word can be calculated in consideration of the immediately following word, and as a result, the accuracy of speech recognition can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of an example of a conventional voice recognition device.

FIG. 2 is a diagram illustrating the necessity of holding a candidate for a boundary between words;

FIG. 3 is a block diagram showing a configuration of another example of the conventional voice recognition device.

FIG. 4 is a block diagram illustrating a configuration example of a voice recognition device according to an embodiment of the present invention;

FIG. 5 is a diagram for explaining word connection information.

FIG. 6 is a flowchart for explaining processing of the voice recognition device of FIG. 4;

FIG. 7 is a diagram for explaining a process of a reevaluation unit 15;

FIG. 8 is a block diagram illustrating a configuration example of a computer according to an embodiment of the present invention.

[Explanation of symbols]

1 microphone, 2 AD converter, 3 feature extractor,
11 control unit, 12 feature amount storage unit, 13 word preliminary selection unit, 14 matching unit, 15 reevaluation unit,
16 word connection information storage unit, 17A to 17C acoustic model database, 18A to 18C dictionary database, 19A to 19C grammar database, 1
01 bus, 102 CPU, 103 ROM, 104
RAM, 105 hard disk, 106 output unit, 107 input unit, 108 communication unit, 109 drive, 110 input / output interface, 111 removable recording medium

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Koji Asano, Inventor 6-7-35 Kita-Shinagawa, Shinagawa-ku, Tokyo Inside Sony Corporation (72) Inventor Helmut Lucke 6-35, Kita-Shinagawa, Shinagawa-ku, Tokyo Sony Corporation F term (reference) 5D015 AA05 BB01 HH11 HH16 9A001 HH17

Claims (7)

[Claims] 1. A speech recognition apparatus for calculating a score representing the likelihood of a speech recognition result for an input speech, and recognizing the speech based on the score. Word string selecting means for selecting a word string consisting of a plurality of words connected to the word for which the score has been calculated, from a group of words to be executed, and for the first word which is the first word of the word string, Score calculating means for calculating the score based on one or more words following the head word; and a head word for which the score has been calculated and a word to which the head word is connected and for which the score has been calculated. A speech recognition apparatus comprising: a connection relationship storage unit that stores a connection relationship between the two; and a determination unit that determines a voice recognition result of the voice based on the connection relationship and the score. 2. The method according to claim 1, wherein the score calculating means calculates an acoustic or linguistic score for the first word based on at least one word following the first word in the word string. The speech recognition device according to the above. 3. The speech recognition apparatus according to claim 2, wherein said score calculating means calculates said linguistic score with reference to a probabilistic language model. 4. The speech recognition apparatus according to claim 1, further comprising storage means for storing a group of words to be subjected to the speech recognition. 5. The apparatus according to claim 1, further comprising an extraction unit configured to extract a feature amount of the voice, wherein the score calculation unit calculates the score using the feature amount. Voice recognition device. 6. A speech recognition method for recognizing a speech based on the score, wherein the score representing the likelihood of the speech recognition result is calculated for the inputted speech. A word string selection step of selecting a word string composed of a plurality of words connected to the word for which the score has been calculated from the set of words to be executed; and for the first word that is the first word of the word string, A score calculating step of calculating the score based on one or more words following the head word; and a head word for which the score has been calculated and a word to which the head word is connected and for which the score has been calculated. A connection relationship storing step of storing a connection relationship between, and a determination step of determining a voice recognition result of the voice based on the connection relationship and the score.識方 method. 7. A program for calculating a score representing the likelihood of a speech recognition result for an input speech, and recording a program for causing a computer to perform speech recognition processing for recognizing the speech based on the score. A word string selection step of selecting, from a group of words to be subjected to speech recognition, a word string consisting of a plurality of words connected to the word for which the score has been calculated; and A score calculation step of calculating the score based on one or more words following the first word of the word string, for the first word that is the first word of the first word; A connection relation storing step of storing a connection relation between a connected word and the word for which the score has been calculated; and a sound of the sound based on the connection relation and the score. Recording medium comprising a program and a determination step for determining the identification result is recorded.