JP2975540B2 - Free speech recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a free speech recognition apparatus for recognizing free speech using a statistical language model in which the appearance of redundant words is statistically modeled.

[0002]

2. Description of the Related Art In recent years, research on speech recognition for free utterances has been actively conducted. One of the features of free utterances compared to reading utterances is "Ah,""Eh," etc. during utterance. Many redundant words appear. Several methods for dealing with redundant words in speech recognition processing have already been proposed, as shown below.

First, Murakami et al., "Word trigram.
Extension to Sentence Speech Recognition and Free Utterance Recognition Using ”, IEICE Technical Report, SP93-127, 199
In January 2004, the following first conventional speech recognition apparatus was disclosed. In the first conventional apparatus, in a speech recognition apparatus using a word trigram, redundant words are skipped in a language model while redundant words are recognized in an acoustic model. That is, when connected to a word other than a redundant word, the redundant word is skipped, the value of the trigram of the word is calculated, and the probability of the trigram connected to the redundant word is set to 1.0. By skipping the redundant words, the relation between words appearing between the redundant words can be grasped as a trigram value, and the speech recognition processing is performed to allow the appearance of the redundant words between arbitrary words. Execute.

Also, Kobayashi et al., "Feature Analysis of Appearance Position of Interjections and Evaluation of Speech Recognition System", IEICE Type 2 Workshop, SPREC-92-3, February 1993 Of the related art is disclosed. In the second conventional example, in a continuous speech recognition apparatus based on a top-down type parser of a context-free grammar, a vocabulary of interjections (redundant words) is added to syntactic knowledge of a task described by a context-free grammar. Then, the rewriting rules are modified and used so as to allow the insertion of interjections. This device has an advantage that a position where a redundant word can appear can be specified by the description of the rewriting rule.

[0005]

However, in the first conventional example, it is possible to use a bias of the appearance position such as "redundant words are likely to appear at the beginning of a sentence or between clauses, but are unlikely to appear within a clause". Since it is not possible, there is a high possibility that erroneous recognition in which a redundant word is erroneously inserted at a position where a redundant word does not actually appear will occur. Therefore, there is a problem that the speech recognition rate is reduced.

Further, in the second conventional example, the probability of occurrence of a redundant word cannot be defined in detail because it is not possible to finely express the deviation in the ease of appearance depending on the type and position of the redundant word. Therefore, it is not possible to expect improvement in recognition performance of speech recognition of free speech. Further, since the rules must be manually described, the creation of the language model cannot be automated, and the creation of the language model requires an extremely large amount of time and effort.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems, to automatically generate a statistical language model in consideration of the probability of a redundant word, and to have a higher free speech utterance recognition than the conventional example. It is an object of the present invention to provide a free-speech speech recognition device capable of obtaining a rate.

[0008]

According to a first aspect of the present invention, there is provided a freely uttered speech recognizing apparatus which converts a free uttered speech sentence composed of an input word string by referring to a predetermined statistical language model. In the free speech recognition apparatus for recognizing, the statistical language model includes: a first occurrence probability for a sequence of basic words that are words that do not belong to a predetermined category in the word sequence; Given a sequence of words, a word belonging to the specific category inserted between words of the sequence of basic words including before the first word and after the last word of the sequence of basic words. It is created so as to be determined by the second occurrence probability for a set of zero or more columns of insertion words, and when the basic words are consecutive, it is assumed that zero insertion words are inserted between them. Feature To.

[0009] Further, according to a second aspect of the present invention, there is provided the free speech recognition apparatus according to the first aspect.
The second occurrence probability is such that when the sequence of the basic words is given in the word sequence, the sequence of the inserted words at each position is different from the sequence of the inserted words at other positions. It is determined by a third set of independently determined probabilities of occurrence.

Further, in the free speech recognition apparatus according to the third aspect, in the free speech recognition apparatus according to the second aspect, the third occurrence probability is a probability determined by a chain and a position of a basic word; It is determined by an occurrence probability determined by a sequence of inserted words inserted therein.

According to a fourth aspect of the present invention, in the free speech recognition apparatus, the words belonging to the specific category are words belonging to a specific part of speech. It is characterized by the following. Further, a free speech recognition apparatus according to a fifth aspect is characterized in that in the free speech recognition apparatus according to the fourth aspect, the specific part of speech is an interjection. Still further, according to a sixth aspect of the present invention, in the free speech recognition apparatus according to the fourth aspect, the specific part of speech is an adverb.

