JP2002082690A - Language model generation method, speech recognition method, and program recording medium therefor - Google Patents

Abstract

(57) [Summary] [Problem] A highly accurate symbol chain probability (language model) can be generated for a recognition task (speech content). SOLUTION: A text database 15 for a recognition task.
0 and a plurality of general text databases 160-1
Used to 160-N, determined symbol linkage probability P _T of DB 150, the OTC DB160-n (n = 1 with P _T,
2, ..., determine the test set perplexity PP _n of N), as the PP _n is small, giving greater weight _{W n (0 <W n <} 1) to the DB 160-n, DB 150 and DB1
The symbol chain probability is obtained from 60-1 to 160-N. At this time, for example, when calculating the number of appearances of a certain word A, the DB
Each of the number of appearances of 160-1 to 160-N corresponds to W ₁ to
W _n is multiplied, and the sum of these values and the number of occurrences of DB 150 is defined as the number of occurrences of A.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech such as a sentence uttered by a human as an input signal, recognizes the speech using an acoustic model and a symbol chain probability (language model), and outputs the result as a symbol sequence. The present invention relates to a method for recognizing speech, a method for generating a language model used in the method, and a program recording medium for these.

[0002]

2. Description of the Related Art When a speech input is converted into a symbol string (word string) by speech recognition, a symbol chain probability (language model) relating to a symbol (word) occurrence chain is generated from a large-scale text database and is used. Conventionally, a technique for improving the speech recognition performance has been known. However, the effectiveness was low when the recognition task (speech content) was different from the task of the large-scale text database used to generate the symbol chain probability.

In order to solve such a problem, as disclosed in Japanese Patent Application Laid-Open No. Hei 4-291399, a large-scale text data is generated by using a learning symbol chain probability created from a text database similar to a recognition task. A technique has been proposed in which a symbol chain probability generated from a database is adapted and speech recognition is performed using the adapted symbol chain probability.

[0004]

The conventional technique using the above-mentioned adapted symbol chain probability can realize higher recognition performance than the technique using only the symbol chain probability generated from a large-scale text database. However, since the symbol chain probability generated from a large-scale text database includes information on data far from the recognition task, the estimation of the adapted probability value becomes unstable depending on the symbol chain. was there. In addition, there is a problem that the adapted symbol chain probability has a large storage capacity because it holds all information of a large-scale text database and a text database similar to a recognition task.

Therefore, one object of the present invention is to generate a highly accurate symbol chain probability by assigning a greater weight to texts that are more similar to a recognition task in a large-scale text database, and to use this for recognition. An object of the present invention is to provide a method of generating a language model that can improve recognition performance by using the language model. Another object of the present invention is to provide a symbol chain probability with a small storage capacity by eliminating data having a low degree of similarity to the recognition task (by setting the weight to 0) to reduce information to be held.

[0006]

According to the language model generating method of the present invention, a text database for a recognition task storing text data relating to a task (speech content) to be recognized, and a general database not directly related to the task to be recognized. Weights indicating the relevance (similarity) of each general-purpose text database to the recognition task text database using a plurality of general text databases, and using the recognition task and general text databases, For each word, the weight of the database to which it belongs is given to generate a symbol chain probability.

To determine the weight of each general-purpose text database, the similarity of the text data of each general-purpose text database is verified based on information obtained from the text data of the recognition task text database, and the weight of each general-purpose text database is determined. A large weight is given to a general-purpose text database having a high similarity with the text database. In order to verify the similarity between the text database for the recognition task and the general text database, the symbol chain probability generated from the text data in the text database for the recognition task was used when the text data in each general text database was given. The perplexity (entropy), or the rate at which each general-purpose text database does not contain a word in the recognition task text database (unknown word rate), respectively, or the perplexity and unknown word rate Use a combination. Also when determining the weight W _T of the text database for recognition tasks, a value based on the ratio of the amount of data amount and the plurality of general text database text database for recognition tasks. This weight W
_T may be used in combination with the value based on the perplexity and the unknown word rate.

