JP2000356997A - Statistical language model generator and voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a statistical language model generator, that a statistical language model having superior performance in prediction percision, reliability and robustness is generated by executing prescribed normalizing and smoothing processes to generate a forward-looking N-gram statistical language model. SOLUTION: When a switch SW is switched to an 'a' side, a statistical language model memory 21 is connected to a word hypothesis squeezing section 6 and the section 6 executes a word hypothesis squeezing process while referring to a forward-looking N-gram statistical language model in the memory 21 generated by a language model generating section 20. When the switch SW is switched to a 'b' side, a statistical language model memory 31 is connected to the section 6, the section 6 executes a word hypothsis squeezing process while referring to a fusion N-gram statistical language model in the memory 31 generated by a language model generating section 30. Thus, voice recognition is conducted at an improved voice recognition rate.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a statistical language model generating apparatus for generating a statistical language model based on learning text data, and a statistical language model using the statistical language model.
The present invention relates to a voice recognition device that recognizes a voice signal of an input uttered voice sentence.

[0002]

2. Description of the Related Art In recent years, a method of using a language model has been studied to improve the performance of a continuous speech recognition apparatus. This aims at improving the recognition rate and reducing the calculation time by predicting the next word using a language model and reducing the search space. Recently, N-gram (N-gram) has been widely used as a language model.
m). It learns large-scale text data and statistically gives the transition probability from the previous N-1 words to the next word. Multiple L word strings w ₁ ^L = w
The generation probability P (w ₁ ^L ) of ₁ , w ₂ ,..., W _L is expressed by the following equation.

[0003]

(Equation 1)

[0004] Here, w _t represents a t-th one word of the word string w ₁ ^L, w _i ^j represents the j-th word string from the i-th. In the above _equation 1, the probability P (w _t |
wt _{+ 1-} ^Nt-1 ) is a word sequence wt _{+ 1-} ^Nt-1 composed of N words.
Is the probability that the word w _t will be uttered after is uttered, and similarly, the probability P (A | B) means the probability that the word A will be uttered after the word or word string B has been uttered. Also,
“Π” in Equation 1 represents the probability P (w _t from t = 1 to L
| W _{t + 1−N} ^t−1 ), and so on.

In large vocabulary continuous speech recognition, the word N
Although N-gram statistical language models such as -gram are widely used, the following four points must be sufficiently satisfied. (1) Accuracy of next word prediction (2) Reliability for sparse data (data with a small amount of learning) (3) Compact model size (4) Task (or domain, job to be subjected to speech recognition processing, etc.) Robustness against misalignment

[0006]

The above word N-gra
m has excellent performance with respect to "accuracy of next word prediction", but remains unsatisfactory in terms of "reliability for sparse data" and "robustness against task deviation". on the other hand,
The part-of-speech class N-gram based on the part-of-speech class information (hereinafter referred to as the part-of-speech class N-gram) is considerably inferior to the word N-gram in "accuracy of next word prediction", but excellent in the other three points. . Class N
A method of performing class classification by automatically performing clustering as -gram (for example,
"Hirokazu Masataki et al.," N-gra by maximum posterior probability estimation
Task adaptation of m-language model ", IEICE Transactions, Vol.J81-D-II, pp.2519-25.
25, November 1998 ". ), A statistical language model (hereinafter, referred to as an automatic class 2-gram) has also been proposed, which is used in any of “accuracy of next word prediction”, “reliability for sparse data”, and “compact model size”. Shows good performance, but the classification itself depends on the task,
It is inferior in terms of "robustness to task deviation". Table 1 summarizes the fulfillment of these models for the four requirements.

[0007]

[Table 1] Comparison of basic performance of various language models ―――――――――――――――――――――――――――――――― Prediction accuracy Reliability Model Size Robustness ―――――――――――――――――――――――――――――――― Word 2-gram ◎ × × × ―――― ―――――――――――――――――――――――――――――― Part of speech class 2-gram × ◎ ◎ ◎ ――――――――――― ――――――――――――――――――――――― Automatic class 2-gram ◎ ◎ ◎ × ―――――――――――――――――― ――――――――――――――――

A first object of the present invention is to solve the above-mentioned problems and to generate a statistical language model which cannot reduce the model size but has excellent performance in prediction accuracy, reliability and robustness. It is an object of the present invention to provide a statistical language model generation apparatus capable of performing the above-described processing and a speech recognition apparatus using the same.

A second object of the present invention is to solve the above-mentioned problems and to provide a statistical language model capable of generating a statistical language model having excellent performance in prediction accuracy, reliability, model size and robustness. An object of the present invention is to provide a language model generation device and a speech recognition device using the same.

[0010]

According to a first aspect of the present invention, there is provided a statistical language model generating apparatus comprising a plurality of words based on learning text data in which uttered voice sentences of a predetermined speaker are written. The forward word N is calculated by calculating the first frequency probability that the word to be processed occurs after the word string.
A first part-of-speech class of the first word and the first part-of-speech class based on the learning text data and the part-of-speech class information data including the part-of-speech class information; A statistical language model of a forward-looking part-of-speech class-word N-gram is generated by calculating a second frequency probability that a word to be processed occurs after a word string including a plurality of words connected after the word. A second generation unit, and a forward-looking word generated by the first generation unit using a forward-looking part of speech class-word N-gram statistical language model generated by the second generation unit as prior knowledge; Forward transition formed by interpolating between the first frequency probability and the second frequency probability by the maximum posterior probability estimation method using the N-gram statistical language model as the posterior knowledge A first calculating means for calculating the rate, and a forward N-gram statistic by executing a predetermined normalizing process and a smoothing process on the forward transition probability calculated by the first calculating device. And first processing means for generating a dynamic language model.

According to a second aspect of the present invention, there is provided a statistical language model generating apparatus according to the first aspect, wherein a plurality of words connected before a word to be processed are connected based on the learning text data. A third generation means for generating a statistical language model of the backward word N-gram by calculating a third frequency probability that a word string consisting of the word string occurs, the learning text data, and the part of speech class The backward part of speech class-statistic of word N-gram is calculated based on the information data by calculating the fourth frequency probability that a word string composed of a plurality of words connected before occurs from the part of speech class of the word to be processed. Generating means for generating a statistical language model, and a statistical language model of the backward part-of-speech class-word N-gram generated by the fourth generating means in advance. Using the statistical language model of the backward word N-gram generated by the third generating means as the posterior knowledge, by the maximum posterior probability estimating method, by using the third frequency probability and the fourth frequency probability. Executing a predetermined normalization process and a smoothing process based on the backward transition probability calculated by the second calculation unit, which calculates a backward transition probability by interpolating between And a second processing means for generating a statistical language model of the backward word N-gram.

