JP2002366190A - Statistical language model generator and statistical language model generator - Google Patents

Abstract

(57) [Summary] [Problem] To provide a statistical language model generation device that generates a statistical language model that can improve recognition performance in speech recognition and increase the probability of appearance of words that are likely to be included in utterance contents. Provide a statistical language model generator. SOLUTION: A statistical language model generation device includes a speech recognition unit 7 for recognizing the most recent text data and past mass text data as speech, the latest text data (accumulation unit 5), and the past mass text data. (Accumulating means 3), accumulating means 9 for accumulating recognition results recognized by the voice recognizing means 7, and a first probability weight of n-gram is calculated based on a large amount of past text data. Based on the text data of
a probability weight calculating means for calculating a second probability weight of n-gram, and calculating a third probability weight of n-gram, based on the recognition result, a first probability weight, and a second probability weight; and A language model generating means for generating a statistical language model based on the third probability weight.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a statistical language model generating apparatus and a statistical language model generating program for generating a statistical language model for use in a speech recognition apparatus.

[0002]

2. Description of the Related Art Conventionally, a statistical (stochastic) method has been used as a method for improving speech recognition performance in a speech recognition apparatus.
A method using a language model has been proposed, and typical ones are as follows. The statistical language model is a model in which the relationship between words and phonemes in a language is modeled based on statistics.

(1) A method based on a cache model (R.
Kuhn, R .; De Mori “ACache-Bas
ed Natural Language Model
for Speech Recognition, "I
EEE Trans. PAMI, vol. 12, no.
6, 1990, p. 570-583). This method
N learned from a large number of past manuscripts (text data)
This is a method for improving the speech recognition performance by combining the gram probability value and the word appearance probability in the latest speech recognition result by linear interpolation or the like. It should be noted that the n-gram probability value is an occurrence probability in a word n-gram in which a word sequence is modeled as a Markov chain, that is, the occurrence probability of a certain word is the immediately preceding (n -1) It depends on words. What is linear interpolation?
An n-gram probability value and a lower-order m-gram probability value (m
<N) is interpolated linearly.

(2) Method based on MAP (posterior probability maximization) estimation (Kobayashi, Imai, Ando, “Time Depen
Dent Language Model for B
loadcast News Transcripti
on and Its Postcorrection
n, “ICSLP-1998). This method adds the n-gram probability value of a task to a large number of documents independent of a task by adding a small amount of the document with an appropriate weight obtained by MAP estimation, This is a method to improve the statistical accuracy of the language model and improve the recognition performance.The vocabulary (corpus) for generating the language model includes all the words in a small manuscript and a part of a large manuscript. It should be noted that a task is generally a task, that is, an object to be processed, and an appropriate weight is statistical (probability) ) A numerical value added so as to increase the probability of occurrence of a certain word in a language model. A vocabulary (corpus) is data used as a basis for generating a language model. More than 100,000 words Do it is that of text database.

[0005]

However, in the conventional cache model-based method, a word that is not registered in a vocabulary for generating a language model (as recently used widely) is used in order to use past speech recognition results. Is not taken into account. For this reason, tasks such as news programs (news, etc.) in which one topic is often composed of a small number of sentences, and in which proper nouns (new words) such as personal names, place names, and organization names are likely to appear very frequently However, there is a problem that the recognition performance of speech recognition cannot be improved unless a language model based on a vocabulary in which new words are registered is always used.

In the method based on MAP (maximum posterior probability) estimation, a manuscript written in written language is used instead of the actual utterance content. There is a problem that can not be.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems of the prior art, to improve the recognition performance in speech recognition, and to increase the probability of appearance of words likely to be included in utterance contents. An object of the present invention is to provide a statistical language model generation device for generating a model and a statistical language model generation program.

[0008]

According to a first aspect of the present invention, there is provided a statistical language model generating apparatus for generating a statistical language model, wherein a word whose appearance frequency is predicted to be high is increased. Text data obtaining means for obtaining the most recent text data, including the most recent text data and a larger amount of data than the most recent text data, a voice recognition means for recognizing a large amount of past text data before that as a voice, The latest text data,
A storage unit for storing the past large amount of text data and the recognition result by the voice recognition unit; and a first probability weight of n-gram is calculated based on the past large amount of text data, and the most recent text is calculated. Based on the data,
calculating a second probability weight of n-gram; calculating a third probability weight of n-gram based on the recognition result; a first probability weight; Language model generating means for generating a statistical language model based on the probability weight and the third probability weight.