[0008]

Embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 shows a configuration example of a speech recognition apparatus to which an embodiment of the speech recognition method according to the present invention is applied. A speech recognition unit 110, a symbol chain probability (language model) database 120, a speech standard pattern database 130, and a symbol chain probability generating unit 140 for a recognition task.
, A recognition task text database 150, and a plurality of general text databases 160-1 to 160-N.
And

The voice standard pattern database 130 holds a plurality of voice standard patterns analyzed in advance. The recognition task database 150 stores text data related to the recognition target task (speech content). If the input speech is, for example, a spoken speech of a news program, each word used to transcribe a large number of news is stored in the recognition task text database 150 as text data of the recognition task. General database 160-1
For 160-N, a newspaper article database, a novel database, or the like, which is not directly related to the recognition target task, is used. For example, a large number of words are collected from newspaper articles, homepages, net news, and the like. One general text database 160-n (n = 1, 2,..., N) may be configured for each of the configured sentences.

Symbol chain probability generating section 140 for recognition task
Executes the language model generation method according to the present invention, and performs a recognition task text database 150 and a general text database 1 prior to a recognition process.
From 60-1 to 160-N, a general text database including text data having a high degree of similarity to the recognition task is given a greater weight, and the recognition task text database and a plurality of general text databases are used to perform recognition. A symbol chain probability (language model) capable of narrowing down tasks with high accuracy is generated and stored in the symbol chain probability database 120.
The speech recognition unit 110 narrows down symbol string candidates for the input speech based on information obtained from the symbol chain probability of the symbol chain probability database 120 and the voice standard pattern of the voice standard pattern database 130, and based on the recognition result. Output a symbol string.

FIG. 2 shows an example of the configuration of the symbol chain probability generation section 140 for a recognition task, and an embodiment of a method of generating a language model, that is, a symbol chain probability according to the present invention will be described with reference to FIG. The weight determining unit 210 inputs the text data of each recognition task in the recognition task database 150 and the text data of each general text database 160-n, and inputs the text data of the recognition task and each general text database 160-n. n based on the similarity to the text data of the n.
Determine the weight W _n for _n . Further, to determine the weight W _T of the recognition task text database 150 based on the weight W _i of each general text database. A specific method for determining these weights W _n and W _T will be described later.
Multiple of the general text database the weight W _T in recognition task for the text database 150 160-1 to 160
−N are given weights W _{1 to} W _N , respectively.

The symbol chain probability generation section 220 receives the text data of the weighted recognition task text database 150 and the plurality of weighted general text databases 160-1 to 160-N output by the weight determination section 210. Then, a symbol chain probability (language model) is generated and stored in the symbol chain probability database 120. The basic method of generating this symbol chain probability, that is, a unigram, a bigram, a trigram, and generally an M-gram (M is an integer of 1 or more) is the same as the conventional method, but a text database for a recognition task. 150 and a plurality of general text databases 160-1 to 160-N are considered as one text database, and a symbol chain probability is generated from this text database.
Consider the weight of the text database to which it belongs. For example, if the number of appearances of the word A in the recognition task text database 150 is C _T (A), and the number of appearances in the general text databases 160-1 to 160-N is C ₁ (A) to C _N (A), These are multiplied by the weight of the database and added, and C (A) = W _T · C _T (A) + W ₁ · C ₁ (A) + W ₂ ·
C ₂ (A) +... + W _N · C _N (A) is the number of appearances of the word A, and the number of appearances of another word is obtained in the same manner. The appearance probability (unigram) of a single word of the word A is obtained by dividing the number of appearances C (A) of the word A by the number of appearances ΣC (k) of all words P (A) = C (A) / ΣC (k) Similarly, the appearance probabilities of other words are obtained and stored in the symbol chain probability database 120.

Alternatively, for example, in the case of a bigram, which is the probability that word B occurs after word A, the number of occurrences of word B following word A in the recognition task text database 150 is C _T (A, B ), Assuming that the number of appearances in the general text databases 160-1 to 160-N is C ₁ (A, B) to C _N (A, B), the sum C ( A, B) = W _T · C _T (A, B) + W ₁ · C
₁ (A, B) + W ₂ · C ₂ (A, B) + ... + W _N · C _N
(A, B) is the number of appearances of the word chains A and B, and is divided by the number of appearances C (A) of the word A, P (B | A) = C (A, B) / C
Similarly, the probability of another word chain may be obtained and stored in the symbol probability database 120 using (A) as the bigram probability that the word B follows the word A.