According to a third aspect of the present invention, there is provided a statistical language model generating apparatus, comprising: a first frequency probability that a word to be processed occurs after a word string including a plurality of words; and a part of speech class of the first word. And a forward N-gram including a forward transition probability obtained by interpolating, after a word string including a plurality of words connected after the first word, a second frequency probability at which a processing target word occurs. Based on the transition probability of the statistical language model of the above, the forward N-gram statistical language model transition probability is assigned as a feature amount to each word string connected before the processing target word, First clustering means for performing clustering so as to prevent variation in the feature amount of the class from becoming small and generating class classification information after clustering; and a plurality of words connected before the word to be processed. Backward transition obtained by interpolating the third frequency probability of occurrence of a word sequence consisting of the following and the fourth frequency probability of occurrence of a word sequence consisting of a plurality of words connected before from the part of speech class of the word to be processed Backward Ng with probability
The backward N-gram is used as a feature for each processing word based on the transition probability of the statistical language model of gram.
A clustering means for allocating the transition probabilities of the statistical language model of (a) and performing clustering so that the variation in the feature amount of each class is not reduced, and generating class classification information after clustering; The class classification information generated by the first clustering means and the second clustering means is used as a processing target based on the learning text data in which the spoken sentence has been written, and a class of a word string before the processing target word is used. Generating a statistical language model of the fusion N-gram by calculating the frequency probability of the word to be processed to the class.
And a generation means.

According to a fourth aspect of the present invention, in the statistical language model generating apparatus according to the third aspect, the forward transition probability obtained by the interpolation is generated by the first processing means. Forward-looking N-gram
The transition probability of the backward language obtained by the interpolation is the transition probability of the statistical language model of the backward N-gram generated by the second processing means. The learning text data used in the fifth generation means is the learning text data used in the first to fourth generation means.

Further, the speech recognition apparatus according to the fifth aspect of the present invention is provided with a speech recognition means for recognizing a speech using a predetermined statistical language model based on a speech signal of an input uttered speech sentence. 3. The second processing unit according to claim 2, wherein the voice recognition unit uses a forward-looking N-gram statistical language model generated by the first processing unit according to claim 1. 4. Speech recognition using a backward N-gram statistical language model generated by the fifth embodiment or using a fused N-gram statistical language model generated by the fifth generation means according to claim 3 or 4. It is characterized by doing.

[0015]

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

FIG. 1 is a block diagram of a continuous speech recognition apparatus according to one embodiment of the present invention. The continuous speech recognition device according to this embodiment executes the first language model generation process of FIG. 4 to generate a forward language N-gram statistical language model and a backward N-gram statistical language model. Model generation unit 20 (first embodiment)
And executing the language model generation process of FIG. 7 based on the two generated statistical language models.
Language model generation unit 30 that generates a statistical language model that is a fusion N-gram that fuses two statistical language models
(Second Embodiment).
Here, an embodiment in which the switch SW is switched to the contact a side and the word hypothesis narrowing unit 6 performs the word hypothesis narrowing process using the forward N-gram in the statistical language model memory 21 is referred to as a first embodiment. On the other hand, the switch SW is switched to the contact b side so that the backward N-g in the statistical language model memory 31 is changed.
An embodiment in which the word hypothesis narrowing unit 6 performs word hypothesis narrowing processing using ram is referred to as a second embodiment.

This embodiment discloses a method of generating a new N-gram model that satisfies the four requirements simultaneously by compensating for the disadvantages of the word N-gram, the part-of-speech class N-gram, and the automatic class N-gram. First,
In the first embodiment, first, information of a part-of-speech class is added to a word N-gram by a known maximum posterior probability estimation method (hereinafter, referred to as a MAP estimation method).
"Reliability for sparse data" and "robustness for task deviation" while maintaining "accuracy of next word prediction"
Create a statistical language model with. Subsequently, the transition probability between words obtained by this statistical language model is regarded as a feature amount of the word, and based on this, a “compact model size” is added by automatically performing clustering and class classification. , Generate a statistical language model that is a fused N-gram that simultaneously satisfies the four requirements listed above.

First, a description will be given of a continuous interpolation process between part of speech class information and word information. Consider using the part-of-speech class information to compensate for problems in the word N-gram, “reliability for sparse data” and “robustness for task deviation”. The reliability of the word N-gram depends on the appearance frequency of the N-1 word strings immediately before used for estimation. Therefore, if a word string having a small number of appearances is supplemented with information on the part of speech class, an improvement in reliability can be expected without lowering the accuracy, and at the same time, the part of speech class N-
It is conceivable that the gram can be made more robust against task deviation. The following methods are conceivable as means for realizing this method. (A) Word N is determined by a threshold based on the number of appearances of a word string
A method of switching between the -gram and the part-of-speech class N-gram, (b) a method using a linear combination of both, and (c) a method used for back-off smoothing.

As the opinion of the inventor, as shown in FIG.
Between the part of speech class N-gram and the word N-gram, the probability estimation value using the immediately preceding word string and the probability estimation value using the immediately preceding part of speech class string are calculated according to the appearance frequency of the immediately preceding word string. Continuous interpolation can provide higher accuracy and reliability. Therefore, for this purpose, a part-of-speech class and an N-gram from a word string to a word (hereinafter, part of speech class-word N-
Gram. )think of. Word strings (w ₁ , w ₂ , ...,
w _n−1 , w _n ), the part-of-speech class-word N-gram consists of the part-of-speech class and the word string (c ₁ , w ₂ ,
, W _n-1 ) to the word w _n , and the value is expressed by the following equation. Here, c ₁ is the part of speech class of the word w ₁ .

[0020]

## EQU2 ## p (w _n | c ₁ , w ₂ ,..., W _n-1 )

Then, this part-of-speech class-word N-gra
A known MAP estimation method using m as prior knowledge and the word N-gram as posterior knowledge will be used. When using the MAP estimation method, it is necessary to know in advance what distribution the value of the word N-gram, which is a posterior event, follows.
Then, it is assumed that this distribution, that is, the prior distribution, follows the beta distribution of the following equation.

[0022]

[Equation 3] {1 / B (α, β)} × p ^{( α −1)} (1-p) ^{( β −1)}

The reason for using the beta distribution is that various distribution shapes including a uniform distribution can be represented by changing the parameters α and β. Word strings (w ₁ ,
The number of observations of w ₂ ,..., w _i ) is represented by C (w ₁ , w _{2 ,.}
, The forward transition probability p _MAP (w _n | w ₁ , w ₂ ) from the word string (w ₁ , w ₂ ,..., W _n-1 ) after the MAP estimation processing by the MAP estimation method to the word w _n ,..., W _n-1 ) are as follows.

[0024]

## EQU4 ## p _MAP (w _n | w ₁ , w ₂ ,..., W _n-1 ) = ｛C
(W ₁ , w ₂ ,..., W _n-1 ) + α-1} / {C (w ₁ ,
w ₂ ,..., w _n−1 , w _n ) + α + β−2｝

In the beta distribution, the mean μ and the variance β
Is known to be:

[0026]

## EQU5 ## μ = α / (α + β)

Σ ² = αβ / {(α + β) ² (α + β + 1)}

However, it is difficult to obtain these values by maximum likelihood estimation. First, the number of observations C
(W ₁ , w ₂ ,..., W _n-1 ) and the number of observations C (w ₁ , w ₂ ,
, W _n ) are both 0, that is, the probability estimation value when the posterior knowledge is not given is obtained by substituting the following equation 7 into the above equation 4 and the equation 8 below.

[0028]

C (w ₁ , w ₂ ,..., W _n-1 ) = C (w ₁ , w ₂ ,
…, W _n ) = 0

## _EQU8 ## p _MAP (w _n | w ₁ , w ₂ ,..., W _n-1 ) = (α−
1) / (α + β-2)

When the posterior knowledge is not given, it is appropriate to use the prior knowledge, so that the value of the above equation 8 is equal to the value based on the prior knowledge, and the following equation is obtained.

