JP2002342323A - Language model learning device, voice recognizing device using the same, language model learning method, voice recognizing method using the same, and storage medium with the methods stored therein - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a language model learning device, which is enhanced in recognition accuracy. SOLUTION: The device is provided with object task language data 101, general task language data 102, a similar word couple extracting means 103, a similar word string synthesizing means 104 and a language model generating means 105 for constructing a task adapted language model by reading text data from the respective language data. Then, the similar word couple extracting means 103 reads the respective text data and extracts a similar couple from the combination of words contained in the object task data and words contained in the general task data, the similar word string synthesizing means 104 reads the respective text data and the similar word couple and synthesizes out a word string containing words in an object task, which is not contained in the language data, and the language generating means 105 finds the statistic quantity of the word string by weighting for every text data.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a language model learning device using a probabilistic language model, a speech recognition device using the same, a language model learning method, a speech recognition method using the same, and a speech recognition of those methods. The present invention relates to a storage medium storing a program.

[0002]

2. Description of the Related Art Generally, in speech recognition, a speech signal is converted into a time series of vectors that well represent acoustic characteristics of a speech by using a processing method of a digitized and inputted speech signal. Is performed.

[0003] The collation processing is an acoustic feature vector time series A (= [a ₁ , a ₂ ,...) Composed of K time frames.
a _K ]), and the uttered word sequence W (= [w ₁ , w
₂ ,..., W _M ], (M is the number of words)).

[0004] In the above collation processing, in order to estimate a word string W for which the recognition accuracy is highest, the appearance probability P
The recognized word string W ^* that maximizes (W | A) may be obtained by the following equation (1).

[0005]

(Equation 1)

However, in equation (1), the appearance probability P
It is usually difficult to directly determine (W | A). Therefore, the appearance probability P (W | A) is calculated using Bayes' theorem.
It can be rewritten as the following equation (2).

[0007]

(Equation 2)

Here, when finding the word string W that maximizes the left side of equation (2), the denominator P (A) on the right side does not affect the word string W that is a candidate for recognition. What is necessary is just to find the word string W to be maximized. That is, the recognition word string W
^* Is represented by the following equation (3).

[0009]

(Equation 3)

Here, the probability model giving P (W) and the probability model giving P (A | W) in equation (3) are called a language model and an acoustic model, respectively. In speech recognition, modeling methods that have been actively studied in recent years include:
It is known that an acoustic model is represented by a “Hidden Markov Model” and a language model is represented by a “probabilistic language model”.

The details of these modeling methods are described in, for example, "Basics of speech recognition (upper and lower)" (LR RABI).
NER, B.A. H. JUANG, translated by Furui, 1995,
November, NTT Advanced Technology (hereinafter referred to as "Reference 1"), or "Stochastic Language Model" (Kenji Kita, University of Tokyo Press) (hereinafter referred to as "Reference 2"). .

In these methods, the parameters constituting the probability model are estimated statistically from a large amount of data. That is, in constructing an acoustic model, speech data such as words and sentences from a large number of speakers are collected in advance, and the recognition accuracy and an index well related to the recognition accuracy are improved using a statistical method. An estimation is made.

For example, using the Baum-Welch algorithm, the parameters of the “hidden Markov model” that constitutes the acoustic model are estimated so that the likelihood of the learning data is increased. The method for estimating the acoustic model is described in detail in the lower volume of Document 1 described above.

Similarly, in constructing a language model, the probabilities of appearance of each utterance and words constituting the utterance are calculated from texts such as newspapers and transcripts of conversation in accordance with the structure of the language model.

As the structure of the language model, an “N-gram language model” or “probabilistic context-free grammar” that predicts the appearance probability of a subsequent word using an “n−1-fold Markov model” for the immediately preceding word, or , And combinations thereof are often applied.

In particular, the N-gram language model is widely used because it is effective and the parameter estimating means can be easily realized. Thus, in the following description, the construction of a language model will be described using an N-gram language model as an example.

For example, in an N-gram language model,
When N = 2 (called a bigram language model), P (W) in the above equation (3) is approximated as in the following equation (4).

[0018]

(Equation 4)

The conditional probability P (w _N | w ₁ ,..., W _N−1 ) serving as a parameter of the N-gram language model is calculated based on the frequency C (w ₁ ,・
.., W _N ) are estimated as in the following equation (5).

[0020]

(Equation 5)

However, if the conditional occurrence probability of a word is simply estimated as in the above equation (5), if a word string that does not exist in the learning data is included, the sentence occurrence probability will be “0”.

In order to prevent such a state, a word string that does not appear in the learning text is non-zero (not "0").
A process of assigning probabilities (generally called "smoothing") is performed.

The most common smoothing methods include:
"Back-off smoothing" proposed by Katz. In the back-off smoothing, a probability is assigned to a word string that has not appeared in the learning data, except for a certain ratio (by executing the counting) according to the frequency from the probability estimated by the above equation (5).

As the conditional probability assigned to a word string that does not appear in the learning data, a value estimated by a more rough language model is used. In the method by Katz, as a model coarser than N-gram, N-
One gram is used. The details of this method are shown on page 67 of Reference 2 mentioned above.

In the case of Japanese, since the text is not written separately, the definition of the word is ambiguous. Therefore, in the text, words obtained by dividing the text into consistent parts by some means are defined as words.

That is, a word is a linguistic unit such as a character, a morpheme, or a phrase, a text division based on an entropy criterion, or a combination thereof. This includes cases where language information such as language information is added.

In constructing a language model using the above-described statistical method, a large amount of voice data and text data are required to estimate the parameters of the language model. In particular, since the N-gram language model strongly depends on learning data, a target task (hereinafter, “target task”)
) Requires a large amount of data collection.

However, it is difficult to collect a large amount of text data for each task, and it is desirable to be able to construct a language model from a small amount of text data relating to a target task. Done.

The class language model collects similar words and is treated as the same class (group). The class language model reduces the number of estimated parameters of the language model, and provides an appropriate probability for words that do not exist in the learning data. Assign.

The definition of the relationship between a word and a class is manually determined according to the word or task, or determined based on data, and can be applied to an N-gram language model.

For example, the appearance probability of a sentence in the bigram class language model is represented by (1) the transition probability P between classes
And _{| (c i c i-1} ), (2) the probability a particular word from the class is selected P | as the product of the (w _i c _i), it is defined by the following equation (6) .

[0032]

(Equation 6)

For example, consider the case where 1000 words are divided into 100 classes each consisting of 10 words. At this time, the number of estimated parameters in the case of the word bigram language model is 1000 ² (= 1,000,000).

On the other hand, the number of estimated parameters in the case of the class bigram language model is expressed as the sum of (1) transition between classes and (2) mapping between classes and words, and is 100 ² + 100 × It is reduced to 10 (= 11000).

The correspondence between words and classes may be determined manually or may be obtained by executing word clustering from language data. FIG. 20 is an explanatory diagram showing an example of the class definition. In FIG. 20, a word w, a class c to which the word w belongs, and a probability P (w | c) output from the class c to which the word w belongs are described.

In the class N-gram language model, the estimation of the inter-class transition model is the same as in the case of a normal word N-gram. The method of constructing the class N-gram language model is described in detail in the above-mentioned document 2 starting from page 72.

On the other hand, the task adaptation is to make use of text data other than the target task together to compensate for the lack of learning data. Here, text data including tasks other than the target task will be referred to as general task language data.

Regarding task adaptation, "N-gram
Of vocabulary setting method for task adaptation "(Akinori Ito, Masaki Yoshida, IEICE Technical Report, No. 51)
-58, SP97-25, 1997) (hereinafter referred to as Reference 3)
) Are proposed.

In this method, task adaptation is performed by weighting and adding learning data of a target task and a general task to an N-gram language model.

FIG. 21 is a block diagram schematically showing an apparatus to which the language model construction method for speech recognition described in the above-mentioned reference 3 is applied. In FIG. 21, reference numeral 100 denotes a language model estimating unit that generates a task-adapted language model.

Reference numeral 101 denotes target task language data which accumulates text data of the target task and divides a text representing a sentence to be recognized by the target task into words. 102
Is general task language data, which accumulates text data of general tasks including tasks other than the target task, and divides text representing sentences included in the general tasks into words.

The language model estimating means 100 reads the target task language data 101 and the general task language data 102, performs appropriate weighting processing on each of them, counts up the frequency of word strings, and uses a statistical method to calculate the parameters of the language model. Is estimated.

The weighting process is given for each input. For example, assuming that the word string “I, wa” appears twice in the target task and four times in the general task, if the frequency weight of the target task is “3” and the frequency weight of the general task is “1”, the word The frequency of the column "I, ha" is "10
(= 3 × 2 + 1 × 4) ”.

Note that the weighting coefficients need not be integers. In counting, if necessary, words with low frequency can be removed, and the removed probability can be redistributed with equal probability to the words required for recognition.

From the frequency information “10” thus obtained,
For example, by the back-off smoothing method of Katz,
Estimate probabilities for known and unknown word strings. The frequency weight can be determined using, for example, a deletion estimation method so as to increase the appearance probability of the finally obtained language model with respect to the test data. The deletion estimation method is described on page 49 of the above reference 2.

Next, a learning procedure of a language model by task adaptation based on the conventional apparatus and the conventional method shown in FIG. 21 will be described with reference to the flowchart of FIG. First, the language model estimation unit 100 reads a weight parameter for an input from a weight parameter storage unit (not shown) (step S2201).

Next, the learning text divided into words is read from the target task language data 101 and the general task language data 102, and the frequency of a word string of n words or less weighted according to the weight parameter is obtained (step S2202). .

Finally, smoothing using, for example, Katz's back-off smoothing method is performed to estimate the parameters of the language model (step S2203), and the processing routine of FIG. 22 ends.

The above-described method is applied to the general task language data 102.
By using the text data in combination, it is possible to more appropriately estimate the appearance probabilities of word strings representing various expressions that are difficult to obtain from a small amount of learning data on the target task.

At the same time, the target task language data 10
By weighting 1, a greater probability can be given to a word string appearing in the corpus of the target task, and the recognition accuracy can be improved.

However, in the language model task adaptation method, although it is possible to satisfactorily estimate the appearance probability of a unique word in a target task or a word string appearing in a general task,
Since the combination of words specific to the target task and words that appeared in the general task is not considered, the accuracy of language model parameter estimation deteriorates around words specific to the target task when the text data of the target task is small. There's a problem.

For example, consider a case where the target task is a hotel reservation task and text data uttered by a similar reservation task other than a hotel is used as general task language data 102.

In this case, a word string that appears in general reservation tasks such as “Please, that, please” and a word specific to the target task such as “hotel” are obtained from the general task language data 102 and the target task language data 101, respectively. The appearance probability is estimated according to the frequency.

However, since the number of types of word combinations is very large, if the text data of the target task is small, the word string including the word specific to the target task, such as “Hotel, please, please”, is not enough text. Often not covered by data.

As a result, an inappropriate appearance probability is assigned to the word string, and the recognition accuracy may be reduced. In particular, words specific to the target task are often important in performing the task, and a decrease in recognition accuracy around these words is likely to greatly affect the performance of the entire system.

[0056]

As described above, the conventional language model learning apparatus, the speech recognition apparatus using the same, and the language model learning method and the speech recognition method using the same are as described above. Since the combination with the words that appeared in the task is not considered, if the text data of the target task is small, the accuracy of the language model parameter estimation around the words specific to the target task will deteriorate, and the performance of the entire system will be reduced. There is a problem of having an adverse effect.

The present invention has been made in order to solve the above-described problems. A similar word is obtained from a word specific to a target task and data of a general task, and a word string including a word specific to the task is obtained. A language model learning device and a speech recognition device using the same, and a language model learning method and a speech recognition method using the same, and a storage medium storing the methods are provided by using the language model learning device with an improved recognition accuracy by using the probability of appearance. The purpose is to gain.

[0058]

According to a first aspect of the present invention, there is provided a language model learning apparatus comprising: target task language data in which text data of a target task is accumulated; and text data of a general task including tasks other than the target task. Similar word pair extraction means and similar words for reading text data for language model learning from the accumulated general task language data, target task language data and general task language data, respectively, and constructing a task-adapted language model A similar word pair extracting unit that reads each text data from the target task language data and the general task language data, and includes a word included in the target task text data and a general task text data. Extracts similar word pairs from combinations with words contained in The word string synthesizing unit reads each text data, reads a similar word pair from the similar word pair extracting unit, synthesizes and outputs a word string including a word in the target task not included in the language data, and generates a language model. The means reads the respective text data, reads the word string from the similar word string synthesizing means, and weights each text data to obtain a statistic of the word string, thereby generating a task-adapted language model. .

