JPH08221091A - Voice recognition device - Google Patents

Abstract

PURPOSE: To provide a device capable of using the whole situational limitation and with less storage capacity when a language score based on the probability that a present word is generated is calculated from a beforehand recognized preceding word system by using all of preceding words and using the conditional probability obtained by the operation among the generative probability, the syndetic co-occurrence probability and the non-syndetic co-occurrence probability. CONSTITUTION: From the beforehand recognized word system preserved in a word system preservation part 4, the generative probability of the word stored in a generative probability storage part 7, the syndetic co-occurrence probability stored in a syndetic co-occurrence probability storage part 8 and the non- syndetic co-occurrence probability stored in a non-syndetic co-occurrence probability storage part 9, the language score is calculated in a language score calculation part 5. A recognition part 3 receives the language score from the language score calculation part 5, and calculates a total score by an acoustic score calculated by the recognition part 3 and the language score, and it transfers the word system with the high total score to the word system preservation part 4. The word system preservation part 4 outputs the preserving word system as the recognition result after an input of a voice is ended.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for recognizing a voice in which words are continuously uttered.

[0002]

2. Description of the Related Art Recently, attempts have been made to improve speech recognition performance by using a statistical language model. An example of such a voice recognition device is, for example, Murakami et al.
Sentence speech recognition using trigram and extension to free speech recognition "Technical report of the Institute of Electronics, Information and Communication Engineers, SP93-127 (1994) can be cited. As a statistical language model, a bigram using probability for two consecutive words, There are trigrams using probabilities related to 3 consecutive words, tetragrams using probabilities related to 4 consecutive words, etc. In the left-to-right processing along the time axis of the input, the bigram is composed of the preceding one word. The trigram is a constraint from the preceding two words, the tetragram is a constraint from the preceding three words, and when the bigram is used as the statistical language model, the constraint is too loose.
The recognition result contains many non-sentences that do not hold as Japanese sentences. In order to strengthen the constraint, a global constraint such as trigram, tetragram, ... May be used. On the other hand, in terms of the required storage capacity, a bigram requires a storage capacity of the vocabulary number of the word included in the recognition target squared, a trigram has a storage capacity of the vocabulary number of the cube 3, and a tetragram requires a storage capacity of the vocabulary number of the fourth power. . Currently, Trigram is used as a statistical language model due to the balance between the strength of the constraint and the required memory capacity.

A conventional speech recognition apparatus using a trigram as a statistical language model will be described below.

FIG. 3 shows a schematic structure of a conventional speech recognition apparatus. Reference numeral 1 is a characteristic parameter calculation unit for converting an input speech into a time series of characteristic parameters, and 2 is a word prepared in advance from learning data. A word standard pattern storage unit for storing a standard pattern, 3 is a recognition unit for calculating an overall score for a word sequence based on an acoustic score as a similarity between a feature parameter and a word standard pattern, and a language score by a language score calculation unit, and 4 is a recognition unit. A word sequence storage unit for storing a word sequence having a high total score, 5 is a language score calculation unit for calculating a score based on the probability that the word will occur next from an already recognized preceding word sequence, and 6 is a word score calculation unit. Preceding 2
It is a trigram storage unit that stores the probability of occurring immediately after a word.

The operation of the voice recognition device will be briefly described below. The input voice is converted into the characteristic parameter calculation unit 1
In the recognition unit 3, a matching process is performed between the feature parameter received from the feature parameter calculation unit 1 and the word standard pattern stored in the word standard pattern storage unit 2 (for example, LPC cepstrum). , DP matching) to calculate an acoustic score. On the other hand, the language score calculation unit 5 calculates a language score from the already recognized word sequence stored in the word sequence storage unit 4 and the trigram stored in the trigram storage unit 6, and the recognition unit 3 Is the language score calculator 5
The language score is received from, the total score based on the acoustic score and the language score is calculated, and the word sequence having a high total score is passed to the word sequence storage unit 4. The word series storage unit 4 outputs the stored word series as a recognition result after the voice input is completed.

Next, the method of calculating the language score in the language score calculator 5 will be described in more detail.