Further, in the free speech recognition apparatus according to the present invention, in the free speech recognition apparatus according to any one of claims 3 to 6, the sequence of the basic words is modeled by a word trigram. And a statistical language model is created by modeling the sequence of the inserted words appearing in one place with a bigram of words.

[0013]

According to the first aspect of the present invention, the statistical language model includes:
A first occurrence probability for a sequence of basic words that are words that do not belong to a predetermined category in the word sequence;
Given the sequence of basic words in the word sequence, the specific word inserted between words of the sequence of basic words including before the first word and after the last word of the sequence of basic words A second occurrence probability for a set of zero or more columns of insertion words that are words belonging to a category is created so that if the basic words are consecutive, there are no insertion words between them Is assumed to be inserted. Then, by referring to the statistical language model created in this way, the uttered speech sentence of the free utterance composed of the input word string is speech-recognized.

In the free speech recognition apparatus according to the second aspect, preferably, the second occurrence probability is determined for each position when the sequence of the basic words is given in the word sequence. Is determined by a third set of occurrence probabilities that are determined independently of the sequence of the inserted words at the other positions. Further, in the free speech recognition apparatus according to the third aspect, the third occurrence probability is preferably determined by a probability determined by a chain and a position of a basic word and a sequence of inserted words inserted therein. And the probability of occurrence.

In the speech recognition apparatus according to the fourth aspect, the words belonging to the specific category are preferably words belonging to a specific part of speech, and the words belonging to the specific part of speech are, for example, interjections. Or an adverb. Further, in the speech-recognition apparatus according to claim 7, the sequence of the basic words is modeled by a trigram of words, and the sequence of the inserted words appearing in one place is modeled by a bigram of the words. Is created.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A speech recognition apparatus according to an embodiment of the present invention will be described below with reference to the drawings. The speech recognition apparatus of the present embodiment of FIG. 1 particularly models, for example, a non-redundant word string by a trigram based on the text data for learning a free utterance composed of a word string stored in the text data memory for learning 11. And a language model creator 10 for modeling a redundant word sequence appearing in one place with a bigram to generate a statistical language model in consideration of the probability of the redundant word and storing the statistical language model in the statistical language model memory 7. And The speech recognition apparatus includes a microphone 1, a feature extraction unit 2, a buffer memory 3, and an HMM, which is an acoustic model in a hidden Markov model (hereinafter referred to as HMM) memory 5, based on input uttered speech data. A phoneme matching unit 4 that executes phoneme matching processing with reference to the phoneme matching unit 4 and outputs phoneme data, and One pass D based on phoneme data from the phoneme matching unit 4.
One pass DP speech recognition unit (hereinafter referred to as speech recognition unit) 6 that executes speech recognition with reference to a statistical language model in a statistical language model memory 7 using a P (Viterbi search) algorithm.

First, a redundant word in the embodiment according to the present invention is defined as follows. In the present embodiment, the criterion of a redundant word is defined as "a self-contained word that is used by being inserted into a sentence when it is stagnant and whose removal does not affect the grammar and meaning of the sentence". The definition of the redundant word is, for example, a word such as "ano" or "erto" and includes interjections in the part of speech classification. Instead of the redundant word, an adverb such as "very" or "hiyouni" may be used. Words that express emotion, call, response, and greeting, such as "yes" and "hello", are classified as inspirational verbs and are not included in redundant words in this embodiment.
Also, it does not include stagnation or restatements such as "ki" and "sai" in "ki, I want to hear" and "I say, say."

As described above, in this embodiment, it is assumed that removing redundant words does not affect sentences. Therefore, a sentence including a redundant word is modeled as a redundant word being stochastically inserted into a chain of non-redundant words. Consecutive redundant words are considered together as a redundant word string. In general, redundant words may be inserted in the middle of a word,
In this case, we'll exclude it. When modeling a statistical language model in the language model creating unit 10, the following 2
We make two assumptions. (Assumption 1) The probability that a redundant word or a redundant word string is inserted at a certain position in a non-redundant word string, and the probability of occurrence of a redundant word or a redundant word string when it is inserted, is the redundant word or the redundant word at another point. It does not depend on the presence or absence of redundant word strings or their contents. (Assumption 2) When a redundant word or redundant word string is inserted into a sentence, the occurrence probability of the redundant word or redundant word string does not depend on the non-redundant word string.