According to this configuration, the latest text data including the word whose appearance frequency is predicted to be high is obtained by the text data obtaining means, and the latest text data and a large amount of the previous text data are obtained by the voice recognition means. The text data is recognized as voice, and the storage unit stores the latest text data, past large-volume text data, and the recognized recognition result. Then, the respective n-gras are calculated by the probability weight calculating means.
The probability weight for m is calculated, and a language model is generated by the language model generating means based on the calculation result.

The most recent text data that is expected to appear more frequently is, for example, a manuscript provided for a broadcast program immediately before or immediately after a broadcast, a newspaper or a magazine just before or immediately after a release, etc. Articles and the like. Further, the past large amount of text data includes, for example, manuscripts provided for broadcast programs for several years to several decades, a brown corpus, a LOB corpus, and the like.

According to a second aspect of the present invention, there is provided a statistical language model generating apparatus according to the first aspect, further comprising a recognition result correcting means for correcting a recognition result recognized by the speech recognition means. The probability weight calculating means calculates a third probability weight of n-gram based on the corrected recognition result.

[0012] According to this configuration, the result of the recognition of the text data as speech is corrected by the recognition result correcting means, and based on the corrected recognition result, the probability weight calculating means calculates the third n-gram. Is calculated.

According to a third aspect of the present invention, there is provided a statistical language model generation program, comprising: a computer for obtaining text data that includes the most recent text data including a word whose appearance frequency is predicted to be high; Speech recognition means for recognizing a large amount of past large-volume text data as voice having a larger data amount than the latest text data as speech, the latest text data, and the large-volume text data of the past, and the speech recognition means. Storage means for storing the recognized recognition result, based on the large amount of past text data,
m, a second probability weight of n-gram is calculated based on the latest text data, and a third probability weight of n-gram is calculated based on the recognition result. Probability weight calculation means to calculate, the first probability weight, and the second probability weight, and, based on the third probability weight, based on the language model generation means to generate a statistical language model, Features.

According to this configuration, the latest text data including the word whose appearance frequency is predicted to be high is obtained by the text data obtaining unit, and the latest text data and the large amount of the previous text data are obtained by the voice recognition unit. The text data is recognized as voice, and the storage unit stores the latest text data, past large-volume text data, and the recognized recognition result. Then, the respective n-gras are calculated by the probability weight calculating means.
The probability weight for m is calculated, and a language model is generated by the language model generating means based on the calculation result.

[0015]

Embodiments of the present invention will be described below in detail with reference to the drawings. (Statistical Language Model Generator: First Embodiment) FIG.
FIG. 3 shows a functional explanatory diagram of the first embodiment of the statistical language model generation device. As illustrated in FIG. 1, the statistical language model generation device 1 includes a main control unit, a storage unit, a display output unit, an input unit, an external connection unit, and the like, which are not illustrated.
The past news manuscript accumulation means 3, the latest report manuscript accumulation means 5, the speech recognition means 7, the recognition result accumulation means 9, and the language model calculation means 11 are functionally realized.

The statistical language model generation device 1 is a device for generating a statistical language model to be used at the time of speech recognition in a speech recognition device (not shown) based on a large amount of text data. In this embodiment, the statistical language model generation device 1 is a general computer, and each main control unit (not shown), storage unit, display output unit, input unit, and external connection unit are a CPU, Memory, hard disk,
It is composed of a keyboard and the like.

The past news manuscript accumulating means 3 is a database stored (accumulated) in a storage unit, not shown, in which a large amount of past text data described in the claims is accumulated. The past news manuscript accumulation means 3 stores a large amount of past news manuscripts in a text file format (text data). This text file is a file in which a space is interposed between each word constituting the manuscript.

In this embodiment, the punctuation marks included in the news manuscript are integrated into the word immediately before the punctuation marks and handled. Further, to supplement, in this specification, the terms storage, accumulation, and accumulation are described as having substantially no difference.

The most recent reporter manuscript accumulation means 5 is a program developed in the main control unit (not shown) and a database stored in the storage unit, and includes a most recent word including a word whose appearance frequency is predicted to be high. The text data acquiring means for acquiring text data and the latest text data acquired are integrated. The latest report manuscript accumulation means 5 first obtains a reporter manuscript provided for the latest broadcast program (especially a news program). The acquisition method is performed by an operator (operator) inputting a news manuscript into the statistical language model generation device 1 or reading the news manuscript using an OCR or the like, and inputting the read result via an external connection unit.