The word sequence uttered as an input is w ₁ , w ₂ ,
When w ₃ , w ₄ , w ₅ , w ₆ , the language score for this word sequence is

[0008]

[Equation 1]

For example, it is represented by a logarithmic value. Where P
(w₁, w₂, w₃, w_Four, w_Five, w₆) Is the word sequence w₁, w₂, w₃, w_Four, w_Five, w₆Raw
Occurrence probability, which is the probability of occurrence of the first word P (w₁) And No. 1
The second word is generated under the condition that the eye word is generated.
Conditional probability P (w₂| w₁) And the first and second
3rd term on the condition that the word sequence of
The conditional probability P (w₃| w₁, w₂) And 1
The word sequence from the 1st word to the 3rd word occurred
Conditional probability that the fourth word will occur under the conditions
P (w_Four| w₁, w₂, w₃) And the first to fourth words
The fifth word occurs under the condition that the sequence occurs
Conditional probability P (w_Five| w₁, w₂, w₃, w_Four) And the first
With the condition that the word sequence from the word to the fifth has occurred
Conditional probability P (w that the sixth word below occurs₆| w
₁, w₂, w₃, w_Four, w_Five) And the product. And Trigra
When using the program, the probability of occurrence of the first word P (w₁) And No. 1
The second word is generated under the condition that the eye word is generated.
Conditional probability P (w₂| w ₁) And the first word
3 under the condition that the second word sequence has occurred
Conditional probability P (w₃| w₁, w₂)
And the word sequence from the second word to the third word occurred
Conditional that the fourth word occurs under the condition
Probability P (w_Four| w₂, w₃) And the 3rd word through the 4th
The fifth word is generated under the condition that the word sequence has occurred.
Conditional probability P (w_Five| w₃, w_Four) And the fourth simple
Under the condition that the word sequence from the fifth word to the fifth word occurs
Conditional probability that the sixth word occurs in P (w₆| w_Four,
w_Five) And the product. According to this formula,
In the middle stage, already recognized word sequences (trigram
In this case, the last two words of the preceding words are said to have been spoken.
The probability that the current word will occur immediately after
For example, if the fourth word is P (w_Four| w₂, w
₃), And if it is the fifth word as shown in FIG. 4 (b), P (w_Five| w₃, w_Four)
However, it works as a linguistic constraint.

[0010]

However, the above-mentioned conventional configuration has a problem that the global constraint cannot be used because the word that is more than two words before cannot be used as the constraint for the current word. Further, when trying to solve this problem by using tetragrams, pentagrams, ... Instead of trigrams, the storage capacity of the fourth, fifth, ... This cannot be achieved when targeting a large number of vocabularies.

The present invention solves the above-mentioned conventional problems, and an object of the present invention is to provide a voice recognition device which can utilize global restrictions and has a small storage capacity.

[0012]

To achieve this object, the speech recognition apparatus of the present invention calculates a language score based on the probability that the next current word will occur from the already recognized preceding word sequence. At this time, the language score calculation unit is characterized by using all the preceding words and using conditional probabilities obtained by calculation of the occurrence probability, the concatenated co-occurrence probability, and the non-concatenated co-occurrence probability.

[0013]

With this configuration, the influence from all the preceding words is reflected in the recognition score of the current word, so global constraints can be used, and the necessary information is the square of the vocabulary of the recognition target word. Since it can be stored with the storage capacity of, a voice recognition device with a small storage capacity can be realized.

[0014]

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

FIG. 1 is a block diagram showing a schematic configuration of an embodiment of a voice recognition device according to the present invention. In FIG. 1, 1 is a feature parameter calculation unit that converts input speech into a time series of feature parameters, 2 is a word standard pattern storage unit that stores a word standard pattern created from learning data in advance, and 3 is a feature parameter. A recognition unit that calculates an overall score for a word sequence based on an acoustic score, which is the degree of similarity to a word standard pattern, and a linguistic score calculated by a linguistic score calculation unit, 4 is a word sequence storage unit that stores a word sequence with a high overall score, and 5 is When calculating a language score based on the probability that the word will occur next from the already recognized preceding word sequence, all of the preceding words are used, and the occurrence probability, the concatenated co-occurrence probability, and the non-concatenated co-occurrence probability are used. The language score calculation unit is characterized by using the conditional probability obtained by the calculation of, the occurrence probability storage unit for storing the probability of occurrence of a certain word, and the reference numeral 8 Is a concatenated co-occurrence probability storage unit that stores the probability of occurring immediately after another certain word, and 9 is a non-concatenated co-occurrence probability storage unit that stores the probability that the word occurs two or more times behind another certain word. is there.

The operation of the speech recognition apparatus configured as above will be briefly described. The input voice is converted into a characteristic parameter (for example, LPC cepstrum) by the characteristic parameter calculation unit 1, and the recognition unit 3 stores the characteristic parameter received from the characteristic parameter calculation unit 1 and the word standard pattern storage unit 2. By performing a matching process (for example, DP matching) with a certain word standard pattern,
Calculate the acoustic score. On the other hand, the already-recognized word sequence stored in the word sequence storage unit 4 and the occurrence probability storage unit 7
Occurrence probability and concatenated co-occurrence probability storage unit 8 stored in
From the concatenated co-occurrence probability stored in and the non-concatenated co-occurrence probability stored in the storage unit, the language score calculation unit 5 calculates a language score, and the recognition unit 3 calculates the language score. The language score is received from the unit 5, the total score is calculated (eg, sum or linear sum) from the acoustic score and the language score, and the word sequence having a high total score is passed to the word sequence storage unit 4. The word series storage unit 4 outputs the stored word series as a recognition result after the voice input is completed.