Formulation of these two assumptions is as follows. The non-redundant word string Sh in the sentence S is represented by the following equation 1.

[0020]

## EQU1 ## Sh = w ₁ , w ₂ ,..., W _n

The redundant word sequence g _i inserted between the non-redundant word w _i and the non-redundant word w _{i + 1} is represented by the following equation (2).

[0022]

## EQU2 ## g _i = g _i1 , g _i2,.

Here, the redundant word sequence g _i is a sequence of zero or more continuous redundant words, and includes a non-redundant word w _i and a non-redundant word w _{i + 1.}
When there is no redundant word between, it is expressed by the following equation (3).

## EQU3 ## g _i = φ

Here, w ₀ and w _{n + 1} represent the beginning and end of sentence S, respectively. The random variable corresponding to the redundant word string g _i and G _i. Hereinafter, when the random variables are clear, they are omitted. At this time, the occurrence probability of the sentence S is expressed by the following Expression 4.

[0025]

P (S) = P (Sh, g ₀ , g ₁ ,..., G _n ) = P (Sh) · P (g ₀ , g ₁ ,..., G _n | Sh)

Here, the first term P (Sh) of the second equation of Expression 4
Is the non-redundant word sequence Sh in the sentence of the input free speech
, And the second term P (g ₀ , g ₁ ,
.., G _n | Sh) is given between the words of the non-redundant word string Sh including the front of the first word of the non-redundant word string Sh and the back of the last word when the non-redundant word string Sh is given. Occurrence probability for a set of zero or more columns of redundant words to be inserted. Here, when non-redundant words continue, it is considered that 0 redundant words are inserted between them. Next, when the assumption condition 1 is formulated, it is expressed by the following equation (5).

[0027]

(Equation 5)

Here, Π on the right side of Expression 5 indicates (n + 1) occurrence probabilities P (g _i | S when i is changed from 0 to n.
h) shows the calculation of the product of As can be seen from Equation 5, the probability P
(g ₀ , g ₁ ,..., g _n | Sh) is determined independently of the redundant word string inserted at each position in the sentence of the freely uttered speech from the redundant word string at other positions. Is determined by a set of occurrence probabilities. Here, the probability random variable G _i redundant word string in the non-redundant word string Sh is not φ P (G _i ≠ φ│Sh)
Is expressed as P _i (Sh), when g _i ≠ φ, that is, when there is no redundant word string g _i , the following equation 6 is obtained.

[0029]

P (g _i | Sh) = P _i (Sh) · P (G _i = g _i | S
h, G _i ≠ φ)

[0030] Here, the occurrence probability P (g _i │Sh) includes a occurrence probability P _i (Sh) when random variable G _i redundant word string in the non-redundant word string Sh is not phi, redundant word or redundant word Occurrence probability P (G _{i of} a redundant word or redundant word sequence when a sequence is inserted
= G _i | Sh, G _i ≠ φ).

Further, from assumption 2, the probability of occurrence of a redundant word or redundant word string when a redundant word or redundant word string is inserted in a sentence does not depend on non-redundant word strings. That is, since the occurrence probability P (G _i = g _i | Sh, G _i ≠ φ) does not depend on the non-redundant word string Sh and the number i of the order, the occurrence probability P
If (G _i = g _i | Sh, G _i ≠ φ) is expressed as P (g) (g ≠ φ), the following equation 7 is obtained.

[0032]

P (g _i | Sh) ≒ 1−P _i (Sh), when g _i = φ ≒ P _i (Sh) · P (g _i ), g _i ≠ φ

From equation (7), the occurrence probability P (g _{i of} a redundant word or redundant word string when a redundant word or redundant word string is inserted in a sentence
| Sh) is determined by the occurrence probability P _i (Sh) determined by the non-redundant word sequence Sh and the position and the occurrence probability P (g _i ) determined by the redundant word sequence g _i inserted therein. It is understood that it is done.