[0030]

(Α-1) / (α + β-2) = p (w _n | c ₁ ,
w ₂ , ..., w _n-1 )

Here, the probability p (w _n | c ₁ , w ₂ ,..., W
_n-1 ) is the probability p (w _n | w _i , w ₂ ,
.., W _n-1 ) (where w _i ∈c _i ) is obtained as a weighted average μ _h . Here, in order to easily find a solution for the parameters α and β, the relationship between the average μ of the prior distribution of Equation 5 and the weighted average μ _h in Equations 8 and 9 is expressed by Equations 10 and 11 below. This is formally applied to the above equation (6), and the following equation (12) is assumed.

[0032]

## EQU10 ## α → α-1

[Equation 11] α + β → α + β-2

(12) {(α-1) (β-1)} / ｛(α + β-
2) ² (α + β-1)｝ = σ _h ²

Here, σ _h ² is the weighted variance of the word N-gram. From Equations 8, 9, and 12, the parameters α and β can be obtained according to the following equations.

[0034]

Α-1 = ｛μ _h ² / (1-μ _h )｝ σ _h ² -μ _h

Α + β−2 = [{μ _h (1-μ _h )} / σ _h ² ] −1

The weighted average μ _h and the weighted variance σ _h ² can be calculated by being defined by the following equations.

[0036]

(Equation 15)

(Equation 16)

The parameter α calculated by the method described above
Using the above equation (4) on the basis of
The forward probability after P estimation is determined.

Next, the backward transition probability will be described. The backward transition probability p _MAP (w _n−1 , w _n ) from the word w _n after the MAP estimation processing by the MAP estimation method to the word string (w _n−1 , w _n−2 ,..., W ₂ , w ₁ ) _{-2, ..., w 2, w} 1 | w n) is given by the following equation.

[0039]

## _EQU17 ## p _MAP (w _n−1 , w _n−2 ,..., W ₂ , w ₁ | w _n )
= {C (w _n , w _n−1 ,..., W ₂ , w ₁ ) + α−1} / {C
(W _n ) + α + β−2｝

Here, a probability estimation value when no posterior knowledge is given is obtained by substituting the following equation (18) into the above equation (17).

[0041]

C (w _n , w _n−1 ,..., W ₂ , w ₁ ) = C (w _n ) = 0

(19) p _MAP (w _n−1 ,..., W ₂ , w ₁ | w _n ) = (α
-1) / (α + β-2)

When the posterior knowledge is not given, it is appropriate to use the prior knowledge, so the value of the above Expression 19 is equal to the value based on the prior knowledge, and the following equation is obtained.

[0043]

(Α-1) / (α + β-2) = p (w _n−1 ,
…, W ₂ , w ₁ | c _n )

The parameters α and β are calculated by the above equation (13).
And Equation 14 where the weighted average μ
_h and the weighted variance σ _h ² can be calculated by being defined by the following equation.

[0045]

(Equation 21)

(Equation 22)

The parameter α calculated by the method described above
And β using the above equation (17), M
The backward probability after AP estimation is determined.

The N obtained after the MAP estimation thus obtained is
The value of -gram is closer to the value of the word N-gram in a word string in which the number of times of observation of the word N-gram is large, and is close to the value of the part of speech class-word 2-gram in a word string of few words.
This model is considered to be a model having three of "prediction accuracy of next word", "reliability for sparse data", and "robustness against task deviation".

Next, the normalization processing and the smoothing processing will be described. Since the probability after the MAP estimation is calculated independently for each word string, there is no guarantee that the sum of the transition probabilities for each preceding word string is 1. For this reason, it is necessary to perform correction so that this becomes 1. N-g after MAP estimation
Even in ram, since a probability value cannot be assigned to a word string to which no prior knowledge is given, it is necessary to allocate some probability value by smoothing. As the smoothing method, preferably, one of the following two methods is used. (1) Smoothing processing by MAP estimation (2) Back-off smoothing processing

First, the smoothing process based on the MAP estimation will be described. Probability p which is prior knowledge in MAP estimation
(W _n | c ₁ , w ₂ ,..., W _n-1 ), the probability p (w _n | w ₂ ,
, W _n-1 ), a probability value can be assigned even when no prior knowledge is given (for example, prior art document 2 "Go Kawabata et al.," Back-off smoothing of N-gram language model based on distribution ", IEICE Technical Report, SP93-95, pp. 1-6, December 1995
Month ". ). Further, the probability p (w _n | w ₂ ,.
w _n-1 ) is the probability p (w _n | c ₂ ,..., w _n-1 ) and the probability p
(W _n | c ₂ ,..., W _n-1 ) is the probability p (w _n | w ₃ ,.
_n-1 ), the MAP estimation is repeated sequentially, and finally the word 1-gram (or 0-gr) is obtained.
Since the MAP estimation using the prior knowledge am) is performed, probability values can be assigned to all word strings. Note that in this case, the weighted average μ
_As the value of _h , the probability estimation value p _MAP (w _n |
c ₁ , w ₂ ,..., w _{n -1} ). The normalized probability value is normalized with respect to each preceding word string according to the following equation so that the sum of transition probabilities to the next word becomes 1. Here, V represents the number of words.

[0050]

(Equation 23)

Next, the back-off smoothing processing will be described. As a smoothing method, a known back-off smoothing method (for example, Prior Art Document 3 “SM Katz,” Estimation
of Probabilities from Sparse Data for the Language
Model Component of a Speech Recognitizer ”, IEEE T
ransaction on Acoustics, Speech and Signal Process
ing, pp. 400-401, 1987 ". )), The sum of the probability values assigned to the unlearned columns is equal to the sum of the probability values assigned to the unlearned columns by a known back-off smoothing method in a model used as prior knowledge. I do. That is, the back-off coefficient b (w ₁ , w ₂ ,..., W _n-1 ) after the MAP estimation is equal to the back-off coefficient b (c ₁ , w ₂ ,..., W _n-1 ) in the prior knowledge. And After that, a discount coefficient is calculated so that the sum of the transition probabilities to the succeeding word becomes 1. Finally, the transition probability after the MAP estimation after the back-off smoothing processing is given by the following equation.

(1) When C (w ₁ , w ₂ ,..., W _n )> 0, that is, when posterior knowledge is given

[0053]

## _EQU24 ## p _MAP (w _n | w ₁ , w ₂ ,..., W _n-1 ) = d
(W ₂ ,..., W _n−1 , w _n ) × ｛C (w ₁ , w ₂ ,.
w _n−1 , w _n ) + α−1} / {C (w ₁ , w ₂ ,...,
w _n-1 ) + α + β-2｝

(2) C (c ₁ , w ₂ ,..., W _n )> Cut
In the case of (n), that is, when no ex post knowledge is given

[0055]

[Number 25] _{p MAP (w n | w 1} , w 2, ..., w n-1) = d
(W ₂ ,..., W _n−1 , w _n ) × (α−1) / ΔC (w ₁ , w
₂ ,..., W _n-1 ) + α + β-2｝

(3) C (c ₁ , w ₂ ,..., W _n ) ≦ Cut
In the case of (n), that is, when even prior knowledge is not reliable

[0057]

(26) p _MAP (w _n | w ₁ , w ₂ ,..., W _n-1 ) = b
(W ₁ , w ₂ ,..., W _n-1 ) × p _MAP (w _n | w ₂ ,.
_n-1 )

Here, Cut (n) is a cutoff coefficient at the order n of N-gram, and d (w ₂ ,..., W
_n−1 , w _n ) is a discount coefficient and b (w ₁ , w _n )
w ₂ ,..., w _n-1 ) are back-off coefficients.