In the latest report manuscript stacking means 5, after the reporter manuscript is obtained, a certain correction is automatically made or the operator proofreads it as text data.
It is stored in a latest text database in a storage unit (not shown). Each sentence in the reporter manuscript is converted into a text file in units of one topic, and this text file has a space between each word constituting the reporter manuscript, similarly to the past news manuscript accumulation means 3. It is configured as a sandwich.

The voice recognition means 7 recognizes a text file as voice (reads out the text file). The speech recognition means 7 is a general text-to-speech conversion engine or the like. This text-to-speech conversion engine is equipped with a dictionary storing hundreds of thousands of words. First, the past news manuscript accumulation means 3 and the latest report manuscript accumulation means From the text file No. 5, the words included in this file are recognized. Then, the recognized words are synthesized by, for example, a waveform superposition method or the like, and output in a PCM output format.

The recognition result accumulating means 9 is a database stored (accumulated) in a storage unit (not shown), and stores the past news manuscript accumulating means 3 and the latest report manuscript in the recognition result recognized by the voice recognizing means 7. The text file of the accumulating means 5 is referred to, and a date and a time are given as a time stamp for each sentence of the recognition result and are accumulated.

The language model calculating means 11 is based on the past mass text data by the past news manuscript accumulating means 3, the text data of the latest report manuscript by the latest report manuscript accumulating means 5, and the recognition result by the voice recognition means 7. And a program for generating a statistical language model. In this embodiment, a bigram model is used as a language model (for an n-gram model including a bigram model, see, for example, “Speech Recognition by Probabilistic Model”, Seiichi Nakagawa, IEICE, pp.109). I have.

In the language model calculating means 11, various calculations are performed in the following order based on mathematical expressions described later. First of all, the past of a large amount text data, and text data of the most recent press manuscript, on the basis of the recognition result, for the linear interpolation (linear interpolation the bigramP of the language model _0, P _1, P ₂ is, for example, "spoken language Processing ”, co-authored by Kita, Nakamura, and Nagata, Morikita Publishing, pp. 29), the weighted language model is

[0025]

(Equation 1) Represented by

In Equation 1, y _n and y _n-1 are words registered in the vocabulary. The probability P (y _n | y _n-1 ) is
After word y _n-1 is uttered, it means a probability of words y _n is uttered. Generally, in an n-gram language model, n
The larger the is, the longer the continuous word string is handled, and the higher the recognition accuracy of the next word is. However, instead of increasing the recognition accuracy, text data including an enormous amount (n times the number of words) of vocabulary is required. λ indicates a probability weight in each language model, and V indicates a vocabulary.

The word y of the weighted language model_n, Y
_n-1If the bigram for
The language model generated by the word model generation device 1
When provided to a voice recognition device (not shown),
During speech recognition, those y _nAnd y_n-1Combination with
Is more likely to appear. In other words, when the device recognizes voice
Then, for a sentence to be read out (a sentence recognized by speech), b
The value of the probability weight λ is determined so that the product of i.g.
It will be good if it is set. Alternatively, evaluation data (voice
Entropy (eg, “probability model”)
Speech Recognition by Le ", Seiichi Nakagawa, IEICE, p
p. 111, and Equation 2) are minimized.
It suffices if the value of the weight λ is determined.

[0028]

(Equation 2)

In Equation 2, N is in the evaluation text.
Indicates the total number of words in (text data)
The text is the word string y = y of the evaluation data₁y_Two・ ・ ・ ・ ・
・ Y _NIt is assumed that Note that λ in this equation is the period
Waiting value maximization algorithm (EM algorithm (EM algorithm)
For the algorithm, for example, "speech language processing", north,
Nakamura, Nagata co-author, Morikita Publishing, pp. 31)))
And it is obtained by repeated calculation using Equation 3.
It is.

[0030]

(Equation 3)

In Equation 3, while updating λ _i , calculation is repeated until the entropy for the evaluation text converges. By this calculation, the probability weight λ for each language model can be automatically obtained. However, it is usually difficult to obtain the optimal probability weight λ for the text to be read because the content of the evaluation text is unknown. For this reason, a transcript of the known utterance content (converted into text data) relating to the evaluation text is prepared in advance, and the value of the probability weight λ is experimentally obtained using this.