The method of calculating the language score in the language score calculation unit 5 of the speech recognition apparatus operating as described above will be described in more detail.

The word sequence uttered as an input is w ₁ , w ₂ ,
When w ₃ , w ₄ , w ₅ , w ₆ , the language score for this word sequence is

[0019]

[Equation 2]

For example, it is represented by a logarithmic value. Where P
(w ₁ , w ₂ , w ₃ , w ₄ , w ₅ , w ₆ ) is the occurrence probability of the word sequence w ₁ , w ₂ , w ₃ , w ₄ , w ₅ , w ₆ and is the first word Occurrence probability P (w ₁ ) and conditional probability P (w ₂ | w ₁ ) that the second word occurs under the condition that the first word occurs, and the first word to the second 3 under the condition that the word sequence of
Conditional probability P (w ₃ | w ₁ , w ₂ ) that the th word occurs
And the conditional probability P (w ₄ | w ₁ , w ₂ , w ₃ ) that the 4th word occurs under the condition that the 1st word to the 3rd word sequence occur, and the 1st word Conditional probability P (w ₅ | w ₁ , w ₂ , w ₃ , w ₄ ) that the fifth word occurs under the condition that the fourth to the fourth word sequence occurs, and 1
Conditional probability that the 6th word occurs under the condition that the 5th word sequence up to the 5th word occurs
It is represented by the product of P (w ₆ | w ₁ , w ₂ , w ₃ , w ₄ , w ₅ ). Here, it is assumed that the influence of the previously recognized preceding word sequence on the current word independently influences each of the preceding words to the current word, and further, each word of the preceding word sequence changes to the current word. Influence from the immediately preceding word (represented by the concatenated co-occurrence probability P _con (w _i | w _j )) and the influence from two or more previous words (non-concatenated co-occurrence probability P _sep (w _i | w _{j It} can be approximated by the following formula from Bayes' theorem.

[0021]

(Equation 3)

Here, P _con (w _i | w _j ) is the probability that the word w _i occurs immediately after the word w _j , and P _sep (w _i | w _j ) is two or more words behind the word w _j. The probability that the word w _i occurs, P (w _i ) is the probability that the word w _i occurs.

The difference from the conventional example is that, in the middle of the recognition, the already recognized word sequence is uttered,
Immediately thereafter, when using the trigram of the conventional example when obtaining a language score based on the probability that the word w _i occurs,

[0024]

[Equation 4]

According to the present embodiment,

[0026]

(Equation 5)

It is approximated by For example, in this embodiment, as shown in FIG.
As in (a), the language score for the fourth word is 1
Non-joint co-occurrence probability P _{s ep} (w ₄ | w ₁ ) that the fourth word occurs after the second word and non-joint co-occurrence probability P that the fourth word occurs after the second word The product of _{se p} (w ₄ | w ₂ ) and the concatenation co-occurrence probability P _con (w ₄ | w ₃ ) in which the fourth word occurs immediately after the fourth word immediately after the third word is the fourth Is the logarithmic value of the value obtained by dividing the occurrence probability P (w ₄ ) of the word by. Similarly, as shown in FIG. 2B, the linguistic score for the fifth word is the non-joint co-occurrence probability P _sep (w ₅ | w ₁ ) in which the _fifth word occurs after the first word. Non-joint co-occurrence probability P _sep (w ₅ | w ₂ ) in which the fifth word occurs after the second word and non-joint co-occurrence probability P _{sep in} which the fifth word occurs after the third word The product of (w ₅ | w ₃ ) and the concatenated co-occurrence probability P _con (w ₅ | w ₄ ) that occurs immediately after the fourth word, P _con (w ₅ | w ₄ ), is the occurrence probability P (w _{5 of the fifth} word. ) Is the logarithmic value of the value divided by the third power.

The number of sentences to be recognized by the speech recognition apparatus according to this embodiment (sentence and word sequence have the same meaning) and the number of sentences to be recognized by the conventional speech recognition apparatus are shown in comparison with (Table 1). .

[0029]

[Table 1]

This (Table 1) shows a trigram in the conventional example from the data of 162 sentences collected from conversations at the airport such as "There is nothing to declare.""Where is the counter at ANA?" In the present embodiment, concatenated co-occurrence probabilities, non-concatenated co-occurrence probabilities, and occurrence probabilities are obtained, all word sequences whose occurrence probabilities are not 0 with respect to the word sequences obtained according to the recognition method are obtained, and the number of words included in the word sequence It is a summary.