A language model learning apparatus according to a second aspect of the present invention provides a general task language data in which text data of a target task is collected and a general task text data of a general task including tasks other than the target task. A target task word classifier, a general task word classifier, and a language model generator for constructing a task-adapted language model from language data and target task language data and general task language data; The word classifying means reads the text data of the target task from the target task language data, replaces the word with the class indicated in the class definition, and outputs classified first text data for language model learning, The general task word classifier converts general task language data into general task texts. The data is read, the word is replaced with the class indicated in the class definition, and the classified second text data for language model learning is output. The language model generating means converts the first and second text data into The language model is generated by reading and calculating the statistic of the word string by weighting each text data.

A language model learning apparatus according to a third aspect of the present invention provides a general task in which target task language data in which text data of a target task is accumulated and text data of general tasks including tasks other than the target task are accumulated. Target task word classifying means, general task word classifying means, similar word pair extracting means, similar word string synthesis for constructing a task-adapted language model from language data and target task language data and general task language data Means and a language model generating means, the target task word classifying means reads text data of the target task from the target task language data, replaces the word with the class indicated in the class definition, and sets a class for language model learning. And outputs the converted first text data. The general task word classifying means outputs The general task text data is read from the task language data, the word is replaced with the class indicated in the class definition, the second text data classified for language model learning is output, and the similar word pair extraction unit includes: The first and second text data are read, and a similar word pair is extracted from a combination of a word included in the text data of the target task and a word included in the text data of the general task. And the second text data, read a similar word pair from the similar word pair extracting means, synthesize and output a word string including a word in the target task not included in the language data, and The first and second text data are read, and a word string is read from the similar word string combining means. By determining the statistics of word sequences Te weighting, and it generates a task adapted already language model.

A language model learning apparatus according to a fourth aspect of the present invention provides a general task in which target task language data in which text data of a target task is accumulated and text data of general tasks including tasks other than the target task are accumulated. To construct a task-adapted statistical language model from language data, an initial language model created using text data prepared in advance, target task language data, general task language data, and initial language model, A similar word pair extracting unit and a similar word probability correcting unit, wherein the similar word pair extracting unit reads text data for language model learning from the target task language data and the general task language data, and converts the text data into text data of the target task. Combination of words included and words included in general task text data The similar word probability correction unit reads the similar word pair from the similar word pair extraction unit, reads the initial language model, and performs smoothing of the appearance probability of the word appearing in the target task. , To generate a task-adapted statistical language model.

A language model learning apparatus according to a fifth aspect of the present invention provides a general task in which target task language data in which text data of a target task is accumulated and text data of general tasks including tasks other than the target task are accumulated. Target task word classifying means for constructing a task-adapted class language model from language data, a pre-created initial class language model, target task language data, general task language data, and an initial class language model; General task word classifying means, similar word pair extracting means and similar word probability correcting means, the target task word classifying means reads the text data of the target task from the target task language data, To the first classified language model for language model learning The general task word classifying means reads the general task text data from the general task language data, replaces the word with the class indicated in the class definition, and outputs the class data for the language model learning. 2, the similar word pair extraction means reads the first and second text data, and performs similarity analysis based on a combination of a word included in the text data of the target task and a word included in the text data of the general task. A word pair is extracted, and the similar word probability correction unit reads the similar word pair from the similar word pair extraction unit,
A task-adapted class language model is generated by reading an initial class language model and smoothing the appearance probability of words appearing in a target task.

According to a sixth aspect of the present invention, there is provided a language model learning apparatus according to the first or the fourth aspect, wherein the similar word extracting means comprises: a language model generating means for calculating distance; Including a threshold determination means,
The language model generating means for distance calculation reads text data for language model learning from the target task language data and the general task language data, obtains a weight of each text data, obtains a statistic of the word string, and calculates the distance. Statistical language model, and the statistical inter-word distance calculating means reads the statistical language model from the distance calculating language model generating means, and calculates a statistical language model for a word pair consisting of words extracted from each text data. The above statistical distance is obtained as an inter-word distance, and the threshold value judging means reads the word pair and the inter-word distance from the statistical inter-word distance calculating means and outputs a word pair exceeding a predetermined threshold. It is.

According to a seventh aspect of the present invention, in the language model learning apparatus according to the first or fourth aspect, the similar word extracting means includes a language model for calculating a distance, a statistical inter-word distance calculating means, and a threshold. Including the value determination means, the language model for distance calculation is created using text data prepared in advance, and the statistical inter-word distance calculation means reads the language model for distance calculation and extracts it from each text data. For a word pair composed of words, a statistical distance on a distance calculation language model is determined as an inter-word distance, and the threshold value determination unit reads the word pair and the inter-word distance from the statistical inter-word distance calculation unit, and determines a predetermined distance. Is output.

According to an eighth aspect of the present invention, in the language model learning apparatus according to the third or fifth aspect, the similar word extracting means includes: a language model generating means for calculating a distance; Including a threshold determination means,
The language model generating means for distance calculation reads the first and second text data from the target task word classifying means and the general task word classifying means, obtains a weight of each text data and obtains a statistic of the word string, A statistical language model for distance calculation is generated, and the statistical inter-word distance calculation means reads the statistical language model from the distance calculation language model generation means, and performs statistical processing on a word pair consisting of words extracted from each text data. The statistical distance on the statistical language model is determined as the inter-word distance, and the threshold determining unit reads the word pair and the inter-word distance from the statistical inter-word distance calculating unit, and determines a word pair exceeding a predetermined threshold. Output.

According to a ninth aspect of the present invention, in the language model learning apparatus according to the third or fifth aspect, the similar word extracting means includes a class language model for distance calculation, a statistical interword distance calculating means, The class language model for distance calculation includes threshold value determination means, and the class language model for distance calculation is created using text data prepared in advance. The first and second text data are read from the word classifying means and the general task word classifying means, and the statistical distance on the distance calculating class language model is calculated for the word pair consisting of the words extracted from each text data. The distance is determined as the distance,
It reads word pairs and inter-word distances from the statistical inter-word distance calculation means and outputs word pairs exceeding a predetermined threshold.

In the language model learning apparatus according to a tenth aspect of the present invention, in any one of the sixth to ninth aspects, the statistical inter-word distance calculating means may calculate the Euclidean distance on the N-gram language model. To measure the inter-word distance.

Further, in the language model learning apparatus according to claim 11 of the present invention, in any one of claims 6 to 9, the statistical inter-word distance calculating means calculates the cross entropy on the N-gram language model. To measure the inter-word distance.

A speech recognition apparatus according to a twelfth aspect of the present invention is a speech recognition apparatus using the language model learning apparatus according to any one of the first to eleventh aspects, wherein the language model or the class language model is used. Is used for speech recognition.

A language model learning method according to a thirteenth aspect of the present invention includes a first step of collecting text data of a target task and a step of collecting text data of a general task including tasks other than the target task. Step 2, similar word pair extraction step, similar word string synthesizing step for reading language model learning text data from the target task language data and the general task language data and constructing a task-adapted language model And a language model generation step. The similar word pair extraction step reads each text data from the target task language data and the general task language data, and includes words included in the text data of the target task and text data of the general task. A similar word pair is extracted from a combination with a word, The column synthesizing step reads each text data and a similar word pair, synthesizes and outputs a word string including a word in the target task not included in the language data, and the language model generating step converts each text data and the word string into A task-adapted language model is generated by reading and calculating the statistic of a word string by weighting each text data.

The language model learning method according to claim 14 of the present invention includes a first step of accumulating text data of the target task and a step of accumulating text data of general tasks including tasks other than the target task. And a target task word classifying step, a general task word classifying step, and a language model generating step for constructing a task-adapted language model from the target task language data and the general task language data. The task word classifying step reads the text data of the target task from the target task language data, replaces the word with the class indicated in the class definition, and outputs classified first text data for language model learning. The general task word classifying step converts general task language data into general tasks. It reads the text data of the click,
The word is replaced with the class indicated in the class definition, and the classified second text data for language model learning is output. The language model generating step reads the first and second text data and reads each text. A language model is generated by weighting each data to obtain a statistic of a word string.

In the language model learning method according to the present invention, a first step of accumulating text data of a target task and a first step of accumulating text data of a general task including tasks other than the target task. Step 2, a target task word classifying step, a general task word classifying step, a similar word pair extraction step, and a similar word string for constructing a task-adapted language model from the target task language data and the general task language data. The target task word classifying step reads the text data of the target task from the target task language data, replaces the word with the class indicated in the class definition, and performs language model learning. Outputting the first text data in a class,
The general task word classifying step reads the general task text data from the general task language data, replaces the word with the class indicated in the class definition, and outputs the classified second text data for language model learning. The similar word pair extracting step reads the first and second text data and extracts a similar word pair from a combination of a word included in the text data of the target task and a word included in the text data of the general task, The similar word string synthesizing step reads the first and second text data and the similar word pair, synthesizes and outputs a word string including a word in the target task that is not included in the language data, and includes a language model generating step. , The first and second text data and the word string are read and weighted for each text data, and the statistics of the word string are read. By obtaining, and generates a task adapted already language model.

Further, in the language model learning method according to claim 16 of the present invention, the first step of accumulating text data of the target task and the step of accumulating text data of a general task including tasks other than the target task. Step 2; a third step of creating an initial language model using text data prepared in advance; and target task language data, general task language data, and the initial language model.
To build a task-adapted statistical language model,
A similar word pair extraction step and a similar word probability correction step, wherein the similar word pair extraction step reads text data for language model learning from the target task language data and the general task language data, respectively, and converts the text data into text data of the target task. A similar word pair is extracted from the combination of the included word and the word included in the text data of the general task. The similar word probability correction step reads the similar word pair and the initial language model, and reads the appearance probability of the word appearing in the target task. Is performed to generate a task-adapted statistical language model.

The language model learning method according to claim 17 of the present invention further comprises a first step of collecting text data of the target task and a step of collecting text data of a general task including tasks other than the target task. Step 2, a fourth step of creating an initial class language model in advance, and a target task for constructing a task-adapted class language model from the target task language data, general task language data, and the initial class language model Word classifying step, general task word classifying step,
A similar word pair extraction step and a similar word probability correction step, the target task word classifying step reads text data of the target task from the target task language data, replaces the word with the class indicated in the class definition, and sets the language The first text data for classifying for model learning is output, and the general task word classifying step includes:
The general task text data is read from the general task language data, the word is replaced with the class indicated in the class definition, the second text data classified for language model learning is output, and the similar word pair extraction step includes , First
And the second text data is read, and a similar word pair is extracted from a combination of a word included in the text data of the target task and a word included in the text data of the general task. By reading the initial class language model and smoothing the appearance probability of words appearing in the target task,
A task-adapted class language model is generated.

According to a eighteenth aspect of the present invention, there is provided the language model learning method according to the thirteenth or sixteenth aspect, wherein the similar word extracting step includes the step of generating a language model for distance calculation and the step of calculating statistical inter-word distance. And a threshold determination step.The language model generation step for distance calculation reads text data for language model learning from the target task language data and the general task language data, and weights each text data to form a word string. A statistical language model for distance calculation is generated by calculating a statistic, and a statistical word-to-word distance calculating step reads the statistical language model and generates a statistical language model for a word pair composed of words extracted from each text data. The statistical distance on the model is determined as the word-to-word distance, and the threshold value determination step includes Reading between distance, and outputs a word pair exceeding a predetermined threshold.

According to a nineteenth aspect of the present invention, in the method for learning a language model according to the thirteenth or sixteenth aspect, the similar word extracting step includes the steps of: creating a language model for calculating a distance; and calculating a statistical inter-word distance. And a threshold value determination step, wherein the distance calculation language model creation step creates a distance calculation language model using text data prepared in advance, and the statistical inter-word distance calculation step includes the distance calculation language model. Is read, and for a word pair consisting of words extracted from each text data, a statistical distance on a distance calculation language model is determined as an inter-word distance.The threshold value determination step reads the word pair and the inter-word distance, A word pair exceeding a predetermined threshold is output.

According to a twentieth aspect of the present invention, there is provided the language model learning method according to the fifteenth or seventeenth aspect, wherein the similar word extracting step includes the step of generating a language model for distance calculation, the step of calculating statistical inter-word distance. And a threshold value determination step. The distance calculation language model generation step reads the first and second text data, obtains a statistic of a word string by weighting each text data, and obtains a distance calculation statistic. Generating a statistical language model, and calculating a statistical inter-word distance, the statistical language model is read, and for a word pair consisting of words extracted from each text data, the statistical distance on the statistical language model is calculated as the inter-word distance. And the threshold determination step reads the word pair and the inter-word distance and outputs a word pair exceeding a predetermined threshold. A.