In the case of the backward probability, the smoothing process can be performed using the same equation. Here, the value of the weighted average μ in Equation 20 is
Value p _MAP after MAP estimation (w _n−1 ,..., W ₂ , w ₁ | c _n )
Is used. In the backward MAP estimation, the statistical language model of the backward part of speech class-word N-gram is used as prior knowledge, and the statistical language model of the backward word N-gram is used as posterior knowledge.

Next, a method of using part-of-speech class information and automatic class classification in the second embodiment will be described. The two statistical language models of the forward N-gram and the backward N-gram obtained by the language model generation unit 20 of the first embodiment are automatically clustered (automatic class classification) to obtain a model. We will reduce the size. Whether or not the automatic class N-gram is robust against a task shift is considered to be determined by whether or not a feature amount assigned to each word is robust against a task shift when performing automatic class classification. In the normal automatic class classification, the value of the word N-gram is used as the feature amount of the word. However, since this feature amount lacks robustness against task deviation, the obtained automatic class N-gram is also robust against task deviation. It is thought to be lacking. Since the statistical language model of N-gram after the MAP estimation obtained by the language model generation unit 20 of the first embodiment has robustness against task deviation, this value is used as a feature amount of a word. It is expected that an automatic class that is robust against a task shift is obtained by the class N-gram using this class, and the statistical language model obtained by the language model generation unit 20 of the first embodiment is obtained. On the other hand, it can be expected that "compact model size" will be added.

Next, regarding the class classification of the preceding word string
explain. A certain word string (w₁, W_Two, ..., w_n-1Think)
When this word string (w₁, W_Two, ..., w_n-1) From other
Word w _nIs given by the following equation.

(27) p _MAP (w _n | w ₁ , w ₂ ,..., W _n-1 )

At this time, all V words in the vocabulary set are considered as transition destination words, and these transition probability sequences can be considered as vectors of length V. Since this vector can be considered to represent a connection attribute for the next word of the word string (w ₁ , w ₂ ,..., W _n-1 ), the word string is classified based on this vector. Here, the following procedure is used as a method of automatic class classification.

<Step SS1> One word string is classified into one class. <Step SS2> A feature value V (X) is assigned to the class X of each word string. Here, the feature amount V (X) is a vector having the transition probability from the class X to the next word as an element. <Step SS3> A pair of classes that minimizes the merge cost (U _new −U _old ) necessary for merging (that is, the variance of the vector after merging is not reduced) is selected and integrated to 1 And two classes. Here, U _new is the variation or variance of the vector in the class after merging, and U _old is the variance or variance of the vector in the class before merging, and can be calculated as follows.

[0064]

[Equation 28]

(Equation 29)

Here, C _new is the class after merging, C _old is the class before merging, and D (Vc, V
w) represents the square of the Euclidean distance between the vectors Vc and Vw. <Step SS4> The clustering process is performed by repeating the procedures of steps SS2 and SS3 until the number of classes reaches a predetermined number.

Next, the class classification of the succeeding word will be described.
I will tell. Word string (w₁, W_Two, ..., w _n-1) From other words
w_nTransition probability to the word string (w₁, W_Two, ..., w_n-1)
Even if it is parse (when there is not enough learning amount), MAP
A more reliable value could be assigned by the rule. I
However, the following word w_nIs reliable when sparse
The problem has not been solved. Therefore, a certain word w_nAhead of
The word string (w₁, W_Two, ..., w_n-1) Probability p
(W₁, W_Two, ..., w_n-1| W_n) To the word w_nPart of speech
Word strings (w₁, W_Two, ..., w_n-1) Comes out
Probability p (w₁, W_Two, ..., w_n-1| C_nA) prior knowledge
It is assumed that interpolation is performed using MAP estimation. MAP estimation
The procedure itself uses the method described above. This gives
Probability p after MAP estimation_MAP(W₁, W_Two, ..., w_n-1| W
_n) Is a vector with the word
w_nPreceding word string (w₁, W_Two, ..., w_n-1Trust in)
Represents connection attributes that have potential. Based on this vector
The subsequent words are also classified. Obtained in this way
Preceding word string class and succeeding word class (both processing
Elephant words) together with word information
The class also takes into account information
Class N-gram generated from the word N-gram
N-gra that fuses the part-of-speech class N-gram
We call it m.

Further, the class based on the part of speech will be supplementarily described. Part-of-speech class information is often used as a class index in the class N-gram. Therefore, the present inventor first performed a performance comparison between the class bigram and the word bigram, which were classified based on the part-of-speech class information. The training set has a total of 459,383 words,
The difference is 7,221 words, and the class used is 158 classes based on part-of-speech class information (stored in the part-of-speech class information data memory 14 in the present embodiment).
The experimental results show that the class bigram has a perplexity of 31.53 in a test set with a total of 6,826 words.
Whereas the word bigram has 18.51.
It can be seen that there is a considerable difference between the two.
It is considered that the reason is that the part of speech information used for the class classification represents the entire property of the word, and thus it cannot be said that the connectivity of important words is purely expressed in the N-gram.

In view of this situation, the present inventor has invented a multi-class N-gram fusion N-gram shown below. Here, the difference between the case where class classification based on part of speech information is performed and the case where class classification is performed focusing only on the connectivity of words is considered. As an example, the following three verb conjugation endings, which are one of the parts of speech used in all sections, will be considered. (A) "ki: ending, five columns, continuous form" (b) "shi: ending, five columns, continuous form" (c) "ku: ending, five columns, closing form" In order to have parts of speech (here, they are classified into three parts of speech), they are classified into different classes in the class classification based on part of speech information. In this case, three types of connections are to be considered for what word comes before or after. However, if we only consider what words come before them, "ki" and "ku" can be considered the same, and if we only consider what words come after, "Ki" and "shi" can be considered the same. Therefore, if the words are preceded or deceased individually, two classes of connections can be considered if each class is assigned separately. Classification can be performed more efficiently.

In this embodiment, a class relating to the backward connectivity (from connectivity) of the former word, which is the former, is called a to class (backward class), and the forward connectivity of the subsequent word, which is the latter, is referred to as the to class. (Connectivity of to)
It will be referred to as an m class (forward class), and each word will be considered to have two class attributes (part of speech attribute) of a to class and a from class. According to this, the above three words can be expressed in the to class in two classes: (a) “ki, ku: ending, five columns” (b) “shi: ending, five columns” In the from class, it can be expressed in two classes: (a) "Ki-shi: ending, 5-dan, continuous form". (b) "K: ending, 5-dan, ending form". This concept can be extended as it is even when N ≧ 3. In this case, each word has N class attributes. This class attribute is called a multiple class, and the N-gra
m is called a multi-class N-gram. The appearance probability of the multi-class bigram when N = 2 is expressed by the following equation.