For example, the occurrence probability P (Sh) of the non-redundant word sequence Sh when the non-redundant word sequence Sh is modeled by a word trigram as a statistical language model, and the redundant word appearing in one place The occurrence probabilities P (g) of the redundant word string g when the string g is modeled by a word bigram (bigram) are expressed by the following equations 8 and 9, respectively.

[0035]

(Equation 8)

(Equation 9)

Here, Π on the right side of Equation 8 is the calculation of the product for i, and Π on the right side of Equation 9 is the calculation of the product for j. In addition, the probability P _i (Sh) that a redundant word is inserted between the non-redundant word w _i and the non-redundant word w _{i + 1 in} the non-redundant word sequence Sh is determined only by the non-redundant words w _i and w _{i + 1} . and is determined, the probability P _i (Sh) and _{α (w i, w i +} 1) putting a, the probability P (S) sentence S, rewritten as shown following Equation 10, the number 11 be able to.

[0037]

(Equation 10)

[Equation 11]

Here, Π on the right side of Expression 10 indicates that i is 0 to n
Π in the second expression of Expression 11 is the calculation of the product related to j. The probability _{P (w i + 1 │w i} -1, w i), the probability _{α (w i, w i +} 1) and the probability _{P (g j │g j-1} ) is for learning of free speech consisting of a string of words Estimated from text data. As described above, the language model creation unit 10 models the non-redundant word string Sh using the word trigram, and obtains a statistical language model by converting the redundant word string g appearing in one place into a bigram model of the word. And this is stored in the memory 7.

The configuration and operation of the speech recognition apparatus using the statistical language model of the present embodiment created as described above will be described.

In FIG. 1, a speaker's uttered voice is input to a microphone 1 and converted into a voice signal, and then input to a feature extracting unit 2. After performing A / D conversion on the input audio signal, the feature extraction unit 2 performs, for example, LPC analysis, and performs 34-dimensional feature parameters including logarithmic power, 16th-order cepstrum coefficient, Δlogarithmic power, and 16th-order Δcepstrum coefficient. Is extracted. The time series of the extracted feature parameters is input to the phoneme matching unit 4 via the buffer memory 3.
HMM in HMM memory 5 connected to phoneme matching unit 4
Is composed of a plurality of states and arcs indicating transitions between the states. Each arc has a transition probability between states and an output probability for an input code. The phoneme matching unit 4 performs phoneme matching processing based on the input data, and outputs phoneme data to the speech recognition unit 6.

A statistical language model memory 7 which stores a predetermined statistical language model including a trigram and a bigram of a word in advance is connected to the speech recognition unit 6. Voice recognition unit 6
Refers to the statistical language model in the statistical language model memory 7 and processes the input phoneme data from left to right in the left-to-right direction without backtracking using a predetermined One pass DP algorithm to obtain a higher occurrence. The speech recognition processing is executed by determining the word of the probability as the speech recognition result data, and the determined speech recognition result data (character string data) is output.

As described above, in the speech recognition apparatus according to the present embodiment, the types of words and phrases whose appearance does not affect the grammar and the meaning of the sentence even when redundant words are removed are used as linguistic constraints. Since the expression can be expressed more finely, the recognition performance of the freely uttered speech including such a phrase can be improved, and a higher speech recognition rate can be obtained as compared with the conventional example. In addition, since the linguistic constraint can be automatically learned as a statistic from a large amount of text data for learning, if there is text data for learning, a statistical language model can be easily and automatically used, for example, using a digital computer. Can be created in a typical way.

In the above embodiment, the feature extracting unit 2, the phoneme checking unit 4, the speech recognizing unit 6, and the language model creating unit 10 can be constituted by, for example, a digital computer.

In the above embodiment, the statistical language model of the non-redundant word string is modeled by the word trigram. However, the present invention is not limited to this. Instead of the word trigram,
A context-free grammar with probability may be used.