Next, FIG. 3 shows an example of a processing procedure in the weight determining section 210 in the symbol chain probability generating section 140 for the recognition task shown in FIG. 2, and the operation is initialized to n = 1 (S1). ), Recognition task text database 15
0 text data and general text database 16
A weight wn for the general text database 160- _n is determined from the text data of 0-n (S2).
It is checked whether n = N (S3). If n = N, n is incremented by 1 and the process returns to step S2 (S4). If n = N, that is, all general text databases 160-1 to 160-1
After determining the weight W ₁ to _W-N for 60-N, it determines a weight W _T for recognition task text database 150 (S5).

A specific example of determining the weight w _n will be described with reference to FIG. For example, when weighting is performed based on perplexity, the symbol chain probability _PT is generated by the symbol chain probability generation unit 410 using the text data of the recognition task text database 150, and the text database weight calculation unit 420 generates the symbol chain probability _PT. calculate the test set perplexity for the general text database 160-n of symbols chain probability P _T, determine the value W _n of weights based on the value of the perplexity. The test set perplexity PP represents the average number of branches of a word in the information-theoretic sense of the language L, is applied to an evaluation text set (word string, symbol string), and is given by the following equation.

PP = 2^{H (L)} Where H (L) = − Σ_w1(1 / n) P (w₁ ⁿ) LogP
(W₁ ⁿ) And H (L) is the entropy per word
And P (w₁ ⁿ) Is the word string w ₁ ⁿ= W₁... w_nGenerate
Probability. In other words, the text database for the recognition task
Chain probability P generated using the_TUsing,
Text data in the general text database 160-n
For test set perplexity PP
You. The higher the language perplexity, that is, the word
The higher the average number of branches, the more difficult it is to identify words.
Test set parp from the symbol chain probability and evaluation text
When determining the lexity PP, the smaller the value of that PP
The symbol chain probability can express the evaluation text well
Will be. Therefore, the general-purpose text data
Test set purple obtained for database 160-n
If the lexity value is small, this general text data
Database 160-n is a text database for the recognition task.
Will be similar to So general text
Weight W for database 160-n_nIncrease
You.

Further, the general text database can be weighted based on the unknown word rate. In this case is present in the recognition task text database 150 as shown in FIG. 4 (appearing) list of symbols (words), that generates a symbol list L _T by the symbol list generator 430, the text database weight calculator in 420, among the number total of words appearing for general text database 160-n (symbol), calculate the percentage (vocabulary rate) is not included in the symbol list L _T sign (word) of that how many there Then, the weight value Wn of the general text database 160- _n is determined based on the value of the unknown word rate. For example, if the total number of words in the general text database 160-n is 20
In 00, when 100 words in it was an unknown word that is not included in the symbol list L _T, it is the unknown word rate (100
/ 2000) × 100 = 500. About the unknown word rate is low, symbol list L _T and the general text database for general use text database 160-n are many words that overlap in the 160-n is recognition tasks for database 150
It will be similar to, to increase the weight W _n.

In the text database calculation section 420,
Not only in the case of using any of the symbols list L _T and symbolic chain probability P _T, can be used in combination both. For example, a general text database 160
Test set perplexity using symbols linkage probability P _T of -n is smaller than the threshold value, and if the unknown word rate for symbol list L _T is smaller than the threshold of the general text database 160-n weights W _{n is set} to 1; in other cases, the weight W _{n is set} to 0. Note generating symbols list L _T is 3 Manbun in Examples described later, when a total of 1,000,000 words recognition task text database 150, different but word count is about 100,000 words, the 100,000 word in the, it includes many words that do not appear only once during the recognition task database 150, from those words are the low statistical reliability, not included in the symbol list L _T as it did not appear If also there, it said in the experiments described below was the symbol list L _T from many word frequency of occurrence up to 20,000 words. The 20,000 words with the highest frequency account for about 99% of the total 1 million words in the database 150.

Next, the text database 1 for the recognition task
An example of processing for determining the weight w _T of 50 will be described with reference to FIG. For example, the recognition task text database 15
Seeking 0 data amount L _T in a text data amount calculation section 510, the general text database 160-1～160-
Obtains the full text data amount C _D N-text data amount calculator 520, weight calculator 53 these ratios C _D / C _T
Calculated in 0, it gives the weight W _T of the recognition task text database 150 based on the calculation result. When considering the weight W _n as the total text data amount C _D of the general text database 160 - 1 to 160-N C
Determined by calculating the ^{_{D = Σ n = 1 N W}} n C n. C _n is the base number of words of the general text database 160-n.