[0070]

P (Wn | Wn-1) ｎP (Ct (Wn) |
Cf (Wn-1)) × P (Wn | Ct (Wn))

Here, Ct represents the to class to which the word to be processed belongs, and Cf represents the from class. The number of parameters at this time is class N-gra
m is the square of the number of classes plus the number of words, whereas to
The number of classes × the number of classes + the number of words.

Further, the statistical language model generation processing executed by the language model generation units 20 and 30 will be described with reference to the flowcharts of FIGS.

FIG. 4 is a flowchart showing a first language model generation process executed by the language model generation unit 20 of FIG. In FIG. 4, first, a forward language model generation process (FIG. 5) is executed in step S1, and then a backward language model generation process (FIG. 6) is executed in step S2, and the first language model generation process ends. I do.

FIG. 5 is a flowchart showing a forward language model generation process which is a subroutine of FIG. FIG.
First, at step S11, based on the learning text data stored in the learning text data memory 13 and the corpus in which the uttered voice sentence is newly written,
After generating the statistical language model of the forward word N-gram, in step S12, based on the learning text data in the learning text data memory 13 and the part of speech class information data in the part of speech class information data memory 14, Generate a statistical language model of the forward part of speech class-word N-gram. Next, in step S13, for each pair of the part of speech class and the word string,
6 and the parameters α and β for the MAP estimating method are calculated. In step S14, M is calculated for each word string by using Equation 4 based on the calculated parameters α and β.
A forward transition probability after AP estimation is calculated. Further, the forward transition probability calculated in step S15 is subjected to normalization processing and smoothing processing to generate a forward language model, and the statistical language model memory 2
1 and the program returns to the main routine.

FIG. 6 is a flowchart showing the backward language model generation process which is a subroutine of FIG. FIG.
First, in step S21, a backward word N-gram is generated based on the learning text data in the learning text data memory 13, and then in step S21.
At 22, the learning text data in the learning text data memory 13 and the part-of-speech class information data memory 14
A backward part-of-speech class-word N-gram is generated on the basis of the part-of-speech class information data. Next, in step S23, for each pair of the part-of-speech class and the word string, MAP is performed using Expressions 13, 14, 21, and 22.
The parameters α and β for the estimation method are calculated, and in step S24, the backward transition probability after the MAP estimation is calculated for each word string using Equation 17 based on the calculated parameters α and β. Further, by performing a normalization process and a smoothing process on the backward transition probability calculated in step S25, a backward language model is generated and stored in the statistical language model memory 22.
Return to the original main routine.

FIG. 7 is a flowchart showing the second language model generation processing executed by the language model generation unit 30 of FIG. In FIG. 7, first, at step S31
In step S32, a clustering process for the forward-looking language model (FIG. 9) is executed, and in step S32, a clustering process for the forward-looking language model (FIG. 9) is executed.
At 33, the fusion language model generation process (FIG. 10) is executed, and the second language model generation process ends.

FIG. 8 is a flowchart showing the forward language model clustering process (step S31) which is a subroutine of FIG. 8, first, in step S41, based on the transition probability of the forward-looking language model in the statistical language model memory 21, the above-mentioned transition probability is used as a feature amount for each word string connected before the processing target word. Here, one class is assigned to one word string. Next, in step S42, the pairs of the classes are merged to select the pair of the classes with the minimum merge cost, and are integrated into one class.
It is determined whether or not a predetermined class number threshold (for example, 200 or 500) has been reached. Here, if the determination is NO, the process returns to step S42, while if the determination is YES, the process proceeds to step S44, where the class classification information of the word string preceding the processing target word after clustering is stored in the temporary memory, and the process returns to the original main routine. .

FIG. 9 is a flowchart showing the backward language model clustering process (step S32) which is a subroutine of FIG. In FIG. 9, step S
At 51, based on the transition probability of the backward language model in the statistical language model memory 22, the above-mentioned transition probability is assigned to each processing target word as a feature amount, where one word string is classified into one class. . Then, step S
At step 52, the pairs of the classes are merged to select the pair of the class with the minimum merge cost, and are integrated into one class. At step S53, it is determined whether or not a predetermined class number threshold is reached. Here, if the determination is NO, the process returns to step S52, whereas if the determination is YES, the process returns to step S54.
Then, the class classification information of the processing target word after clustering is stored in the temporary memory, and the process returns to the original main routine.

FIG. 10 is a flowchart showing a fusion language model generation process which is a subroutine of FIG. FIG.
In step S61, first, in step S61, the class classification information of the word string preceding the clustering target word and the class classification information of the clustering target word stored in the temporary memory are processed. Next, in step S62, based on the learning text data in the learning text data memory 13, the classes of the word string preceding the processing target word are extracted from the class of the word string preceding the processing target word using each of the two classes of class classification information. By calculating the frequency probabilities for the classes, the word N-
A statistical language model of the fusion N-gram in which the gram and the part-of-speech class N-gram are fused is generated, stored in the statistical language model memory 31, and the process returns to the original main routine.

Next, the configuration and operation of the continuous speech recognition apparatus shown in FIG. 1 will be described. In FIG. 1, the phoneme HMM in the phoneme hidden Markov model (hereinafter, referred to as HMM) memory 11 connected to the word matching unit 4 is represented by including each state, and each state includes the following information. Having. (A) state number, (b) acceptable context class, (c) list of preceding and succeeding states, (d) parameters of output probability density distribution, and (e) self-transition probability and transition to succeeding state probability. Note that the phoneme HMM used in the present embodiment is generated by converting a predetermined speaker-mixed HMM because it is necessary to specify which speaker each distribution originates from. Here, the output probability density function is a Gaussian mixture distribution having a 34-dimensional diagonal covariance matrix. Further, the word dictionary in the word dictionary memory 12 connected to the word matching unit 4 stores a symbol string indicating a reading represented by a symbol for each word of the phoneme HMM in the phoneme HMM memory 11.

In FIG. 1, a uttered voice of a speaker 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 word matching unit 4 via the buffer memory 3.

The word collating unit 4 uses the one-pass Viterbi decoding method to store the phoneme HMM memory 1 based on the characteristic parameter data input via the buffer memory 3.
A word hypothesis is detected by using the phoneme HMM in 1 and the word dictionary in the word dictionary memory 12, and the likelihood is calculated and output. Here, the word matching unit 4 determines, for each state of each HMM at each time,
The likelihood within a word and the likelihood from the start of utterance are calculated. The likelihood is individually provided for each word identification number, word start time, and difference between preceding words. Further, in order to reduce the amount of calculation processing, a grid hypothesis having a low likelihood among the total likelihoods calculated based on the phoneme HMM and the word dictionary is reduced. The word collating unit 4 outputs the resulting word hypothesis and likelihood information to the word hypothesis narrowing unit 6 via the buffer memory 5 together with time information (specifically, a frame number, for example) from the utterance start time. .

Here, when the switch SW is switched to the a side, the statistical language model memory 21 stores the word hypothesis narrowing section 6
And the word hypothesis refining unit 6 is connected to the language model generation unit 2
0 while performing a word hypothesis narrowing process by referring to a statistical language model that is a forward-looking N-gram in the statistical language model memory 21 generated when the switch SW is switched to the b side. The memory 31 is connected to the word hypothesis narrowing unit 6, and the word hypothesis narrowing unit 6 is configured to use the statistical language model memory 31 generated by the language model generation unit 30.
The word hypothesis narrowing process is performed with reference to the statistical language model which is the fusion N-gram in the above.