It is desirable that the number of sentences recombined by the information (trigram, concatenation co-occurrence probability, etc.) extracted from the example sentences is small, and the closer to the number of sentence examples, the better. This is because the smaller the number of recognition targets, the less the possibility of erroneous recognition, and the more advantageous the recognition.

As is clear from this (Table 1), the number of sentences to be recognized is closer to the number of example sentences in this embodiment,
It works strongly as a constraint to limit the recognition target. Looking at the increased number of sentences, which represents the excessive generation of recognition target sentences, the number of 51 sentences in the present embodiment is suppressed to less than half of the 119 sentences in the conventional example.

Looking at the contents of the overgenerated sentence,
You can see the following. In the conventional example, since the constraint is a local constraint, even a sentence that does not hold globally can be a target of recognition if the local constraint is satisfied. For example, "Please bring this baggage to the taxi stand to the ANA counter.""Where does the money exchange come from?""What time does it start in New York?" 55 like this
Non-sentences (sentences that do not hold as Japanese) are included in the recognition target. On the other hand, most of the sentences other than the example sentences in the recognition target of the present embodiment are sentences that are considered to have omitted a certain word sequence (for example, "How about New York?"), And non-sentences that have no meaning at all are "airplane". Where can I get the ticket? ”" Where can I get the ticket? " As described above, the reason why the number of non-sentences is small in this embodiment is that only words that co-occur with all preceding words can be connected to the preceding word series, and therefore the sentence to be recognized is a word pair extracted from one sentence. Consists of co-occurring word pairs, that is, the words in the sentence can be expected to be a semantically consistent group of words, and
This is because all the two consecutive words extracted from one sentence are the two consecutive words that can exist, so that it can be expected that the sentence is syntactically correct.

By the way, in terms of storage capacity, recognition
Since the number of vocabulary of words in the target is 256, the conventional example is 25
6³= 16777216 storage capacity is required, but in this embodiment
Is 256 ²× 2 + 256 = 131328, 1/100 or less of the conventional example
It is enough if there is enough storage capacity.

As described above, according to this embodiment, all the preceding words are used when calculating the language score based on the probability that the next current word will occur from the already recognized preceding word sequence. , A global score constraint can be utilized by providing a language score calculator characterized by using conditional probabilities obtained by calculation of occurrence probabilities, concatenated co-occurrence probabilities, and non-concatenated co-occurrence probabilities. It is possible to realize an excellent voice recognition device having a small capacity.

[0036]

INDUSTRIAL APPLICABILITY As described above, the present invention utilizes all the preceding words when calculating the language score based on the probability that the current word will occur next from the already recognized preceding word sequence, By providing a linguistic score calculation unit characterized by using conditional probabilities obtained by the calculation of the occurrence probability, the concatenated co-occurrence probability, and the non-concatenated co-occurrence probability, global constraints can be used and the memory capacity can be increased. It is possible to realize an excellent voice recognition device with a small amount of noise.

[Brief description of drawings]

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

FIG. 2A is a conceptual diagram of an operation example when recognizing a fourth word according to the embodiment. FIG. 2B is a conceptual diagram of an operation example when recognizing a fifth word according to the embodiment.

FIG. 3 is a schematic block diagram showing a conventional voice recognition device.

FIG. 4A is a conceptual diagram of an operation example when recognizing a fourth word by a conventional voice recognition device. FIG. 4B is a conceptual diagram of an operation example when recognizing a fifth word by a conventional voice recognition device.

[Explanation of symbols]

 1 feature parameter calculation unit 2 word standard pattern storage unit 3 recognition unit 4 word sequence storage unit 5 language score calculation unit 6 trigram storage unit 7 occurrence probability storage unit 8 concatenated co-occurrence probability storage unit 9 non-concatenated co-occurrence probability storage unit

Claims (1)

[Claims] 1. A feature parameter calculation unit for converting an input voice into a time series of feature parameters representing the feature of the voice, a word standard pattern storage unit for storing a word standard pattern created in advance, and An occurrence probability storage unit that stores an occurrence probability that is a probability that a word occurs, and a concatenation co-occurrence probability storage unit that stores a concatenation co-occurrence probability that is a probability that the first word occurs immediately after the second word. , Using a non-connected co-occurrence probability storage unit that stores a non-connected co-occurrence probability that is a probability that the first word occurs two or more times behind the third word, and a previously recognized preceding word sequence A language score calculation unit that calculates a language score representing a probability that the first word occurs next to the preceding word sequence by calculating the occurrence probability, the concatenated co-occurrence probability, and the non-concatenated co-occurrence probability; The characteristic parameter and the single A recognition unit that calculates an overall score that indicates a likelihood for a word sequence based on an acoustic score that is a similarity to a word standard pattern and the language score, and a result storage unit that stores the word sequence with a high overall score. Speech recognizer.