According to a twenty-first aspect of the present invention, there is provided a method for learning a language model according to the fifteenth or seventeenth aspect, wherein the similar word extracting step includes the step of creating a class language model for distance calculation, A step of creating a class language model for distance calculation includes creating a class language model for distance calculation using text data prepared in advance, and the step of calculating a statistical inter-word distance includes a step of calculating a distance between words. Class language model and the first and second text data are read, and for a word pair consisting of words extracted from each text data, a statistical distance on the distance calculation class language model is obtained as an inter-word distance. The value determination step is
A word pair and an inter-word distance are read, and a word pair exceeding a predetermined threshold is output.

Further, in the method for learning a language model according to claim 22 of the present invention, in any one of claims 18 to 21, the step of calculating the statistical inter-word distance includes the step of:
It measures the distance between words using the Euclidean distance on the Gram language model.

According to a twenty-third aspect of the present invention, in the method for learning a language model according to any one of the eighteenth to twenty-first aspects, the step of calculating the statistical inter-word distance comprises the step of:
It measures the distance between words using cross entropy on a Gram language model.

A speech recognition method according to a twenty-fourth aspect of the present invention is a speech recognition method using any one of the language model learning methods according to the thirteenth to twenty-third aspects. Is used for speech recognition.

A recording medium according to a twenty-fifth aspect of the present invention stores the language model learning method according to any one of the thirteenth to twenty-third aspects.

A recording medium according to a twenty-sixth aspect of the present invention stores a speech recognition method using the language model learning method of the twenty-fourth aspect.

[0084]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiment 1 Hereinafter, Embodiment 1 of the present invention will be described in detail with reference to the drawings. Here, an N-gram language model will be described as an example, but it goes without saying that the present invention is applicable to any statistical language model.

FIG. 1 is a block diagram schematically showing a language model learning apparatus according to Embodiment 1 of the present invention.
3 shows a configuration example of a language model learning device for speech recognition. In FIG. 1, reference numeral 101 denotes target task language data divided into words in a target task, and reference numeral 102 denotes general task language data divided into words in a general task, which are the same as those described above (see FIG. 21). .

Reference numeral 103 denotes similar word pair extracting means, 104 denotes similar word string synthesizing means, and 105 denotes language model generating means. These means 103 to 105 are associated with the target task language data 101 and the general task language data 102. To generate a task-adapted language model.

The language model generating means 105 corresponds to the language model estimating means 100 and generates a task-adapted language model. The similar word pair extracting means 103 and the similar word string synthesizing means 104 constitute a characteristic part of the present invention, unlike the above-described conventional apparatus.

That is, by means 103 and 104, similar general task words are obtained for words specific to the target task, and a word string in which the general task words in the learning text are replaced with similar target task words is synthesized. And by adding it to the language model learning text,
In constructing a language model, even if the text data of the target task is small, the recognition accuracy can be improved.

Hereinafter, the function of each of the units 103 to 105 in FIG. 1 will be specifically described with reference to various models and various data. However, the same functional blocks and models as described above are denoted by the same reference numerals, and detailed descriptions thereof will be omitted.

First, the similar word pair extraction means 103 defines in advance any combination (wT _, wG) of the word wT included in the target task language data 101 and the word wG included in the general task language data 102. The distance between words is calculated based on the distance scale.

At this time, the similar word pair extraction means 103
The calculated value of the inter-word distance is a predetermined threshold value t
h, the similar word pair (wT _, wG)
Is output to the similar word string synthesizing unit 104.

The distance d (wT, wG) between words is determined, for example, by forming a semantic classification corresponding to each word into a tree structure in advance according to the breadth of the concept, and calculating the number of arcs between the meaning nodes corresponding to each word. It is obtained by using it as a distance.

Next, the similar word string synthesizing means 104 generates the target task language data 101 and the general task language data 1
02 is extracted separately, and a similar word pair (wT, wG) read from the similar word pair extracting means 103 is referred to, and each of the word strings of the target task is included in the general task. It is determined whether or not the word wG is included.

As a result, when there is a word string “... WG...” Including the word wG in the general task, subsequently, the word wG in the general task is replaced with the word wT in the target task.
Is determined in the data of the general task or the target task.

As a result, if the word string “... WT...” Does not exist in the data of the general task or the target task, the similar word string synthesizing means 104 outputs the word w of the general task.
A word string "..." in which G is replaced by the word wT of the target task
.. wT... are output to the language model generating means 105.

Finally, the language model generating means 105 generates the target task language data 101, the general task language data 10
2 and the similar word string synthesizing means 104, respectively, reading the text data, assigning appropriate weights to the input frequencies to determine the frequency of the word string, and estimating the parameters of the language model using a statistical method. , Generate a task-adapted language model.

Next, the procedure of learning a language model by task adaptation according to the first embodiment of the present invention shown in FIG. 1 will be described more specifically with reference to the flowchart of FIG.

In FIG. 2, steps S201 to S20
3 is a process executed by the similar word pair extraction unit 103;
Steps S204 to S208 are similar word string synthesizing means 10
4, steps S209 to S211
Is a process executed by the language model generation means 105.

First, the similar word pair extraction means 103 extracts the target task language data 101 and the general task language data 1.
02, the learning text divided into words is read, and a word pair (wT, wG) is created (step S20).
1).

Further, a distance d (wT, wG) is calculated for a combination of the word wT included in the target task language data 101 and the word wG (different from the word wT) included in the general task language data 102 (step). S20
2).

Subsequently, the calculated distance d (wT, wG)
Is compared with a predetermined threshold th, and the distance d (wT, wG)
Is smaller than the threshold th (step S203).

The similar word pair extracting means 103 determines whether the
At 203, d (wT, wG) ≧ th (ie,
If No), the process returns to step S202 and the calculation of the distance d (wT, wG) is repeated, and d (wT, wG)
If determined as <th (that is, Yes), the word pair (wT, wG) at that time is output to the similar word string synthesizing unit 104.

The similar word string synthesizing means 104 reads text data delimited by words from the target task language data 101 and the general task language data 102, and extracts and stores word strings of all n words included in the data ( Step S204).

The word pair (wT, wT) selected from the read word string by the similar word pair extracting means 103
G), a word string “... WG...” Including the general task word wG is extracted (step S20).
5).

Subsequently, of the word strings extracted, a word string “... WT...” Obtained by replacing the general task word wG with the target task word wT is present in the already stored word strings. Is determined (step S206).

In step S206, the word string "..
··· wT... Are present in the already stored word string (that is, Yes), step S20
5, if it is determined that the word string "... wT ..." does not exist (that is, No), the word string "...
.. WT... Are output as text data (step S207).

Next, it is determined whether or not the processing for all the similar word pairs (wT, wG) has been completed (step S20).
8) If it is determined that the processing has not been completed (that is, No), the process returns to step S202, and the processing has been completed (that is, Ye).
If determined to be s), the process proceeds to step S209. Thereby, the processing steps S202 to S207 are executed for all the similar word pairs (wT, wG).

Here, as a specific example, the distance between the word “Yokohama Station” of the target task and the word “Narita Airport” of the general task is smaller than the threshold th, and the respective word strings “to Narita Airport” and “Narita Airport” Consider the case where "Narita Airport" exists in the general text data.

At this time, if the target text data includes the word string “Yokohama Station,” but the word string “from Yokohama Station” does not exist, the similar word string synthesizing means 104 The column "kara Yokohama Station" will be synthesized and output. As a result, a word string expected to appear in the target task is added to the learning text data using the word similarity information.

Next, in FIG. 2, the language model generation means 105 receives a weight parameter from the weight parameter storage means (not shown).
A weight parameter corresponding to each input is read (step S209).

The target task language data 101, general task language data 102, and similar word string synthesizing means 10
From 4, read the training text divided into words,
The frequency of the word string is obtained (step S210). At this time, in the case of the N-gram language model, it is necessary to calculate the frequency for a word string of n words or less.

Further, the language model generating means 105 generates a task-adapted language model by performing smoothing using, for example, Katz's back-off smoothing method and estimating the parameters of the language model (step S211). Then, the processing routine of FIG. 2 ends.

Since the word string including words characteristic of the target task is added to the learning data of the language model thus obtained, the prediction accuracy of the language model for the target task is improved.

Therefore, a high-accuracy language model for speech recognition can be estimated from a large amount of data (general task language data 102) including a task other than the target and a small amount of data relating to the target task (target task language data 101). it can. Then, a task-adapted language model is generated (step S211), and the processing routine of FIG. 2 ends.

The language model obtained as described above is applicable not only to speech recognition but also to character recognition requiring language processing and natural language text processing.

Further, the language model learning apparatus for speech recognition configured as shown in FIG. 1 can be recorded as a program on a recording medium.

That is, the similar word pair extracting means 1 in FIG.
03, a similar word string synthesizing function that performs the same processing as the similar word string synthesizing unit 104, and a language model generating function that performs the same processing as the language model generating unit 105. With the software configured, a language model learning program for speech recognition can be realized.

Embodiment 2 In the first embodiment, the text data from the target task language data 101 and the general task language data 102 are used as they are, but text data classified into classes may be used.

FIG. 3 is a block diagram schematically showing a language model learning device for a speech recognition device according to a second embodiment of the present invention. The same components as those described above (see FIG. 1) are denoted by the same reference numerals. Attaching or adding “A” after the reference numeral will omit detailed description.

In FIG. 3, reference numeral 301 denotes a target task word classifying means, which is inserted between the target task language data 101 and the language model generating means 105A. 30
Reference numeral 2 denotes a general task word classifying unit, which is inserted between the general task language data 102 and the language model generating unit 105A.

The characteristic function in this case is that a target task word classifying means 301 and a general task word classifying means 302 are provided to classify the words of the text corpus of the target task and the general task into classes. By reducing the number of estimated parameters, high-accuracy recognition is enabled even when the data of the target task is small in language model learning.

Hereinafter, the function of each of the means 301 and 302 in FIG. 3 will be specifically described with reference to various models and various data. The word class definition data (not shown) includes, for example, the word w, the class c to which the word w belongs, and the probability P () output from the class c to which the word w belongs, as described above (see FIG. 20). w | c) is described. The word class definition data as shown in FIG. 20 may be created manually or may be created from learning data by calculation.

The target task word classifying means 301 sequentially classifies, among the words of the input target task language data 101, the classes of which are defined according to the word class definition data, and outputs the words to the language model generating means 105A. I do.

The general task word classifying means 302 sequentially classifies the words of the input general task language data 102 whose classes are defined in accordance with the word class definition data, and outputs the class to the language model generating means 105A. I do.

Next, the procedure of learning a language model by task adaptation according to the second embodiment of the present invention shown in FIG. 3 will be described more specifically with reference to the flowchart of FIG.

In FIG. 4, steps S401 to S40
3 is a process executed by the target task word classifying means 301 and the general task word classifying means 302.

Steps S404 to S406 are processes executed by the language model generating means 105A, and correspond to steps S209 to S211 described above (see FIG. 2).

First, the target task word classifying means 301
The general task word classifier 302 reads word class definition data (not shown) (step S401).

The target task word classifying means 301
Reads the target task language data 101, generates a text in which the word is replaced with a class with respect to the word defined by the word class definition, and outputs the text (step S).
402).

Similarly, the general task word classifying means 30
2 reads the general task language data 102, generates a text in which the word is replaced with a class with respect to the word defined by the word class definition, and outputs the text (step S403).

Next, the language model generation means 105A first reads the weight parameter from the weight parameter storage means (not shown) (step S404), and then, converts the target task word classifying means 301 and the general task word classification. The learning text which is a word string including a class is read from the means 302, and the frequencies of the word and the word string are cumulatively calculated by multiplying the learning text by a weight parameter given to each of them (step S405).

Here, in the case of the class N-gram language model, the frequency is calculated for a class string of n words or less in the same manner as described above. Finally, the language model generation means 105A
The calculated frequency is smoothed, the parameters of the language model are estimated, a task-adapted class language model is generated (step S406), and the processing routine of FIG. 4 ends.

A class language model is obtained from the above-described processing procedure and word class definition data (not shown) defined in advance. As described above, a large amount of data including a task other than the target (general task language data 102) and a small amount of data on the target task (target task language data 10)
From 1), a highly accurate language model for speech recognition can be estimated.

The language model obtained in this way can be used not only for speech recognition, but also for character recognition that requires language processing,
It is also applicable to natural language text processing.