In the above description, the weight W _T need not be given to the recognition task text database 150. That is, the general text databases 160-1 to 160-N
Seeking weight W ₁ to _W-N, the symbol linkage probability by the symbol linkage probability generation unit 220 in FIG. 2 may be determined as described above using the same. In this case, it can be said that W _T = 1 and W _{1 to} W _N are positive numbers of 1 or less. Conversely, without giving weights for General text database 160 - 1 to 160-N, that is to all 1 W ₁ to _W-N, giving one or more of the weight W _T recognition task text database 150, The symbol chain probability may be calculated by the symbol chain probability generation unit 220.

The above-described generation of the language model and speech recognition can also be performed by executing a program by a computer. For example, as shown in FIG. 6, each unit is connected to a bus 670, and a language model generation program
It is installed from a D-ROM, a hard disk or the like or via a communication line, and the CPU 660 executes the language model generation program to
Using the recognition task text database 150 and the general text databases 160-1 to 160-N, as shown in FIG.
0 words generates symbol linkage probability P _T or symbol list L _T from (S1), then, for each of the general text database 160 - 1 to 160-N, weights W _n
And sequentially calculating (S2), then calculates a weight W _T of the recognition task text database 0.99 (S4), then,
These weights W ₁ to _W-N, and generates the word and text database 150 and 160 - 1 to 160-N symbol linkage probability (language model) using the W _T stored in the symbol linkage probability database 120 (S4 ).

Thereafter, voice recognition is performed.
Install the speech recognition program as above
When a voice is input to the input unit 610, the CPU 66
0 executes the speech recognition program, and the symbol chain probability data
Base 120 and voice standard pattern database 130
And performs a speech recognition, and outputs the resulting symbol string to an output unit 63.
Output from 0. Note that the voice input by the input unit 610 is
For example, LPC cepstrum analyzed by linear prediction, LPC
Feature parameters such as delta cepstrum and log power
The vector data is a time series. Storage unit 6
20 is used to generate language models and temporarily store data during speech recognition.
It is used for storing.Example An evaluation experiment was performed to confirm the effects of the present invention. Comment
The input voice data for the price is a male announcement of a news program.
Sir's utterance 129 sentences were used. Text data for recognition task
The database 150 contains about 1 million words of news
The transcript text was used. General text data
As the bases 160-1 to 160-N, newspaper articles, e-
About 150 million collected from websites, net news, etc.
Prepare a database of 100 million words, and in this evaluation experiment
Is one sentence in this text database
General text database 160-1 to 160-N
And Weight W of general text database_nDecision
Has Perplexity PP _nAnd general text data
Perplexite in the bases 160-1 to 160-N
IPP_nIs 40% of the sentences that are lower than the threshold
Sea perplexity PP_nSet the threshold for
Plexity PP_nGeneral-purpose text
Weight W of Kist database 160-n_nOne, threshold
Value of the general text database 160-n
Weight W_nWas set to 0. Also, text data for recognition tasks
Weight W for database 150_TIs the text for the recognition task
Data (number of words) C of the strike database 150_TAnd heavy
General-purpose text database 160-1
60-N data amount (number of words) C_DAnd based on C_D/
C_TAnd

When the evaluation was performed based on the perplexity of the symbol chain probability for the transcribed text of the evaluation speech data (the average number of branches of the word, the smaller the word, the higher the accuracy of the evaluation text is the chain probability). In the case of the symbol chain probability generated only from the text database 150 for use, the number of text data is small, so the value becomes 75, which is a large value. The text database 150 for the recognition task and the large-scale text database group, that is, the general text databases 160-1 to 160 are used. −N, the symbol chain probability is reduced to 42, but the symbol chain probability generated using the weighting of the present invention is 36.
And became even smaller.

When evaluated by a speech recognition experiment, the word error rate was 14.7 in the case of a symbol chain probability generated only from the recognition task text database 150.
%, 11.6% for the symbol chain probability generated from the recognition task text data and the general text databases 160-1 to 160-N, and 11.6% for the symbol chain probability generated using the weighting of the present invention. This was 9.9%, which clearly improved the recognition rate.