The word hypothesis narrowing section 6 refers to the statistical language model in the statistical language model memory 21 or 31 based on the word hypothesis output from the word matching section 4 via the buffer memory 5, and For a word hypothesis of the same word having the same end time and different start time, the highest likelihood among the total likelihoods calculated from the utterance start time to the end time of the word for each head phoneme environment of the word. After narrowing down the word hypotheses so as to be represented by one word hypothesis having a degree, of the word strings of all the narrowed word hypotheses,
A word string of a hypothesis having the maximum total likelihood is output as a recognition result. The task-adapted statistical language model includes one statistical language model for each task, and the word hypothesis narrowing unit 6 selectively selects a statistical language model corresponding to the task to be subjected to speech recognition. Refer to In the present embodiment, preferably, the first phoneme environment of the word to be processed includes the last phoneme of the word hypothesis preceding the word and the first two phonemes of the word hypothesis of the word.
One phoneme.

[0085] For example, as shown in FIG. 2, the (i-1) th word W _i-1 of the following phoneme string a _1, a _2, ..., come i th word W _i consisting a _n Sometimes, six hypotheses Wa, Wb, Wc, Wd, We, and Wf are assumed as the word hypotheses of the word Wi _-1.
Exists. Here, the former three word hypotheses Wa, W
It is assumed that the final phonemes of b and Wc are / x /, and the final phonemes of the latter three word hypotheses Wd, We and Wf are / y /. The hypothesis with the highest total likelihood (for example, FIG. 2) is the hypothesis in which the end time t _e is equal to the first phoneme environment (the top three word hypotheses in which the _first phoneme environment is “x / a ₁ / a ₂ ” in FIG. 2). 2
Are deleted except for the top hypothesis). Since the fourth hypothesis from the top has a different phoneme environment, that is, since the last phoneme of the preceding word hypothesis is y instead of x,
Do not delete the fourth hypothesis from the top. That is, only one hypothesis is left for each final phoneme of the preceding word hypothesis. In the example of FIG. 2, one hypothesis is left for the final phoneme / x /, and one hypothesis is left for the final phoneme / y /.

In the above embodiment, the head phoneme environment of the word is defined as a sequence of three phonemes including the last phoneme of the word hypothesis preceding the word and the first two phonemes of the word hypothesis of the word. Although defined, the present invention is not limited to this. The phoneme sequence of the preceding word hypothesis including the final phoneme of the preceding word hypothesis, and at least one phoneme of the preceding word hypothesis that is continuous with the final phoneme, And a phoneme sequence that includes a phoneme sequence that includes the first phoneme of the word hypothesis.

In the above embodiment, the feature extraction unit 2
The word matching unit 4, the word hypothesis narrowing unit 6, and the language model generation units 20 and 30 are configured by a computer such as a digital computer, for example, and include buffer memories 3 and 5.
A phoneme HMM memory 11, a word dictionary memory 12,
Learning text data memory 13, part of speech class information data memory 14, statistical language model memories 21,
Each of the reference numerals 2 and 31 includes a storage device such as a hard disk memory.

In the above embodiment, speech recognition is performed using the word collating unit 4 and the word hypothesis narrowing unit 6.
The present invention is not limited to this. For example, the present invention includes a phoneme matching unit that refers to the phoneme HMM 11 and a speech recognition unit that performs speech recognition of a word by referring to a statistical language model using, for example, the One Pass DP algorithm. Is also good.

In the above embodiment, the forward N-gram or the fusion N-gram is used as the statistical language model used in the word hypothesis narrowing unit 6, but the backward N-gram in the statistical language model memory 22 is used. An N-gram may be used, or a forward N-gram and a backward N-gram may be used.
A statistical language model of N-gram having an average value of each transition probability of the gram is generated, and is generated by the word hypothesis narrowing unit 6.
May be used.

[0090]

EXAMPLES <Experiment of First Embodiment and Results thereof> The present inventors conducted an evaluation experiment in order to confirm the effectiveness of continuous interpolation of part of speech information and word information by MAP estimation. The dimension of N-gram in the experiment used 2-gram, and the smoothing used the back-off of the cut-off coefficient 0. The comparison targets were word 2-gram as post-knowledge, part-of-speech class-word 2-gram as pre-knowledge, and a linear combination model thereof. The learning set, which is learning text data and part-of-speech class information data, is a Japanese dialogue sentence consisting of about 260,000 total words and about 4,000 different words.
It is classified into eight categories. The evaluation set includes 16 conversations of the same task as the learning set, 2178 evaluation sets of 2178 words, and 24 conversations of tasks different from the learning set.
Two types of evaluation set B of 55 words were used. Evaluation set A in each model, which is an experimental result by the inventor,
Table 2 shows the perplexity in B.

[0091]

[Table 2] Performance (perplexity) in the evaluation set of each model ―――――――――――――――――――――――――――――――― Test set A Test set B (Same task) (Different tasks) ―――――――――――――――――――――――――――――――――― Words 2-gram 14.39 53.06 ―――――――――――――――――――――――――――――――― Part of speech class-word 2-gram 22.00 58.74 ―――――――――――――――――――――――――――――――― 2-gram 14.00 linearly combined 45.69 ―――――――――――――――――――――――――――――――― MAP estimated 2-gram 13.33 43.76 (First implementation State) ----------------------------------

As is clear from Table 2, the model after the MAP estimation has the lowest perplexity for both evaluation sets, and is particularly remarkable for evaluation set B. For this reason, in the model in which part of speech information and word information are continuously interpolated using MAP estimation, “robustness to task deviation” is added while maintaining the next word prediction accuracy of the word 2-gram. The effect is larger than that of the linear combination model.

<Experiment of Second Embodiment and Results thereof>
MA obtained by the language model generation unit 20 of the embodiment
Part-of-speech N-gram using P estimation and automatic classification
And a word N-gram fusion N-gram model (second
Of the present embodiment). The evaluation conditions are the same as those of the first embodiment, and the comparison target is the word 2-gram and the class 2-gram in which only the automatic class classification is performed. The number of classes used was 500 and 200, respectively. FIG. 11 shows the relationship between the perplexity of the evaluation sets A and B for each model and the number of classes. In both the evaluation sets A and B, the perplexity decreases along with the decrease in the number of classes up to a certain number of classes, and the effect of the class classification using the feature quantity including the part of speech information appears. Particularly, in the evaluation set B, which is a different task, the perplexity is remarkably reduced, indicating that the part of speech information is effectively working as robustness against the task deviation. Table 3 shows the values of perplexity for the classes 200 and 500.

[0094]

[Table 3] Perplexity evaluation of fused 2-gram ―――――――――――――――――――――――――――――――― Set A Test set B (Same task) (Different task) ―――――――――――――――――――――――――――――――――― Word 2 −gram 14.39 53.06 ―――――――――――――――――――――――――――――――― Automatic clustering 500 15.12 59.10 Performed 2-gram ――――――――――――――――――――――― 200 16.62 58.01 ―――――――――――― ―――――――――――――――――――――― Fusion 2-gram 500 13.95 45.32 (Second embodiment) ―――――――――― ―――― ――――――――― 200 15.44 46.51 ――――――――――――――――――――――――――――――――――

As is evident from Table 3, the fused 2-gra
m is class 2-gram by automatic classification, word 2
-Shows low perplexity for both. In particular, for the evaluation set B, the word 2-gram
In contrast, the 15% perplexity has decreased, indicating the robustness to the task deviation. The logical parameter size at this time is less than 2% of the word 2-gram, and the number of entries is also １ ／. It can be seen that the four requirements described in the section of the prior art are simultaneously satisfied.