Further, the language model learning apparatus for speech recognition shown in FIG. 3 can be recorded as a program on a recording medium.

That is, a target word classifying function for performing the same processing as the target task word classifying means 301 in FIG. 3, a general word classifying function for performing the same processing as the general task word classifying means 302, and a language Model generation means 10
A software configured with a language model generating function for performing the same processing as that of 5A can realize a language model learning program for speech recognition.

Embodiment 3 In the second embodiment, only the language model generating means 105A is used.
Similar word pair extracting means and similar word string synthesizing means similar to those in the first embodiment may be used together.

FIG. 5 is a block diagram schematically showing a language model learning device for a speech recognition device according to Embodiment 3 of the present invention. The same components as those described above (see FIGS. 1 and 3) are used. The same reference numeral is given, or “B” is appended after the reference numeral, and the detailed description is omitted.

The characteristic function in this case is to provide a target task word classifying means 301 and a general task word classifying means 302 according to a single class definition, and to classify the words to reduce the number of parameters of the language model. In addition, by providing the similar word pair extracting means 103B and the similar word string synthesizing means 104B, high-precision recognition is possible even when the data of the target task is small when constructing the language model.

Next, the procedure of learning a language model by task adaptation according to the third embodiment of the present invention shown in FIG. 5 will be described more specifically with reference to the flowchart of FIG.

In FIG. 6, steps S601 to S60
3 corresponds to steps S401 to S403 described above (see FIG. 4).
, Respectively, and steps S604 to S614
Correspond to steps S201 to S211 described above (see FIG. 2).

First, the target task word classifying means 301
The general task word classifying means 302 reads word class definition data (not shown) (step S601).

The target task word classifying means 301 reads the target task language data 101, and generates and outputs text in which words are replaced with classes with respect to the words defined by the word class definition (step S602).

The general task word classifying means 302
Reads the general task language data 102, generates and outputs a text in which words are replaced with classes with respect to the words defined in the word class definition (step S60).
3).

The similar word pair extracting unit 103B outputs the class cT included in the target task language data and the class c included in the general task language data from the target task word classifying unit 301 and the general task word classifying unit 302.
A list of word class pairs (cT, cG) composed of a combination with G (different from class cT) is created and stored (step S604).

The similar word pair extracting means 103B calculates the distance d between the word class pair for the class cT included in the target task language data and the class cG (different from the class cT) included in the general task language data. (CT, c
G) is determined (step S605), and it is determined whether the threshold value is smaller than a predetermined threshold thc (step S606).

In step S606, d (cT, c
G) ≧ thc (ie, No), the process returns to step S605, and if it is determined that d (cT, cG) <thc (ie, Yes), the class pair (c
T, cG) as a similar word pair
4B (step S606).

The similar word string synthesizing means 104B reads the learning text data divided into classes from the target task word classifying means 301 and the general task word classifying means 302, and converts them into a class string having a length of n or less. The data is stored separately (step S607).

The word classifying means 301 and 3
Class pair (cT, cG) selected by the similar word pair extracting means 103B based on the class sequence read from "02".
Of the general task class cG are extracted (step S60).
8).

Further, the similar word string synthesizing means 104B
By referring to the class sequence read and stored from each of the word classifying means 301 and 302, the class sequence "..." in which the class cG of the general task is replaced by the class cT of the target task.
It is determined whether or not “cT...” exists in the target task language data 101 or the general task language data 102 (step S609).

If it is determined in step S609 that the class string “... CT...” Exists in each language data 101 or 102 (ie, Yes), the process returns to step S608, and the class string does not exist ( That is,
If it is determined as No), the class string “... CT.
Are combined and output as learning text data (step S610).

Next, it is determined whether or not the processing has been completed for all the similar class pairs (step S611). If it is determined that the processing has not been completed (that is, No), the process proceeds to step S611.
Returning to 605, if it is determined that the processing has been completed (that is, Yes), the processing proceeds to the processing steps (S612 to S614) by the language model generating unit 105B. Thus, the above processing is repeatedly executed for all the similar word class pairs (cT, cG).

The language model generation means 105B first reads the weight parameters from the weight parameter storage means (not shown) (step S612), and subsequently, the target task language data 101, the general task language data 102, and the similar word string synthesis From the means 104B, a learning text which is divided into words by weighting the frequency with the weight parameter is read (step S613).

Further, by performing frequency smoothing, the parameters of the language model are estimated (step S61).
4), the processing routine of FIG. 6 ends. A task-adapted class language model can be obtained by the above-described processing procedure and word class definition data (not shown) defined in advance.

As described above, a high-precision language model for speech recognition can be learned from a large amount of data including a task other than a target task and a small amount of data relating to the target task.

The language model thus obtained can be applied not only to speech recognition but also to character recognition requiring language processing, text processing in a natural language, and the like.

The language model learning apparatus for speech recognition shown in FIG. 5 can be recorded on a recording medium as a program.

That is, the target word classifying function for performing the same processing as the target task word classifying means 301 in FIG. 5 and the general word classifying function for performing the same processing as the general task word classifying means 302 are similar. Word pair extraction means 10
3B, a similar word string synthesizing function that performs the same processing as the similar word string synthesizing unit 104B, and a language model generating function that performs the same processing as the language model generating unit 105B. With the software configured, a language model learning program for speech recognition can be realized.

Embodiment 4 FIG. It should be noted that the first to the first embodiments
In 3, the language model generating means 105, 105A, or 105A is used to generate a task-adapted language model.
Although B was used, an initial language model created in advance and a similar word probability correction unit that executes smoothing of the word appearance probability may be used.

FIG. 7 is a block diagram schematically showing a language model learning device for a speech recognition device according to a fourth embodiment of the present invention. The same components as those described above (see FIG. 1) are designated by the same reference numerals. The detailed description is omitted.

In FIG. 7, reference numeral 701 denotes an initial language model,
Reference numeral 702 denotes a similar word probability correction unit. The similar word probability correction unit 702 generates a task-adapted statistical language model based on the similar word pair from the similar word pair extraction unit 103 and the prior language model from the initial language model 701.

The characteristic function in this case is that the similar word pair extracting unit 103 and the similar word probability correcting unit 702 are provided, and the characteristic of the similar word appearing in the text data of the general task is reflected on the word specific to the target task. Therefore, when constructing a statistical language model, high-precision recognition is possible even if the data of the target task is small.

Hereinafter, the function of each means in FIG. 7 will be specifically described with reference to various models and various data. The initial language model 701 is composed of a statistical language model whose parameters are estimated by a well-known conventional method or the method of the first embodiment.

The similar word probability correcting means 702 reads a similar word pair between the target task and the general task from the initial language model 701 and the similar word pair extracting means 103, and conditionally generates a word string including the word of the target task. Correct the probability. In the correction process of the word string appearance probability at this time, a conditional appearance probability of a word string including words of similar general tasks is used.

The probability assigned by the similar word probability correcting means 702 is obtained as the appearance probability of a word string that has not appeared in the learning text data, and is a part of the probability that has been excluded (discounted) from the conditional probability of the word string that has appeared. is there. That is, the conditional occurrence probability of a word string existing in the learning text data is stored while being equal to the initial language model 701.

Next, the procedure of learning a language model by task adaptation according to the fourth embodiment of the present invention shown in FIG. 7 will be described more specifically with reference to the flowchart of FIG.

In FIG. 8, steps S801 to S80
3 and S805 are performed in step S2 described above (see FIG. 2).
01 to S203 and S208, respectively. Steps S806 to S812 are processes executed by the similar word probability correction unit 702.

First, the similar word pair extracting means 103 extracts the target task language data 101 and the general task language data 1.
02, the text for learning divided into words is read (step S801), and the word wT included in the target task language data and the word wG included in the general task language data are read.
(Different from wT), the distance d (wT, wG)
(Step S802).

Subsequently, it is determined whether or not the distance d (wT, wG) between words is smaller than a threshold value th (step S).
803), d (wT, wG) ≧ th (that is, No)
Is determined, the process returns to step S802, and d (wT, w
G) If it is determined that <th (that is, Yes), the word pair (wT, wG) at that time is added to the similar word pair (step S804).

Thereafter, it is determined whether or not the above processing has been completed for all the word pairs (step S805). If it is determined that the processing has not been completed (ie, No), the process returns to step S802, and has been completed (ie, step S805). , Yes), the process proceeds to the next processing step S806. As a result, calculations for all word pairs are sequentially performed, and a list of created similar word pairs (wT, wG) is output to the similar word probability correction unit 702.

First, the similar word probability correcting means 702 reads the initial language model 701 (step S806).
Subsequently, for the similar word pair (wT, wG) read from the similar word pair extraction unit 103, the initial language model 70
1, the conditional probabilities PwG (w _n | w ₁ ,..., W) including the general task word wG among the conditional probabilities defined in
_n-1 ) is taken out (step S807).

Next, for each of the extracted conditional probabilities, a conditional probability PwT (w _n | w ₁ ,..., W, where the general task word wG is replaced by the target task word wT.
_n-1 ) is determined in the initial language model 701 (step S808).

If it is determined in step S808 that the conditional probability PwT (w _n | w ₁ ,..., W _n-1 ) is not defined in the initial language model 701 (ie, No), the unknown probability is determined. A part of the probabilities removed for the word string is assigned to correct the conditional probabilities (step S80).
9) The process proceeds to the next determination step S810.

On the other hand, if the conditional probability PwG has been defined, and it is determined in step S808 that the conditional probability PwT has been defined (ie, Yes),
The process immediately proceeds to the next determination step S810.

At this time, the probability corrected in step S809 is, for example, the same word history (w ₁ ,...)
, W _n-1 ) as the minimum value of the conditional probabilities.

Next, it is determined whether or not there is another conditional probability of a word string including the general word wG (step S8).
10) If it is determined that a word string including the general word wG exists (that is, Yes), the process returns to step S808.

On the other hand, if it is determined in step S810 that there is no other conditional probability including the general word wG (ie, No), all word pairs (wT, w
Regarding G), it is determined whether or not the execution of the above processing has been completed (step S811).

If it is determined in step S811 that the processing of all the word pairs has not been completed (ie, No), the process returns to step S807, and has been completed (ie, Ye).
If determined to be s), the process proceeds to the next processing step S812.

As a result, the above processing is executed for the word string including all the general words wG, and for the word pair (wT, wG) including all the general words wG. Finally, the sum of the probabilities removed from the language model for the unknown word string is normalized so that the sum of the probabilities of the language model becomes "1" (step S812), and the processing routine of FIG. 8 ends. .

If the conditional probability is not defined, the probability given by a simple language model is usually used. For example, an N-gram language model according to Katz's backoff refers to a low-order N-1 gram language model and is assigned a small probability,
Since the accuracy of the probability is low, if there is a word string including a similar word of the target task, a larger probability than the actual word is estimated.

Other conditional probabilities Pw containing the general word wG
G is processed in the same manner in step S810, and the processing in steps S806 to S810 is performed in step S811 on all similar word pairs (wG, wT).
Is executed for

As described above, the similar word probability correcting means 702
Is used, smoothing is performed on words that have similar properties between the general task and the target task using the appearance probability of words of the general task, and a more accurate model for speech recognition can be estimated. it can.

Note that the language model obtained in this way can be applied to character recognition and text processing that require language processing, as described above.

The language model learning apparatus for speech recognition shown in FIG. 7 can be recorded as a program on a recording medium. That is, the similar word pair extracting means 1 in FIG.
A language model learning program for speech recognition is realized by software composed of a similar word pair extraction function performing the same processing as that performed in step S03, and a similar word probability correction function performing the same processing as performed by the similar word probability correcting unit 702. be able to.

Embodiment 5 FIG. In the fourth embodiment, the text data from the target task language data 101 and the general task language data 102 are used as they are. However, as in the third embodiment (see FIG. 5), the text data is classified into classes. May be used.

FIG. 9 is a block diagram schematically showing a language model learning device for a speech recognition device according to Embodiment 5 of the present invention. The same components as those described above (see FIGS. 5 and 7) are used. The same reference numerals are given and detailed descriptions are omitted.

In FIG. 9, reference numeral 901 denotes an initial class language model, which is the initial language model 701 described above (see FIG. 7).
Is connected to the similar word probability correcting means 702.

The characteristic functions in this case are similar word pair extracting means 103B, target task word classifying means 301,
A general task word classifying unit 302 and a similar word probability correcting unit 702 are provided to reflect a characteristic of a similar class appearing in text data of a general task on a class specific to the target task. However, the object is to generate a class language model with further improved recognition accuracy from the initial class language model 901.