The number of parameters of the symbol chain probability is about 10 million in the case of the symbol chain probability generated from the recognition task text database 150 and the general text databases 160-1 to 160-N. In the case of the symbol chain probability generated using weighting, about 4
It was 600,000, considerably reduced.

[0026]

As described above, according to the present invention, the following first and second effects can be obtained. The first effect is that a high-precision symbol chain probability can be generated for a recognition task by weighting data similar to the recognition task text data from a general large-scale database group.

The second effect is that, by setting the weight for data having low similarity to the recognition task to 0 at the time of weighting, a symbol chain probability with high accuracy and small storage capacity can be generated.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an apparatus to which a voice recognition method according to the present invention is applied.

FIG. 2 is a block diagram showing a configuration of a generation unit to which a language model (symbol chain probability) generation method of the present invention is applied.

FIG. 3 is a flowchart showing a processing flow of a language model generation method according to the present invention.

FIG. 4 is a block diagram showing a configuration example of a weight determining unit of the general text database.

FIG. 5 is a block diagram showing a configuration example of a weight determination unit of a text database for a recognition task.

FIG. 6 is a diagram showing a configuration example when a language model generation method and a speech recognition method according to the present invention are executed by a computer.

FIG. 7 is a flowchart showing an example of a processing procedure of a language model generating method according to the present invention.

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Shoichi Matsunaga, Inventor 2-3-1 Otemachi, Chiyoda-ku, Tokyo Nippon Telegraph and Telephone Corporation (72) Inventor Go Go Kawabata 2-3-3, Otemachi, Chiyoda-ku, Tokyo No. 1 Nippon Telegraph and Telephone Corporation F-term (reference) 5D015 HH23

Claims (10)

[Claims] 1. A recognition task text database storing text data relating to a recognition target task, and a plurality of general text databases storing general text data not directly related to the recognition target task. A weight indicating the relation between each of the general text databases with respect to the task text database is determined. Using the recognition task text database and the plurality of general text databases, a symbol (word) of interest belongs to a database to which it belongs. A language model generation method characterized by generating a symbol chain probability (language model) by giving a weight. 2. A symbol chain probability is determined using the recognition task text database, and the weight is calculated for each general text database based on perplexity (entropy) using the symbol chain probability. The language model generation method according to claim 1, wherein: 3. The weight of each of the general text databases is calculated based on the ratio (unknown word rate) of symbols (words) in the database that are not included in the recognition task text data. 2. The language model generating method according to claim 1, wherein 4. A symbol chain probability is determined using the recognition task text database, and a perplexity (entropy) is determined for each general text database using the symbol chain probability. Of the symbols (words) in the database that are not included in the recognition task text data (unknown word rate), and for each of the general text databases, the perplexity and unknown word rate are calculated. 2. The language model generation method according to claim 1, wherein the weight is obtained from the following. 5. The weight of the recognition task text database based on a ratio of the text data amount of the recognition task text database to the text data amount of the plurality of general text databases. The language model generation method according to any one of 1 to 4. 6. A recognition task text database storing text data relating to a recognition target task and a plurality of general text databases storing general text data not directly related to the recognition target task. The task text database is given a greater weight than the plurality of general text databases, and the recognition task is performed for the symbol (word) of interest using all of the recognition task text database and the plurality of general text databases. A language model generation method characterized by generating a symbol chain probability (language model) by giving a weight to a text database. 7. The method of generating a symbol chain probability by assigning a weight to a target symbol includes assigning a weight of the database to the number of appearances in each database, and using the total value as the number of occurrences in the entire database. The method according to claim 1, wherein a probability is generated. 8. The method of generating a symbol chain probability by assigning a weight to an occurrence frequency or a symbol chain probability in each database for a symbol of interest is given a weight of the database, and the total integrated value of the symbol is calculated in the entire database. 8. The language model generation method according to claim 1, wherein a symbol chain probability is obtained as the appearance frequency or the symbol chain probability. 9. A speech recognition method for recognizing an input speech by using an acoustic model and a symbol chain probability (language model) and outputting the symbol as a sequence of symbols (words). A speech recognition method characterized by using a language model generated by the method according to any one of the first to third aspects. 10. A recording medium on which a program for causing a computer to execute the method according to claim 1 is recorded.