Next, the fused 2-gra of the second embodiment
The evaluation in m continuous word recognition was performed. The experimental conditions were
The conditions are the same as those described above, and the evaluation results are shown.

[0097]

[Table 4] Evaluation of word misrecognition rate of fusion 2-gram (second embodiment) ―――――――――――――――――――――――――――― ―――――― Test set A Test set B (Same task) (Different tasks) ――――――――――――――――――――――――――――― ――――― Word 2-gram 14.34% 32.76% ――――――――――――――――――――――――――――――――― ― Fusion 2-gram 12.37% 23.07% (200 class) ――――――――――――――――――――――――――――――――― ― Fusion 2-gram 12.02% 23.51% (500 class) ――――――――――――――――――――――――――――――――― ―

As is clear from Table 4, the fused 2-gra
The word misrecognition rate of m is lower than that of the evaluation set A or B, both words 2-gram, as in the case of perplexity. The decrease rate is as large as 16%, compared with 3% for the same task, and is remarkable as 28% for different tasks. From this, it can be seen that the fusion 2-gram is very effective for the same task and different tasks even in continuous word recognition.

As described above, according to the present embodiment, as a method of giving the reliability to sparse data and the robustness against task deviation to the word N-gram, the part-of-speech class N-gram is used as prior knowledge. A method using maximum posterior probability estimation was proposed. By using this method, a reliable value based on the value of the part-of-speech class N-gram can be given to a word pair having a small number of appearances. At the same time, the property of the part-of-speech class N-gram, which is robust against task deviation, can be inherited.

Further, by regarding the transition probability between words of this model as a feature amount of words and performing automatic class classification based on this, the parameter size can be reduced without deteriorating performance. This model uses the word Ng
2% of ram, 50% of entries, 3% for the same task, 15% low perplexity for different tasks, and 16% and 28% for word recognition in continuous word recognition Only low. From this, it was confirmed by experiments that the reliability of sparse data and the robustness against task deviation can be added and the model size can be reduced while maintaining the next word prediction accuracy of the word N-gram.

Therefore, as described in detail above, the forward N generated by the language model generation unit 20 of the first embodiment.
Statistical language models that are -gram cannot reduce the model size, but have good performance in prediction accuracy, reliability and robustness. Also, by performing speech recognition using this statistical language model, speech recognition can be performed with an improved speech recognition rate as compared with the conventional example.

The statistical language model, which is a fusion N-gram generated by the language model generation unit 30 of the second embodiment, has excellent performance in prediction accuracy, reliability, model size, and robustness. Also, by performing speech recognition using this statistical language model, speech recognition can be performed with an improved speech recognition rate as compared with the conventional example.

[0103]

As described above in detail, according to the statistical language model generating apparatus according to the first aspect of the present invention, based on the learning text data in which the uttered voice sentence of a predetermined speaker has been written, a plurality of First generation means for generating a statistical language model of a forward word N-gram by calculating a first frequency probability that a word to be processed occurs after a word string consisting of the words Based on the data and the part-of-speech class information data including the part-of-speech class information, the word to be processed is added after the part of speech class of the first word and the word string including a plurality of words connected after the first word. Second generation means for generating a forward-looking part-of-speech class-word N-gram statistical language model by calculating a second frequency probability of occurrence, and forward generation generated by the second generation means. Using the statistical language model of the part-of-speech class-word N-gram as prior knowledge, and the maximum posterior probability estimation using the statistical language model of the forward-looking word N-gram generated by the first generation means as posterior knowledge First calculating means for calculating a forward transition probability obtained by interpolating between the first frequency probability and the second frequency probability by the method, and the forward transition probability calculated by the first calculating means. And a first processing unit for generating a forward-looking N-gram statistical language model by executing predetermined normalization processing and smoothing processing. Therefore,
The generated statistical language model that is the generated forward N-gram cannot reduce the model size,
It has excellent performance in prediction accuracy, reliability and robustness. Also, by performing speech recognition using this statistical language model, speech recognition can be performed with an improved speech recognition rate as compared with the conventional example.

Further, according to the statistical language model generating device of the present invention, in the statistical language model generating device of the present invention, based on the learning text data,
Third generation means for generating a statistical language model of the backward word N-gram by calculating a third frequency probability that a word string composed of a plurality of words connected before the word to be processed occurs; Based on the learning text data and the part-of-speech class information data, a backward frequency is calculated by calculating a fourth frequency probability that a word string composed of a plurality of words connected before occurs from the part-of-speech class of the word to be processed. A fourth generation unit for generating a statistical language model of the part-of-speech class-word N-gram, and prior knowledge of a backward part-of-speech class-statistical language model of the word N-gram generated by the fourth generation unit Using the statistical language model of the backward word N-gram generated by the third generating means as the posterior knowledge, by the maximum posterior probability estimation method. Calculating a backward transition probability by interpolating between the frequency probability of the second and fourth frequency probabilities, and a predetermined value based on the backward transition probability calculated by the second calculating means. And a second processing unit that generates a statistical language model of the backward word N-gram by executing the normalization process and the smoothing process. Therefore, the generated generated backward N-gr
The statistical language model that is am cannot reduce the model size, but has excellent performance in prediction accuracy, reliability, and robustness. Also, by performing speech recognition using this statistical language model, speech recognition can be performed with an improved speech recognition rate as compared with the conventional example.

According to the statistical language model generating apparatus according to the third aspect of the present invention, the first frequency probability that a word to be processed occurs after a word string composed of a plurality of words, A forward N including a forward transition probability obtained by interpolating a part of speech class and a second frequency probability of occurrence of a word to be processed after a word string including a plurality of words connected after the first word; Based on the transition probability of the statistical language model of -gram, for each word string connected before the word to be processed, the forward-looking N-gra
first clustering means for allocating the transition probabilities of the statistical language model of m and performing clustering so that the variation in the feature amount of each class is not reduced, and generating class classification information after clustering; The third occurrence of a word string consisting of multiple words connected before
Of the backward N-gram including the backward transition probability obtained by interpolating the frequency probability of the word to be processed and the fourth frequency probability of the occurrence of a word string composed of a plurality of words connected before the part of speech class of the word to be processed. Based on the transition probability of the statistical language model, the backward N-
assigning the transition probabilities of the statistical language model of gram,
Based on second clustering means for performing clustering so that the variation in the feature amount of each class does not become small and generating class classification information after clustering, and learning text data in which uttered voice sentences of a predetermined speaker are written down With the class classification information generated by the first clustering unit and the second clustering unit as a processing target, the frequency probability from the class of the word string before the word to be processed to the class of the word to be processed is calculated. Fifth generation means for generating a statistical language model of the fusion N-gram by calculating. Here, preferably, the forward transition probability obtained by the interpolation is a transition probability of the forward N-gram statistical language model generated by the first processing means, and the backward transition probability obtained by the interpolation is used. The transition probability is a transition probability of a backward N-gram statistical language model generated by the second processing means, and the learning text data used in the fifth generation means is the first to fourth text data. Is the text data for learning used in the generation means. Therefore, the generated statistical language model, which is the generated fused N-gram, has excellent performance in model size, prediction accuracy, reliability, and robustness. Also, by performing speech recognition using this statistical language model, speech recognition can be performed with an improved speech recognition rate as compared with the conventional example.