The function of each means in FIG. 9 will be specifically described below in relation to various models and various data. The initial class language model 901 is a statistical class language model whose parameters are estimated by a well-known conventional method or the method according to the second or third embodiment.

The probability assigned by the similar word probability correcting means 702 is a part of the probability that is excluded (discounted) from the conditional probability of a word class sequence that has appeared for a word class sequence that has not appeared in the learning text data. The conditional occurrence probabilities of the word classes included in the learning text data are stored.

For example, when the conditional probability P (c _n | c ₁ ,..., C _n-1 ) for the word class is changed, the probability is set so as to be larger than the original conditional probability of the word class sequence. Assigned.

Next, the procedure of learning a language model by task adaptation according to the fifth embodiment of the present invention shown in FIG. 9 will be described more specifically with reference to the flowchart of FIG.

In FIG. 10, steps S1001-S100
Steps 1003 to 1003 correspond to steps S601 to S100 described above (see FIG. 6).
603, and corresponds to steps S1004 to S1004.
Step S1015 corresponds to steps S801 to S801 described above (see FIG. 8).
This corresponds to S812.

First, the target task word classifying means 301
The general task word classifying means 302 reads word class definition data (not shown) (step S1001).

The target task word classifying means 301 reads the target task language data 101, and generates and outputs text in which words are replaced with classes with respect to the words defined by the word class definition (step S1002).

The general task word classifying means 302
Reads the general task language data 102, generates and outputs text in which words are replaced with classes with respect to the words defined by the word class definition (step S100).
3).

Next, the similar word pair extracting means 103B reads the class strings through the target task word classifying means 301 and the general task word classifying means 302 (step S1004).

The class cT included in the target task language data and the class cG included in the general task language data
(Different from cT), the distance d (cT, cG)
(Step S1005), and the distance d (c) between the classes
(T, cG) is smaller than the threshold value thc (step S1006).

In step S1006, d (cT,
If it is determined that cG) ≧ thc (that is, No), the process returns to step S1005. If it is determined that d (cT, cG) <thc (that is, Yes), the class pair (cT, cG) at that time is similar. Add to class pair (step S
1007).

Hereinafter, through determination step S1008,
The above processing is sequentially performed for all the class pairs, and a list of the created similar class pairs (cT, cG) is output to the similar word probability correction unit 702.

Next, the similar word probability correcting means 702 first reads the initial class language model 901 (step S1009), and then sequentially reads the similar class pairs (cT, cG) from the similar word pair extracting means 103B (step S1009). Step S1010).

Also, of the conditional probabilities defined in the initial class language model 901, the class cG of the general task
Conditional probability PcG (c _n | c ₁ ,...,
c _n-1 ), the conditional probability PcT of replacing the general task class cG with the target task class cT
It is determined whether (c _n | c ₁ ,..., C _n-1 ) is defined in the learning data (step S1011).

In step S1011, the conditional probability PcT (c _n | c ₁ ,..., C _{n- 1} ) is not defined in the initial class language model 901 (ie, No).
Is determined, a part of the probabilities removed for the unknown class sequence is assigned to correct the conditional probability (step S1012), and the process proceeds to the next determination step S1013.

On the other hand, the conditional probability PcG is defined, and in step S1011 the conditional probability PcT
Is determined (that is, Yes), the process immediately proceeds to the next determination step S1013.

At this time, the probability corrected in step S1012 is, for example, the same as the class history (c ₁ ,.
.., C _n-1 ) (step S1012).

Hereinafter, the same processing is performed for other conditional probabilities PcG including the class cG via step S1013. Also, through step S1014,
The processes in steps S1006 to S1010 are executed for all the similar class pairs (cG, cT).

Finally, the similar word probability correction means 702
The task-adapted class language model is generated by normalizing the back-off probability so that the sum of the probabilities of the class language model becomes 1 (step S1015), and the processing routine of FIG. 10 ends.

As described above, the similar word pair extracting means 103B and the similar word probability correcting means 702 are provided in addition to the word classifying means 301 and 302, and a word class having similar properties between the general task and the target task is provided.
By performing smoothing using the appearance probability of a word class of a general task, a class language model for speech recognition can be estimated with high accuracy.

The class language model obtained in this way can be applied to character recognition that requires language processing, text processing of natural language, and the like.

Further, the language model learning apparatus for speech recognition shown in FIG. 9 can be recorded as a program on a recording medium.

That is, the similar word pair extracting means 1 shown in FIG.
A similar word pair extraction function that performs the same processing as that of the target task word classifier 301, a target task word classifier that performs the same processing as the target task word classifier 301, and a general task that performs the same processing as the general task word classifier 302 A software comprising a word classifying function and a similar word probability correcting function for performing the same processing as the similar word probability correcting means 702 can implement a language model learning program for speech recognition.

Embodiment 6 Although the functional configuration of the similar word pair extracting means is not specifically described in the above-described Embodiment 1, it may be configured as shown in FIG. 11, for example.

FIG. 11 is a functional block diagram showing a specific configuration example of the similar word pair extracting means 103C used in the language model learning apparatus for speech recognition according to the sixth embodiment of the present invention. Are denoted by the same reference numerals or by adding a “C” after the reference numerals, and a detailed description thereof will be omitted.

In FIG. 11, reference numeral 1101 denotes a statistical inter-word distance calculating means; 1102, a threshold value determining means;
Is a language model generating means for calculating distance in the similar word pair extracting means 103C.

The characteristic function in this case is that the similarity word pair extraction means 103C includes a distance calculation language model generation means 11
05, a statistical inter-word distance calculating unit 1101 and a threshold determining unit 1102 are provided, and based on a statistical distance scale according to the language data, the inter-word distance d () between the word wT of the target task and the word wG of the general task. wT, wG) to select a word pair to determine a similar word pair with high accuracy.

Hereinafter, the function of each means in FIG.
A detailed description will be given in relation to various models and various data. In the similar word pair extraction unit 103C, the statistical inter-word distance calculation unit 1101 extracts the language model estimated from the distance calculation language model generation unit 1105, and is extracted from the target task language data 101 and the general task language data 102. For each different word pair, the inter-word distance based on the language model is obtained, and the word pair and the inter-word distance are output.

The threshold value judging means 1102 calculates the word pair and the statistical inter-word distance by the statistical inter-word distance calculating means 11.
01, and the word pair (wT, wG) is output when the inter-word distance is equal to or smaller than a certain threshold value.

At this time, the statistical inter-word distance calculating means 11
01 is, for example, using a Euclidean distance in a conditional probability of an N-gram language model as a calculation method of a statistical inter-word distance regarding the word wT in the target task and the word wG in the general task, as in the following equation (7). The statistical inter-word distance D ₁ (wT, wG) is obtained.

[0219]

(Equation 7)

However, in the expression (7), V is a population of the vocabulary x of the language data (word), and represents all the vocabulary included in the language model.

Also, statistical inter-word distance calculating means 1101
Uses the Euclidean distance using the conditional probability of the preceding word with respect to the succeeding word, and can obtain the statistical inter-word distance D ₂ (wT, wG) as in the following equation (8).

[0222]

(Equation 8)

Further, not only the above equations (7) and (8) can be used individually, but also the sum of the equations (7) and (8) can be used.

Also, statistical inter-word distance calculating means 1101
Is, for example, using the cross entropy related to the word wT, and as shown in the following equation (9), the statistical inter-word distance D ₃
(WT, wG) can be obtained.

[0225]

(Equation 9)

As in the case of using the Euclidean distance, the conditional probability of the preceding word with respect to the succeeding word can be used as shown in the following equation (10).

[0227]

(Equation 10)

In addition, not only the above equations (9) and (10) can be used individually, but also the sum of equations (9) and (10) can be used.

Furthermore, the above-mentioned statistical measure and linguistic information can be used in combination. For example, when the word represents a morpheme and the parts of speech of the two words are not the same, the distance can be set to infinity and excluded from the similar word candidates.

Next, the operation of the similar word pair extraction means 103C in the task adaptation shown in FIG. 11 according to the sixth embodiment of the present invention shown in FIG. 11 will be described more specifically with reference to the flowchart in FIG. In FIG.
Steps S1203 to S1207 are as described above (see FIG. 2).
Steps S201 to S203, S207 and S20
8 respectively.

First, the language model generating means for distance calculation 11
05 reads the target task language data 101 and the general task language data 102 (step S1201),
The language model parameters are estimated from the input text data (step S1202).

Also, statistical inter-word distance calculating means 1101
Creates a word pair (wT, wG) composed of an arbitrary combination of the word wT included in the target task and the word wG included in the general task (step S1203), and estimates by the language model generating means for distance calculation 1105 The statistical distance d (wT, wG) is calculated on the language model to be executed (step S1204).

Subsequently, the threshold value judging means 1102 calculates the distance d (wT, wG) of the word pair (wT, wG) obtained from the statistical inter-word distance calculating means 1101 as the threshold value th.
Then, it is determined whether or not the distance d (wT, wG) is smaller than the threshold th (step S1205).

In step S1205, d (wT,
If it is determined that wG) ≧ th (that is, No), the process returns to step S1204; if it is determined that d (wT, wG) <th (that is, Yes), the word pair (w
T, wG) as a similar word pair (step S1)
206).

Hereinafter, the above processing is performed for all the word pairs (wT, wG) through the end determination step S1207.

As described above, the similar word pair extracting means 103C
In, a highly accurate similar word pair can be determined by estimating a language model and using a distance scale based on statistics.

The language model obtained in this way can be applied to character recognition that requires language processing, natural language text processing, and the like. Further, the function of the similar word pair extraction means 103C in FIG. 11 can be recorded on a recording medium as a program.

That is, a language model generating function for performing the same processing as the distance calculating language model generating means 1105 in FIG. 11, and a statistical inter-word distance calculating function for performing the same processing as the statistical inter-word distance calculating means 1101 With a software including a threshold determination function for performing the same processing as the threshold determination unit 1102, a similar word pair extraction program of a language model learning apparatus for speech recognition can be realized.

In FIG. 11, the language model generating means for distance calculation 1105 is used, but as shown in FIG.
A language model for distance calculation 1301 may be used. FIG.
The language model 1301 for distance calculation in the similar word pair extraction means 103D is the same as the initial language model 701 described above (see FIG. 7), and is created in advance.

Here, the similar word pair extracting means 10
Although the input data to 3C is a word, a word class may be used instead of a word as shown in FIG. FIG.
In, the language model generating means for calculating distance 1105E and the statistical inter-word distance calculating means 1101E in the similar word pair extracting means 103E fetch word classes from the respective word classifying means 301 and 302. Also in this case, a class pair can be extracted as described above.

Further, in FIG. 14, a language model for distance calculation 1105E is used, but a class language model for distance calculation 1501 may be used as shown in FIG. In FIG. 15, similar word pair extraction means 103F
The class language model for distance calculation 1501 is similar to the above-described initial class language model 901 (see FIG. 9) and is created in advance.

Seventh Embodiment In the first to sixth embodiments, only the language model learning device is focused on and the speech recognition device is not specifically mentioned. For example, the speech recognition device is shown in FIG. May be configured.

FIG. 16 is a block diagram schematically showing a speech recognition apparatus using a language model according to a seventh embodiment of the present invention.
6 shows a case where a language model generated by the method described in Example 6 is used.

In FIG. 16, reference numeral 1601 denotes an acoustic feature extracting means; 1602, an acoustic model; 1603, an acoustic collating means; 1604, a word dictionary;
Reference numeral 06 denotes a language matching unit.

The language model 1605 is constructed using the language model learning apparatus and method described in the first, fourth and sixth embodiments. The characteristic function in this case is that the language model 1
A linguistic collation unit 1606 using 605 is provided to enable high-accuracy speech recognition even when the data of the target task is small.

In the following, the function of each means in FIG.
A detailed description will be given in relation to various models and various data. First, the acoustic feature extraction unit 1601 performs A / D conversion on an input speech waveform, extracts the speech waveform for each analysis time frame, and converts it into a vector of parameters that favorably represents speech features such as mel cepstrum.

The acoustic model 1602 expresses the properties of the acoustic feature vector in a speech recognition unit (phonemes, words, etc.) by using a HMM, for example, by means of a probability distribution, a state transition, or the like.

The acoustic matching means 1603 collates the acoustic feature vector of the phoneme obtained from the acoustic feature extracting means 1601 with the acoustic model 1602, and outputs a score representing the degree of collation.

The word dictionary 1604 contains the acoustic model 1602
And the correspondence between words and words that are linguistic units. The language model 1605 is obtained from the language model learning device, and describes connection information of a word to be recognized. For example, the transition between words is described by using a word N-gram language model by using (n-1) Markov models. Express in the process.