[0045]

As described above in detail, according to the present invention, there is provided a free speech recognition apparatus for recognizing a free speech utterance composed of an input word string by referring to a predetermined statistical language model. The statistical language model includes a first occurrence probability for a sequence of basic words that are words that do not belong to a predetermined category in the word sequence, and a case where the sequence of basic words is given in the word sequence. A sequence of zero or more inserted words that are words belonging to the specific category and inserted between words of the sequence of basic words including before the first word and after the last word of the sequence of basic words. And the second occurrence probability for the set of..., Where if the basic words are consecutive, 0
It is assumed that the inserted words have been inserted. Therefore, for example, even if the redundant words, such as insertion words, are removed, the type or appearance position of a phrase that does not affect the grammar and meaning of the sentence can be expressed more finely as a linguistic constraint. It is possible to improve the recognition performance of the freely uttered voice and obtain a higher voice recognition rate than the conventional example. In addition, since the linguistic constraint can be automatically learned as a statistic from a large amount of text data for learning, if there is text data for learning, a statistical language model can be easily and automatically used, for example, using a digital computer. Can be created in a typical way.

[Brief description of the drawings]

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

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... Feature extraction part, 3 ... Buffer memory, 4 ... Phoneme collation part, 5 ... One pass DP speech recognition part, 6 ... Hidden Markov model (HMM) memory, 7 ... Statistical language model memory, 10 ... Language model creation unit 11: Text data memory for learning.

Continuation of the front page (56) References The Acoustical Society of Japan Fall Meeting, 1994, I-Q1-Q-2, "Study of redundant word processing in statistical language models" p. 145-146 (Received by the NDL on November 1, 1994) Transactions of the Institute of Electronics, Information and Communication Engineers, Vol. J77-A No. 2, Feburary 1994, "A Method of Processing Unwanted Words in Freely Spoken Sentences Using Garbage HMM" (Published February 25, 1994) Proceedings of the Acoustical Society of Japan Spring Meeting 1993 I-4 -12 “Examination of Garbage HMM for Understanding Freely Spoken Sentences” p. 41-42 (issued on March 17, 1993) IEICE Technical Report [Voice] Vol. 92, No. 410, SP92-127, "Examination of use of Gar page HMM for comprehending free utterance sentences" p. 33-40 (issued on January 19, 1993) IEICE Technical Report [Voice] Vol. 91, No. 396, SP91-100, “Study of acoustic and linguistic problems in speech recognition of free speech” p. 71-78 (Issued December 20, 1991) IEICE Technical Report [Language Understanding and Communication] Vol. 398, NLC91-57, “Study on acoustic and linguistic problems in speech recognition of free speech” p. 71-78 (issued on December 20, 1991) IEICE Technical Report [Voice] Vol. 93, no. 426, SP93-127, "Sentence Speech Recognition Using Word Trigram and Extension to Free Speech Recognition" p. 71-78 (Issued January 20, 1994) (58) Fields investigated (Int. Cl. ⁶ , DB name) G10L 3/00 537 G10L 3/00 531 G10L 3/00 535 JICST file (JOIS)

Claims (7)

(57) [Claims] 1. A free utterance speech recognition apparatus for recognizing a speech utterance of a free utterance composed of an input word string with reference to a predetermined statistical language model, wherein the statistical language model includes: A first occurrence probability for a sequence of basic words that do not belong to a predetermined specific category; and when the sequence of basic words is given in the word sequence, the first occurrence of the sequence of basic words in the sequence of basic words is determined. A second occurrence probability for a set of zero or more columns of inserted words that are words belonging to the specific category and inserted between words of the basic word sequence including the preceding and last words. A free speech recognition apparatus characterized in that, when the basic words are continuous, it is regarded that 0 insertion words are inserted between them. 2. The second occurrence probability is such that, when the sequence of the basic words is given in the word sequence, the insertion of another portion with respect to the sequence of the inserted words inserted therein at each position is performed. 2. The free speech recognition apparatus according to claim 1, wherein the third set of occurrence probabilities is determined independently of the word sequence. 3. The third occurrence probability is determined by a probability determined by a chain and a position of a basic word and an occurrence probability determined by a sequence of inserted words inserted therein. The free speech recognition device according to claim 2. 4. The speech recognition apparatus according to claim 1, wherein the words belonging to the specific category are words belonging to a specific part of speech. 5. The speech recognition apparatus according to claim 4, wherein the specific part of speech is an interjection. 6. The free speech recognition apparatus according to claim 4, wherein the specific part of speech is an adverb. 7. A statistical language model is created by modeling the sequence of basic words with a trigram of words and modeling the sequence of insertion words appearing at one place with a bigram of words. 7. The freely uttered speech recognition device according to any one of 3 to 6.