Further, according to the speech recognition apparatus of the fifth aspect of the present invention, the speech recognition means for recognizing the speech using a predetermined statistical language model based on the speech signal of the input uttered speech sentence. In a speech recognition apparatus provided with the second aspect, the speech recognition means uses a forward-looking N-gram statistical language model generated by the first processing means according to claim 1. Using a statistical language model of the backward N-gram generated by the processing means, or using a statistical language model of the fusion N-gram generated by the fifth generation means according to claim 3 or 4; Recognize voice. Therefore, by performing speech recognition using the statistical language model, speech recognition can be performed with an improved speech recognition rate as compared with the conventional example.

[Brief description of the drawings]

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

FIG. 2 is a timing chart showing a process of a word hypothesis narrowing section 6 in the continuous speech recognition device of FIG.

FIG. 3 is a graph illustrating an interpolation process of a probability estimation value performed by a language model generation unit 20 of FIG. 1;

FIG. 4 is a flowchart illustrating a first language model generation process executed by a language model generation unit 20 of FIG. 1;

FIG. 5 is a flowchart showing a forward language model generation process which is a subroutine of FIG. 4;

FIG. 6 is a flowchart showing a backward language model generation process which is a subroutine of FIG. 4;

FIG. 7 is a flowchart illustrating a second language model generation process executed by the language model generation unit 30 of FIG. 1;

8 is a flowchart showing a clustering process of a forward-looking language model, which is a subroutine of FIG. 7;

FIG. 9 is a flowchart showing a backward language model clustering process which is a subroutine of FIG. 7;

FIG. 10 is a flowchart showing a fusion language model generation process which is a subroutine of FIG. 7;

11 is an experimental result of a fusion 2-gram language model (second embodiment) and a word 2-gram (conventional example) generated by the language model generation unit 30 of FIG. It is a graph which shows the relationship of plexity.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... Feature extraction part, 3, 5 ... Buffer memory, 4 ... Word collation part, 6 ... Word hypothesis narrowing part, 11 ... Phoneme HMM memory, 12 ... Word dictionary memory, 13 ... Text data memory for learning , 14 ... part-of-speech class information data memory, 20, 30 ... language model generator, 21, 22, 31 ... statistical language model memory, SW ... switch.

 ──────────────────────────────────────────────────続 き Continuing from the front page (72) Inventor Yoshinori Sakasaka 5 Sanraya, Inaya, Seika-cho, Sagara-gun, Kyoto F-term in ATR Speech Translation and Communication Research Laboratories Co., Ltd. 5B091 AA15 CB12 CC01 CC16 EA01 5D015 HH23

Claims (5)

[Claims] 1. A first frequency probability that a word to be processed occurs after a word string composed of a plurality of words, based on learning text data in which uttered voice sentences of a predetermined speaker are written. First generation means for generating a statistical language model of the forward-looking word N-gram; a part-of-speech class of the first word based on the learning text data and part-of-speech class information data including part-of-speech class information; By calculating a second frequency probability that a word to be processed occurs after a word string consisting of a plurality of words connected after the first word, a forward-looking part-of-speech class—a statistical language of the word N-gram A second generation unit for generating a model; and a forward-looking part-of-speech class-word N-gram statistical language model generated by the second generation unit as prior knowledge, Using the statistical language model of the forward-looking word N-gram generated by the generating means as the posterior knowledge and interpolating between the first frequency probability and the second frequency probability by the maximum posterior probability estimation method. First calculating means for calculating a forward transition probability; and performing forward normalization and smoothing processing on the forward transition probability calculated by the first calculating means, thereby obtaining a forward N- a first processing means for generating a statistical language model of the gram. 2. The statistical language model generating apparatus according to claim 1, wherein a third frequency at which a word string including a plurality of words connected before the word to be processed occurs based on the learning text data. By calculating the probability, the backward word N-
a third generation unit for generating a statistical language model of the gram; a plurality of words connected before the part of speech class of the word to be processed based on the learning text data and the part of speech class information data; A fourth generation unit that generates a statistical language model of the backward part-of-speech class-word N-gram by calculating a fourth frequency probability that the word string occurs; and a backward direction generated by the fourth generation unit. Using the statistical language model of the part-of-speech class-word N-gram as prior knowledge, and the maximum posterior probability estimation using the statistical language model of the backward word N-gram generated by the third generation means as posterior knowledge A second calculating means for calculating a backward transition probability obtained by interpolating between the third frequency probability and the fourth frequency probability by the method, Second processing means for generating a statistical language model of the backward word N-gram by executing a predetermined normalization process and a smoothing process based on the backward transition probability calculated in the above manner. A statistical language model generation device. 3. A first frequency probability that a word to be processed occurs after a word string including a plurality of words, a part of speech class of the first word, and a plurality of words connected after the first word. The word to be processed based on the transition probabilities of the forward-looking N-gram statistical language model including the forward transition probabilities obtained by interpolating the second word probability at which the word to be processed occurs after the word sequence The forward N-gram statistical language model transition probability of the forward N-gram is assigned as a feature amount to each word string connected before, and clustering is performed so that the variation in the feature amount of each class is not reduced. First to generate class classification information after clustering
Clustering means, a third frequency probability that a word string consisting of a plurality of words connected before the word to be processed occurs, and a word string consisting of a plurality of words connected before the part of speech class of the word to be processed Is generated as a feature amount for each processing word based on the transition probability of the backward N-gram statistical language model including the backward transition probability obtained by interpolating the fourth frequency probability of occurrence of A second clustering means for allocating a transition probability of a statistical language model of "gram" and performing clustering so that the variation of the feature amount of each class is not reduced, and generating class classification information after clustering; Generated by the first clustering means and the second clustering means based on the learning text data in which the uttered voice sentence of A statistical language model of the fusion N-gram is generated by calculating the frequency probability of the class of the word to be processed from the class of the word string before the word to be processed, using the class classification information as the processing target. A statistical language model generation device comprising: a fifth generation unit. 4. The statistical language model generation device according to claim 3, wherein the forward transition probability obtained by interpolation is obtained by calculating a forward N-gram statistical language model generated by the first processing means. The transition probability is the transition probability of the backward N-gram statistical language model generated by the second processing means, and is used in the fifth generation means. A statistical language model generation device, wherein the learning text data is learning text data used in the first to fourth generation means. 5. A speech recognition apparatus comprising speech recognition means for recognizing a speech using a predetermined statistical language model based on a speech signal of an input uttered speech sentence, wherein the speech recognition means comprises: A forward-looking N-gram statistical language model generated by the first processing unit according to claim 2 is used to generate a backward N-gram statistical language model generated by the second processing unit according to claim 2. A speech recognition apparatus characterized in that speech recognition is performed using the statistical language model of the fused N-gram generated by using the fifth generation means according to claim 3 or 4.