The language matching means 1606 is the sound matching means 1
A collation score between the acoustic feature amount and the acoustic model is received from 603, and a process is performed by referring to the word dictionary 1604 and the language model 1605 to determine a word string having the highest score among the word strings to be recognized as a recognition result.

Next, the procedure of speech recognition based on the seventh embodiment of the present invention shown in FIG. 16 will be described more specifically with reference to the flowchart of FIG. First, the speech recognition device shown in FIG. 16 includes an acoustic model 1602 and a word dictionary 1604 prepared in advance,
Embodiments 1, 4, and 6 (FIGS. 1, 2, 7, 8,
Language model 16 generated according to FIGS. 11 to 13)
05 is read (step S1701).

The acoustic feature extraction means 1601 performs A / D on the input speech to be recognized, reads a speech frame obtained by dividing a certain time interval (step S1702), and uses a signal processing method for the speech frame to be processed. An acoustic feature vector that favorably represents a speech feature such as a cepstrum is extracted (step S1703).

Subsequently, the sound matching means 1603 checks the sound feature vector obtained in step S1703 with the sound model 1602 to obtain a sound matching score (step S1704).

Next, the language matching means 1606 refers to the word dictionary 1604 and the language model 1605 and accumulates acoustic matching scores for the words to be recognized (step S1705).

The language matching means 1606 determines whether or not the last frame of the target voice has been reached while executing the above-described matching process for each frame (step S1706), and has not reached the last frame of the target voice. (Ie, N
If it is determined as o), the process returns to step S1702.

If it is determined in step S1706 that the last frame of the target voice has been reached (that is, Yes), it is considered that the matching has been completed, and the one having the best score at this time is recognized as the recognition result. Is output (step S1707), and the processing routine of FIG. 17 ends.

As described above, by using the language model 1605, a high-precision language model is constructed from a large amount of data including a task other than the target task and a small amount of data relating to the target task. Can be realized.

Eighth Embodiment In the seventh embodiment, the first, fourth and fourth embodiments are described.
Although the language model generated by the sixth embodiment is used, the class language model generated by the second, third, fifth, and sixth embodiments may be used.

FIG. 18 is a block diagram schematically showing a speech recognition apparatus using a language model according to the eighth embodiment of the present invention. The apparatus and method described in the second, third, fifth and sixth embodiments are described. 2 shows a case where a language model generated by is used.

In FIG. 18, each means 1601 to 160
Reference numeral 4 is the same as that described above (see FIG. 16), and the language matching means 1606A corresponds to the language matching means 1606 described above. Reference numeral 1801 denotes a class definition representing the correspondence between classes and words in the language model, and reference numeral 1802 denotes a class language model that gives the probability of occurrence of the class.

The class language model 1802 is the same as that of the second, third, fifth and sixth embodiments (FIGS. 3 to 6, FIG. 9, FIG. 10, FIG.
4, see FIG. 15).

The characteristic function in this case is that by providing a language matching means 1606A using a class language model 1802, high-precision speech recognition can be performed even when the data of the target task used for learning is small. It is made possible.

Next, the procedure of speech recognition based on the eighth embodiment of the present invention shown in FIG. 18 will be described more specifically with reference to the flowchart in FIG. In FIG. 19, steps S1901 to S1907 correspond to steps S1701 to S1707 described above (see FIG. 17), respectively.

First, the acoustic model 16 prepared in advance
02, along with the word dictionary 1604 and the class definition 1801, the class language model 1802 generated according to the second, third, fifth and sixth embodiments is read (step S19).
01).

The acoustic feature extraction means 1601 A / Ds the input speech to be recognized, reads a speech frame obtained by dividing a certain time section (step S1902), and uses the signal processing method for the subject speech frame to execute An acoustic feature vector that well represents a speech feature such as a cepstrum is extracted (step S1903).

Subsequently, the sound matching means 1603 checks the obtained sound feature vector with the sound model 1602,
An acoustic matching score is obtained (step S1904).

Next, the language matching means 1606A refers to the word dictionary 1604, the class definition 1801 and the class language model 1802, and
The acoustic matching score is accumulated (step S190)
5).

Thereafter, the above-described collation processing is executed for each frame through step S1906, and when the final frame of the target voice is reached and the collation is completed, the one having the highest score is determined as the recognition result. Is output (step S1907), and the processing routine in FIG. 19 ends.

As described above, by using the class language model 1802, high-accuracy speech recognition can be realized from a large amount of data including a task other than the target and a small amount of data relating to the target task.

[0270]

As described above, according to the first aspect of the present invention, general task text data of target tasks and text data of general tasks including tasks other than the target tasks are collected. Similar word pair extraction means, similar word string synthesis means for reading text data for language model learning from task language data, target task language data and general task language data, respectively, and constructing a task-adapted language model And a language model generating unit, wherein the similar word pair extracting unit reads each text data from the target task language data and the general task language data, and includes the words included in the target task text data and the general task text data. A similar word pair is extracted from the combination with the word, and the similar word string synthesizing unit, Along with reading the text data, a similar word pair is read from the similar word pair extraction unit, a word string including a word in the target task not included in the language data is synthesized and output, and the language model generation unit converts each text data. In addition to reading, a word string is read from the similar word string synthesizing means and weighted for each text data to obtain a statistic of the word string, thereby generating a task-adapted language model, thereby improving recognition accuracy. There is an effect that a language model learning device can be obtained.

According to claim 2 of the present invention, target task language data in which text data of a target task is collected, general task language data in which text data of general tasks including tasks other than the target task are collected, A target task word classifier, a general task word classifier, and a language model generator for constructing a task-adapted language model from the target task language data and the general task language data; Reads the text data of the target task from the target task language data, replaces the word with the class indicated in the class definition, and converts the first
The general task word classifying means reads the general task text data from the general task language data, replaces the word with the class indicated in the class definition, and classifies for general language model learning. Second
Is output, and the language model generating means reads the first and second text data and weights each text data to obtain a statistic of a word string, thereby generating a language model. Thus, there is an effect that a language model learning device with improved recognition accuracy can be obtained.

According to claim 3 of the present invention, target task language data in which text data of a target task is collected, general task language data in which text data of general tasks including tasks other than the target task are collected, Target task word classifying means, general task word classifying means, similar word pair extracting means, similar word string synthesizing means, and language model for constructing a task-adapted language model from target task language data and general task language data The target task word classifying means reads text data of the target task from the target task language data, replaces the word with the class indicated in the class definition, and classifies the text data into a class for the language model learning. 1. The general task word classifying means outputs general task language data. Read the text data of the general task, replace the words with the classes indicated in the class definition, and output the classified second text data for language model learning. Reading the second text data, extracting a similar word pair from a combination of a word included in the text data of the target task and a word included in the text data of the general task,
The similar word string synthesizing unit reads the first and second text data, reads a similar word pair from the similar word pair extracting unit, and synthesizes a word string including a word in the target task that is not included in the language data. The language model generating means reads the first and second text data, reads the word string from the similar word string synthesizing means, and weights each text data to obtain a statistic of the word string. Since the adapted language model is generated, there is an effect that a language model learning device with improved recognition accuracy can be obtained.

According to claim 4 of the present invention, target task language data in which text data of a target task is collected, general task language data in which text data of general tasks including tasks other than the target task are collected, Similar word pair extraction for constructing a task-adapted statistical language model from an initial language model created using text data prepared in advance and target task language data, general task language data, and initial language model Means and a similar word probability correction means, wherein the similar word pair extraction means reads text data for language model learning from the target task language data and the general task language data, respectively, and extracts words included in the text data of the target task. Similar word pairs are obtained from combinations with words contained in general task text data. The similar word probability correction means extracts the similar word pairs from the similar word pair extraction means, reads the initial language model, and smooths the appearance probability of the words appearing in the target task, thereby obtaining the task-adapted statistics. Since a dynamic language model is generated, there is an effect that a language model learning device with improved recognition accuracy can be obtained.

According to claim 5 of the present invention, target task language data in which text data of a target task is collected, general task language data in which text data of general tasks including tasks other than the target task are collected, Target task word classifying means and general task word class for constructing a task-adapted class language model from a pre-created initial class language model and target task language data, general task language data, and initial class language model Target task word classifier reads text data of a target task from target task language data, and replaces the word with a class indicated in the class definition. To output the first text data classified into classes for language model learning. Then, the general task word classifying means reads the general task text data from the general task language data, replaces the word with the class indicated in the class definition, and classifies the second text data for language model learning. And the similar word pair extraction means outputs the first
And the second text data are read, and a similar word pair is extracted from a combination of a word included in the text data of the target task and a word included in the text data of the general task. By reading similar word pairs from the means, reading the initial class language model, and smoothing the appearance probability of words appearing in the target task, a task-adapted class language model was generated, so recognition accuracy was improved. There is an effect that an enhanced language model learning device can be obtained.

According to a sixth aspect of the present invention, in the first or fourth aspect, the similar word extracting means includes a language model generating means for calculating a distance, a statistical inter-word distance calculating means, and a threshold value determining means. The language model generation means for distance calculation reads text data for language model learning from the target task language data and the general task language data, and obtains a statistic of the word string by weighting each text data. Generating a statistical language model for distance calculation, and the statistical inter-word distance calculating means reads the statistical language model from the distance calculating language model generating means, and for a word pair consisting of words extracted from each text data, The statistical distance on the statistical language model is determined as the inter-word distance. Load the fine word distance, since the output a word pair exceeding a predetermined threshold, the language model learning device with improved recognition accuracy is the effect obtained.

According to a seventh aspect of the present invention, in the first or fourth aspect, the similar word extracting means includes a language model for distance calculation, a statistical inter-word distance calculating means, and a threshold value determining means. Including, the language model for distance calculation is created using text data prepared in advance,
The statistical inter-word distance calculation means reads the distance calculation language model, obtains a statistical distance on the distance calculation language model as a word-to-word distance for a word pair composed of words extracted from each text data, The value judging unit reads the word pair and the inter-word distance from the statistical inter-word distance calculating unit and outputs a word pair exceeding a predetermined threshold, so that a language model learning apparatus with improved recognition accuracy is obtained. Has the effect.

According to an eighth aspect of the present invention, in the third or fifth aspect, the similar word extracting means includes a language model generating means for calculating distance, a statistical inter-word distance calculating means, and a threshold value judging means. The language model generating means for distance calculation reads the first and second text data from the target task word classifying means and the general task word classifying means, and weights each text data for statistical data of the word string. , And a statistical language model for distance calculation is generated. The statistical inter-word distance calculation means reads the statistical language model from the distance calculation language model generation means and generates a word composed of words extracted from each text data. For the pair, the statistical distance on the statistical language model is determined as the inter-word distance, and the threshold value judging means receives the word pair from the statistical inter-word distance calculating means. Load the fine word distance, since the output a word pair exceeding a predetermined threshold, the language model learning device with improved recognition accuracy is the effect obtained.

According to a ninth aspect of the present invention, in the third or fifth aspect, the similar word extracting means includes a class language model for calculating distance, a statistical inter-word distance calculating means, and a threshold determining means. The distance calculation class language model is created using text data prepared in advance, and the statistical inter-word distance calculation means reads the distance calculation class language model and sets the target task word classification means. And reading the first and second text data from the general task word classifying means, and for a word pair consisting of words extracted from each text data, obtains a statistical distance on a distance calculation class language model as an inter-word distance. The threshold determining means reads the word pair and the inter-word distance from the statistical inter-word distance calculating means and exceeds a predetermined threshold Since so as to output a word pair, the effect of the language model learning device with improved recognition accuracy can be obtained.

According to a tenth aspect of the present invention, in any one of the sixth to ninth aspects, the statistical inter-word distance calculation means uses the Euclidean distance on the N-gram language model to calculate the word. Since the distance is measured, there is an effect that a language model learning device with improved recognition accuracy can be obtained.

According to the eleventh aspect of the present invention, in any one of the sixth to ninth aspects, the statistical inter-word distance calculating means uses the cross entropy on the N-gram language model to calculate the word. Since the distance is measured, there is an effect that a language model learning device with improved recognition accuracy can be obtained.

According to a twelfth aspect of the present invention, there is provided a speech recognition apparatus using any one of the first to eleventh language model learning apparatuses, wherein the language model or the class language model is a speech model. Since it is used for recognition, there is an effect that a highly accurate speech recognition device can be obtained.

According to a thirteenth aspect of the present invention, a first step of accumulating text data of a target task, a second step of accumulating text data of a general task including tasks other than the target task, A similar word pair extraction step, a similar word string synthesis step, and a language model generation step for reading text data for language model learning from the target task language data and the general task language data, respectively, and constructing a task-adapted language model. The similar word pair extraction step reads each text data from the target task language data and the general task language data, and executes a similar word pair extraction step based on a combination of a word included in the target task text data and a word included in the general task text data. A similar word pair is extracted, and a similar word string synthesizing step includes: The text data and similar word pairs are read, a word string including the word in the target task not included in the language data is synthesized and output, and the language model generation step reads each text data and word string, and Since the task-adapted language model is generated by calculating the statistic of the word string by weighting, the language model learning method with improved recognition accuracy is obtained.

According to claim 14 of the present invention, a first step of accumulating text data of a target task, a second step of accumulating text data of a general task including tasks other than the target task, A target task word classifying step, a general task word classifying step, and a language model generating step for constructing a task-adapted language model from the target task language data and the general task language data; Reads the text data of the target task from the target task language data, replaces the word with the class indicated in the class definition, outputs the first text data classified for language model learning, and outputs the general task word class The generalization step converts general task language data to general task text data. Reading, replacing the word with the class indicated in the class definition, and outputting classified second text data for language model learning, the language model generating step reads the first and second text data, Since the language model is generated by weighting each text data to obtain the statistic of the word string, a language model learning method with improved recognition accuracy can be obtained.

According to the fifteenth aspect of the present invention, a first step of accumulating text data of a target task, a second step of accumulating text data of a general task including tasks other than the target task, A target task word classifying step, a general task word classifying step, a similar word pair extraction step, a similar word string synthesizing step, and a language model for constructing a task-adapted language model from the target task language data and the general task language data. Generating a target task word classifying step, reads text data of the target task from the target task language data, replaces the word with the class indicated in the class definition, and classifies the target task 1 is output, and the general task word classifying step is The general task text data is read from the task language data, the words are replaced with the classes indicated in the class definition, the second text data classified for language model learning is output, and the similar word pair extraction step includes: The first and second text data are read, and a similar word pair is extracted from a combination of a word included in the text data of the target task and a word included in the text data of the general task.
The first and second text data and the similar word pair are read, a word string including a word in the target task not included in the language data is synthesized and output, and the language model generating step includes the first and second language data. A task-adapted language model is generated by reading text data and word strings and calculating the weight of the word strings by weighting each text data, so a language model learning method with improved recognition accuracy has been developed. There is an effect that can be obtained.

According to claim 16 of the present invention, a first step of accumulating text data of a target task, a second step of accumulating text data of a general task including tasks other than the target task, Create an initial language model using text data prepared in advance 3
And a similar word pair extraction step and a similar word probability correction step for constructing a task-adapted statistical language model from the target task language data, the general task language data, and the initial language model. The pair extraction step reads text data for language model learning from the target task language data and the general task language data, respectively, and extracts the words included in the target task text data and the words included in the general task text data. The similar word probability correction step extracts similar word pairs and generates a task-adapted statistical language model by reading the similar word pairs and the initial language model and smoothing the appearance probability of words appearing in the target task. Linguistic modelling with improved recognition accuracy There is an effect in which the method can be obtained.

According to the seventeenth aspect of the present invention, a first step of accumulating text data of a target task, a second step of accumulating text data of a general task including tasks other than the target task, A fourth step of creating an initial class language model in advance, and a target task word classifying step of constructing a task-adapted class language model from the target task language data, the general task language data, and the initial class language model;
A general task word classifying step, a similar word pair extracting step, and a similar word probability correcting step. The target task word classifying step reads text data of the target task from the target task language data, and executes the class indicated in the class definition. And outputs the first text data classified as a class for language model learning. The general task word classifying step reads the general task text data from the general task language data, and indicates the general task word data in the class definition. The word is replaced with the class, and the classified second text data for language model learning is output. The similar word pair extraction step reads the first and second text data and converts the text data of the target task into text data of the target task. Combination of words included and words included in text data of general tasks The similar word probability correction step reads the similar word pair and the initial class language model, and smoothes the appearance probability of the words appearing in the target task to obtain a task-adapted class language model. Since it is generated, there is an effect that a language model learning method with improved recognition accuracy can be obtained.

According to the eighteenth aspect of the present invention, in the thirteenth aspect or the sixteenth aspect, the similar word extracting step includes the step of generating a language model for calculating a distance, the step of calculating a statistical inter-word distance, and the step of determining a threshold value. The language model generation step for distance calculation includes reading text data for language model learning from the target task language data and the general task language data, and weighting each text data to obtain a statistic of the word string. Generating a statistical language model for distance calculation, and calculating a statistical inter-word distance. The statistical language model reads the statistical language model, and executes a statistical process on the statistical language model for a word pair composed of words extracted from each text data. Is determined as the inter-word distance, and the threshold determination step reads the word pair and the inter-word distance, Since so as to output a word pairs exceeding certain threshold, the language model learning method with improved recognition accuracy is the effect obtained.

According to a nineteenth aspect of the present invention, in the thirteenth or sixteenth aspect, the similar word extracting step includes the step of creating a language model for calculating a distance, the step of calculating a statistical inter-word distance, and the step of determining a threshold value. The step of creating a language model for distance calculation includes creating a language model for distance calculation using text data prepared in advance, and the step of calculating a statistical inter-word distance reads the language model for distance calculation, For a word pair consisting of words extracted from the data, a statistical distance on a distance calculation language model is obtained as an inter-word distance. The threshold value determining step reads the word pair and the inter-word distance, and reads a predetermined threshold value. Is output, a word model learning method with improved recognition accuracy is obtained.

According to a twentieth aspect of the present invention, in the fifteenth or seventeenth aspect, the similar word extracting step includes the step of generating a language model for calculating a distance, the step of calculating a statistical inter-word distance, and the step of determining a threshold value. The step of generating a language model for distance calculation includes reading the first and second text data, obtaining a statistic of the word string by weighting each text data, and generating a statistical language model for distance calculation. The statistical inter-word distance calculating step reads the statistical language model, finds a statistical distance on the statistical language model as a word-to-word distance for a word pair composed of words extracted from each text data, In the value determination step, the word pairs and the inter-word distance are read, and word pairs exceeding a predetermined threshold value are output. The effect of the language model learning method with enhanced is obtained.

According to a twenty-first aspect of the present invention, in the fifteenth or seventeenth aspect, the similar word extracting step includes the step of creating a class language model for calculating a distance, the step of calculating a statistical inter-word distance, and the step of Including a determination step, the distance calculation class language model creation step creates a distance calculation class language model using text data prepared in advance, and the statistical inter-word distance calculation step includes a distance calculation class language model and The first and second text data are read, and for a word pair composed of words extracted from each text data, a statistical distance on a distance calculation class language model is obtained as an inter-word distance. Since word pairs and inter-word distances are read and word pairs exceeding a predetermined threshold are output, The effect of the language model learning method with enhanced 識精 degree is obtained.

According to a twenty-second aspect of the present invention, in any one of the eighteenth to twenty-first aspects, the step of calculating a statistical inter-word distance uses the Euclidean distance on an N-gram language model to calculate a word. Since the distance is measured, there is an effect that a language model learning method with improved recognition accuracy can be obtained.

According to a twenty-third aspect of the present invention, in any one of the eighteenth to twenty-first aspects, the statistical word-to-word distance calculating step uses the cross entropy on the N-gram language model to calculate the word. Since the distance is measured, there is an effect that a language model learning method with improved recognition accuracy can be obtained.

According to claim 24 of the present invention, there is provided a speech recognition method using any of the language model learning methods according to claims 13 to 23, wherein the language model or the class language model is a speech model. Since it is used for recognition, there is an effect that a highly accurate voice recognition method can be obtained.

According to the twenty-fifth aspect of the present invention, there is an effect that a recording medium storing any one of the language model learning methods according to the thirteenth to twenty-third aspects is obtained.

According to the twenty-sixth aspect of the present invention, there is an effect that a recording medium storing a speech recognition method using the language model learning method of the twenty-fourth aspect is obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram schematically showing a language model learning apparatus according to Embodiment 1 of the present invention.

FIG. 2 is a flowchart showing a processing procedure of the language model learning apparatus and method according to the first embodiment of the present invention.

FIG. 3 is a block diagram schematically showing a language model learning device according to a second embodiment of the present invention.

FIG. 4 is a flowchart showing a processing procedure of a language model learning apparatus and method according to Embodiment 2 of the present invention;

FIG. 5 is a block diagram schematically showing a language model learning apparatus according to Embodiment 3 of the present invention.

FIG. 6 is a flowchart illustrating a processing procedure of a language model learning apparatus and method according to Embodiment 3 of the present invention;

FIG. 7 is a block diagram schematically showing a language model learning apparatus according to Embodiment 4 of the present invention.

FIG. 8 is a flowchart illustrating a processing procedure of a language model learning apparatus and method according to Embodiment 4 of the present invention;

FIG. 9 is a block diagram schematically showing a language model learning apparatus according to Embodiment 5 of the present invention.

FIG. 10 is a flowchart showing a processing procedure of a language model learning apparatus and method according to Embodiment 5 of the present invention.

FIG. 11 is a functional block diagram showing a specific example of a similar word pair extraction unit of the language model learning apparatus according to Embodiment 6 of the present invention;

FIG. 12 is a flowchart showing a processing procedure of a similar word pair extracting unit of the language model learning device according to the sixth embodiment of the present invention.

FIG. 13 is a functional block diagram showing a second specific example of the similar word pair extracting means according to the sixth embodiment of the present invention.

FIG. 14 is a functional block diagram showing a third specific example of the similar word pair extracting means according to the sixth embodiment of the present invention.

FIG. 15 is a functional block diagram showing a fourth specific example of the similar word pair extracting means according to the sixth embodiment of the present invention.

FIG. 16 is a block diagram schematically showing a speech recognition device using a language model learning device according to a seventh embodiment of the present invention.

FIG. 17 is a flowchart showing a processing procedure of the speech recognition device using the language model learning device according to the seventh embodiment of the present invention.

FIG. 18 is a block diagram schematically showing a speech recognition device using a language model learning device according to an eighth embodiment of the present invention.

FIG. 19 is a flowchart showing a processing procedure of a speech recognition device using the language model learning device according to the eighth embodiment of the present invention.

FIG. 20 is an explanatory diagram showing an example of a general class definition.

FIG. 21 is a block diagram schematically showing a conventional language model learning device.

FIG. 22 is a flowchart showing a processing procedure by a conventional language model learning apparatus and method.

[Explanation of symbols]

101 target task language data, 102 general task language data, 103, 103B, 103C, 103D, 1
03E, 103F Similar word pair extraction means, 104, 10
4B Similar Word String Synthesis Means, 105, 105A, 105
B language model generating means, 301 target task word classifying means, 302 general task word classifying means and language model generating means, 701 initial language model, 702 similar word probability correcting means, 901 initial class language model, 1101, 1101D, 1101F Statistical word-to-word distance calculation means, 1102, 1102E Threshold determination means, 1105, 1105E Distance calculation language model generation means, 1301 Distance calculation language model, 150
1 Class language model for distance calculation, 1605 language model, 1802 class language model.

Claims (26)

[Claims] 1. A target task language data in which text data of a target task is accumulated, a general task language data in which text data of general tasks including tasks other than the target task are accumulated, the target task language data and the general task language A similar word pair extracting means, a similar word string synthesizing means, and a language model generating means for reading text data for language model learning from data, respectively, and constructing a task-adapted language model; The extracting unit reads each text data from the target task language data and the general task language data, and extracts a similar word pair from a combination of a word included in the target task text data and a word included in the general task text data. The similar word string synthesizing means extracts each of the texts. Reading the similar word pair from the similar word pair extracting unit, and synthesizing and outputting a word string including a word in the target task that is not included in the language data, the language model generating unit includes: Generating the task-adapted language model by reading each text data, reading the word string from the similar word string synthesizing means, and weighting the text data to obtain a statistic of the word string. A language model learning device characterized by the following. 2. A target task language data in which text data of a target task is accumulated, a general task language data in which text data of general tasks including tasks other than the target task are accumulated, the target task language data and the general task language A target task word classifier, a general task word classifier, and a language model generator for constructing a task-adapted language model from data, wherein the target task word classifier includes the target task language data. Read the text data of the target task from
Replacing the word with the class indicated in the class definition and outputting first text data classified for language model learning, wherein the general task word classifying means converts the general task language data into a general task text Read the data,
Replacing the word with the class indicated in the class definition and outputting classified second text data for language model learning; the language model generating means reads the first and second text data; A language model learning apparatus, wherein the language model is generated by obtaining a statistic of a word string by weighting each text data. 3. A target task language data in which text data of a target task is accumulated, a general task language data in which text data of general tasks including tasks other than the target task are accumulated, the target task language data and the general task language A target task word classifying unit, a general task word classifying unit, a similar word pair extracting unit, a similar word string synthesizing unit, and a language model generating unit for constructing a task-adapted language model from data; The task word classifying means reads text data of the target task from the target task language data,
Replacing the word with the class indicated in the class definition and outputting first text data classified for language model learning, wherein the general task word classifying means converts the general task language data into a general task text Read the data,
The word is replaced with the class indicated in the class definition, and the classified second text data for language model learning is output. The similar word pair extracting unit reads the first and second text data. Extracting a similar word pair from a combination of a word included in the text data of the target task and a word included in the text data of the general task, wherein the similar word string synthesizing unit includes the first and second text data. And reads the similar word pair from the similar word pair extraction unit, synthesizes and outputs a word string including a word in the target task that is not included in the language data, and outputs the first language model. And the second text data, and reads the word string from the similar word string synthesizing means. By determining the statistics of the word string put,
A language model learning apparatus for generating the task-adapted language model. 4. Using the target task language data in which text data of the target task is accumulated, the general task language data in which text data of general tasks including tasks other than the target task are accumulated, and text data prepared in advance A similar word pair extraction unit and a similar word probability correction unit for constructing a task-adapted statistical language model from the target language data, the general task language data, and the initial language model. The similar word pair extracting means reads text data for language model learning from the target task language data and the general task language data, respectively, and reads a word included in the text data of the target task and the general task. Similar words from the combination with the words contained in the text data of The similar word probability correction unit reads the similar word pair from the similar word pair extraction unit, reads the initial language model, and performs smoothing of the appearance probability of the word appearing in the target task. A language model learning apparatus for generating the task-adapted statistical language model. 5. A target task language data in which text data of a target task is collected, general task language data in which text data of general tasks including tasks other than the target task are collected, an initial class language model created in advance, From the target task language data, the general task language data and the initial class language model, a target task word classifier, a general task word classifier, and similar word pair extraction for constructing a task-adapted class language model Means and a similar word probability correction means, the target task word classifying means reads text data of the target task from the target task language data,
Replacing the word with the class indicated in the class definition and outputting first text data classified for language model learning, wherein the general task word classifying means converts the general task language data into a general task text Read the data,
The word is replaced with the class indicated in the class definition, and the classified second text data for language model learning is output. The similar word pair extracting unit reads the first and second text data. Extracting a similar word pair from a combination of a word included in the text data of the target task and a word included in the text data of the general task, wherein the similar word probability correction unit includes the similar word pair extraction unit Reading the word pair, reading the initial class language model, and smoothing the appearance probability of the word appearing in the target task, thereby generating the task-adapted class language model. apparatus. 6. The similar word extraction means includes a distance calculation language model generation means, a statistical inter-word distance calculation means, and a threshold value determination means, wherein the distance calculation language model generation means includes the target task language. Reading text data for language model learning from the data and the general task language data, obtaining a statistic of a word string by weighting each text data, generating a statistical language model for distance calculation, The inter-word distance calculation means reads the statistical language model from the distance calculation language model generation means,
For a word pair consisting of words extracted from each of the text data, a statistical distance on the statistical language model is determined as an inter-word distance. The language model learning device according to claim 1, wherein a word pair and the inter-word distance are read, and a word pair exceeding a predetermined threshold is output. 7. The similar word extraction means includes a distance calculation language model, a statistical inter-word distance calculation means, and a threshold value determination means, and the distance calculation language model uses text data prepared in advance. The statistical inter-word distance calculating means reads the language model for distance calculation, and, for a word pair composed of words extracted from each of the text data, statistically calculates the word pair on the language model for distance calculation. Calculating a distance as a word-to-word distance, wherein the threshold value determination means reads the word pair and the word-to-word distance from the statistical word-to-word distance calculation means, and outputs a word pair exceeding a predetermined threshold value. The language model learning device according to claim 1 or 4, wherein 8. The similar word extraction means includes a distance calculation language model generation means, a statistical inter-word distance calculation means, and a threshold value determination means, wherein the distance calculation language model generation means includes the target task word The first and second text data are read from the classifying means and the general task word classifying means, weighted for each text data to obtain a word string statistic, and a statistical language model for distance calculation is generated. The statistical inter-word distance calculating means reads the statistical language model from the distance calculating language model generating means,
For a word pair consisting of words extracted from each of the text data, a statistical distance on the statistical language model is determined as an inter-word distance. The language model learning device according to claim 3, wherein a word pair and the inter-word distance are read, and a word pair exceeding a predetermined threshold is output. 9. The similar word extraction means includes a distance calculation class language model, a statistical inter-word distance calculation means, and a threshold value determination means, wherein the distance calculation class language model includes text data prepared in advance. The statistical inter-word distance calculation means reads the distance calculation class language model and obtains first and second words from the target task word classification means and the general task word classification means. , And for a word pair consisting of words extracted from each of the text data, a statistical distance on the distance calculation class language model is determined as an inter-word distance. The word pair and the inter-word distance are read from the target inter-word distance calculation means, and a word pair exceeding a predetermined threshold is output. The language model learning device according to claim 3 or 5, wherein: 10. The statistical inter-word distance calculating means, wherein N
10. The language model learning apparatus according to claim 6, wherein the inter-word distance is measured using a Euclidean distance on a Gram language model. 11. The statistical inter-word distance calculating means includes:
10. The language model learning apparatus according to claim 6, wherein the inter-word distance is measured using cross entropy on a Gram language model. 12. The speech recognition device using the language model learning device according to claim 1, wherein the language model or the class language model is used for speech recognition. 13. A first step of accumulating text data of a target task, a second step of accumulating text data of a general task including tasks other than the target task, the target task language data and the general task language A similar word pair extracting step, a similar word string synthesizing step, and a language model generating step for reading text data for language model learning from the data and constructing a task-adapted language model, respectively. The extracting step reads each text data from the target task language data and the general task language data, and sets a similar word pair from a combination of a word included in the target task text data and a word included in the general task text data. And the similar word string synthesizing step comprises: The text data and the similar word pair are read, and a word string including a word in the target task that is not included in the language data is synthesized and output.The language model generating step includes: By reading, by weighting for each of the text data to determine the statistics of the word string,
A language model learning method, characterized by generating the task-adapted language model. 14. A first step of accumulating text data of a target task, a second step of accumulating text data of a general task including a task other than the target task, the target task language data and the general task language A target task word classifying step, a general task word classifying step, and a language model generating step for constructing a task-adapted language model from data, wherein the target task word classifying step includes the target task language data. Read the text data of the target task from, replace the word with the class indicated in the class definition,
Outputting first classified text data for language model learning, the general task word classifying step reads text data of a general task from the general task language data, and converts the word into a class indicated in the class definition; Is replaced by
Outputting classified second text data for language model learning; the language model generating step reads the first and second text data, weights each of the text data, and calculates a statistic of a word string. A language model learning method, wherein the language model is generated by calculating 15. A first step of accumulating text data of a target task, a second step of accumulating text data of general tasks including tasks other than the target task, the target task language data and the general task language A target task word classifying step, a general task word classifying step, a similar word pair extracting step, a similar word string synthesizing step, and a language model generating step for constructing a task-adapted language model from data; The task word classifying step reads text data of the target task from the target task language data and replaces the word with the class indicated in the class definition,
Outputting first classified text data for language model learning, the general task word classifying step reads text data of a general task from the general task language data, and converts the word into a class indicated in the class definition; Is replaced by
Outputting classified second text data for language model learning, the similar word pair extracting step reads the first and second text data, and outputs a word included in the text data of the target task; Extracting a similar word pair from a combination of words included in the text data of the general task; and the similar word string synthesizing step reads the first and second text data and the similar word pair and includes the similar word pair in the language data. The language model generating step reads the first and second text data and the word string, and weights the text data for each of the text data. A language model learning method for generating the task-adapted language model by calculating a statistic of the word string. . 16. A first step of accumulating text data of a target task, a second step of accumulating text data of a general task including tasks other than the target task, and initializing using text data prepared in advance. A third step of creating a language model; a similar word pair extraction step and a similarity step for constructing a task-adapted statistical language model from the target task language data, the general task language data, and the initial language model. A word probability correction step, wherein the similar word pair extraction step reads text data for language model learning from the target task language data and the general task language data, respectively, and includes words included in the text data of the target task. And words contained in the text data of the general task Extracting a similar word pair from a set of words, the similar word probability correcting step reads the similar word pair and the initial language model, and performs smoothing of an appearance probability of a word appearing in the target task, thereby performing the task adaptation. A language model learning method characterized by generating a statistical statistical language model. 17. A first step of accumulating text data of a target task, a second step of accumulating text data of general tasks including tasks other than the target task, and a fourth step of preparing an initial class language model in advance. From the target task language data, the general task language data and the initial class language model, to construct a task-adapted class language model, target task word classifying step, general task word classifying step,
A similar word pair extraction step and a similar word probability correction step, wherein the target task word classifying step reads text data of the target task from the target task language data and replaces the word with the class indicated in the class definition. ,
Outputting first classified text data for language model learning, the general task word classifying step reads text data of a general task from the general task language data, and converts the word into a class indicated in the class definition; Is replaced by
Outputting classified second text data for language model learning, the similar word pair extracting step reads the first and second text data, and outputs a word included in the text data of the target task; Extracting a similar word pair from a combination with a word included in the text data of the general task; the similar word probability correction step reads the similar word pair and the initial class language model, and appears a word appearing in the target task. A language model learning method, wherein the task-adapted class language model is generated by performing probability smoothing. 18. The step of extracting a similar word includes a step of generating a language model for calculating a distance, a step of calculating a statistical inter-word distance, and a step of determining a threshold value. Data and the general task language data,
The respective text data for language model learning are read, the weight of each text data is weighted to obtain a statistic of the word string, and a statistical language model for distance calculation is generated. The statistical language model is read, and a statistical distance on the statistical language model is obtained as a word-to-word distance for a word pair composed of words extracted from each of the text data. And reading the word-to-word distance and outputting a word pair exceeding a predetermined threshold value.
Language model learning method described in. 19. The similar word extraction step includes a distance calculation language model creation step, a statistical inter-word distance calculation step, and a threshold value determination step, and the distance calculation language model creation step is prepared in advance. Creating the language model for distance calculation using text data, the statistical inter-word distance calculation step reads the language model for distance calculation, and for the word pair consisting of words extracted from each of the text data, Calculating a statistical distance on the calculation language model as a word-to-word distance; and reading the word pair and the word-to-word distance, and outputting a word pair exceeding a predetermined threshold value. Claim 13 or Claim 16 characterized by the above-mentioned.
Language model learning method described in. 20. The similar word extraction step includes a distance calculation language model generation step, a statistical inter-word distance calculation step, and a threshold value determination step, wherein the distance calculation language model generation step includes the first and second language models. The second text data is read, a weight of each text data is weighted, a statistic of the word string is obtained, a statistical language model for distance calculation is generated, and the statistical inter-word distance calculating step includes: Reading a model, for a word pair consisting of words extracted from each of the text data, determining a statistical distance on the statistical language model as an inter-word distance, the threshold value determining step includes: 18. The method according to claim 15, wherein the inter-distance is read and a word pair exceeding a predetermined threshold is output.
Language model learning method described in. 21. The similar word extraction step includes a distance calculation class language model creation step, a statistical inter-word distance calculation step, and a threshold value determination step, wherein the distance calculation class language model creation step includes: The distance calculation class language model is created using the prepared text data, and the statistical inter-word distance calculation step reads the distance calculation class language model and the first and second text data, and For a word pair composed of words extracted from text data, a statistical distance on the distance calculation class language model is determined as an inter-word distance. The threshold value determining step reads the word pair and the inter-word distance. 15. A word pair exceeding a predetermined threshold value is output. 7
Language model learning method described in. 22. The method according to claim 18, wherein in the statistical inter-word distance calculating step, the inter-word distance is measured using a Euclidean distance on an N-gram language model. Language model learning method described in. 23. The method according to claim 18, wherein in the step of calculating a statistical inter-word distance, the inter-word distance is measured using cross entropy on an N-gram language model. Language model learning method described in. 24. The speech recognition method using the language model learning method according to claim 13, wherein the language model or the class language model is used for speech recognition. 25. A storage medium storing the language model learning method according to any one of claims 13 to 23. 26. A storage medium storing a speech recognition method using the language model learning method according to claim 24.