Next, a text weight w is obtained. The value of this text weight w gives the weighted word frequency. The total number of words in the past mass text data G ₀ is m ₀
If the total number of words in the latest text data G ₁ is m ₁ and the total number of words in the recognition result G ₂ is m ₂ , the text weight w is
It is calculated by Equation 4 using the converged probability weights λ ₀ , λ ₁ , and λ ₂ (the first probability weight, the second probability weight, and the third probability weight of n-gram).

[0033]

(Equation 4)

In equation (4), the number of additions (text weights) w ₁ , w of the latest text data G ₁ and the recognition result G ₂ added to the past mass text data G _0
2 is calculated from the probability weights λ ₀ , λ ₁ , λ ₂ . In Equation 4, the probability weight λ in the statistical language model is normalized to the probability weight in a set of a plurality of text data.

Based on the calculated text weights w ₁ and w ₂ , the most recent text data is weighted by the text weight w ₁ , and the recognition result is weighted by the text weight w ₂ , and is added to the past mass text data. New vocabulary is required. In other words, the frequency f of a word, the frequency f ₀ in the past of a large amount text data G _0, frequency f ₁ of the most recent text data G _1, and the frequency f ₂ of the recognition result G _2,

[0036]

(Equation 5) Becomes

Then, words are vocabulary in descending order of frequency f.
Register with V. However, the maximum number of registered vocabulary
(V_max) Is set, and make sure that this limit is not exceeded.
Registered in. This limits the total number of registered words in the vocabulary.
While the latest text data G ₁Was included in it
Words that appear less frequently are weighted and added to the vocabulary.
Is recorded.

That is, in the language model calculating means 11 of the statistical language model generating apparatus 1, the appearance of a new word (not included in the past mass text data) in the latest text data (the latest news manuscript etc.) Frequency is increased. In addition, since a new vocabulary is determined based on the recognition result by the voice recognition means 7, the speech recognition device (not shown) uses the language model generated by the statistical language model generation device 1. Thereby, the recognition performance at the time of speech recognition is improved. In this embodiment, the generated language model is fed back to the speech recognition means 7 and is used again for speech recognition.

(Statistical Language Model Generation Apparatus: Second Embodiment) FIG. 2 is a functional explanatory diagram of a second embodiment of the statistical language model generation apparatus. In the statistical language model generation device 1A, the same components as those of the statistical language model 1 are denoted by the same reference numerals, and description thereof is omitted.

The recognition result correcting means 13 of the statistical language model generating apparatus 1A is a program for correcting the recognition result of the speech recognition means 7, and for example, when the speech recognition means 7 reads out the text data, If the hiragana reads "Ame wa Furu", it is read as "Raining." In other words, in this case, "a"
Will have an accent. Actually, "candy falls"
If (me is accented with “me”), by analogy with the context before and after “
This is to correct the recognition result.

The correction recognition result accumulating means 15 is a database stored (accumulated) in a storage unit not shown, and accumulates (accumulates) the recognition results corrected by the recognition result correcting means 13. Note that the correction recognition result accumulating unit 15 temporarily stores the recognition result obtained by the voice recognition unit 7 before the correction.

The language model calculation means 11A, like the language model calculation means 11, performs various calculations in the following order. Note that, in this embodiment, a large amount of past text data and the latest text data, and
After recognizing the text data by the voice recognition means 7, the following calculation is performed by the language model calculation means 11A based on the recognition result (evaluation data) corrected by the correction recognition result accumulation means 15.

First, based on the past mass text data, the text data of the most recent reporter's manuscript, and the corrected recognition results, the language models bigramP ₀ , P ₁ ,
P ₂ ′ is represented by linear interpolation (see Equation 1). Then, the probability weight λ of the weighted language model is defined.

If the value of the probability weight λ is determined so that the entropy of the data to be evaluated (past mass text data, text data of the most recent reporter's manuscript, and the corrected recognition result) is minimized, Equation 2 is obtained. Often, the probability weight λ is obtained by repeated calculation by using an expected value maximizing algorithm (see Equation 3).

The converged probability weights λ ₀ , λ ₁ , λ ₂ ′ (n−
The text weights w ₁ and w ₂ ′ are calculated using the first probability weight, the second probability weight, and the third probability weight of the gram (see Equation 4). Calculated text weights w ₁ , w ₂ ′
The most recent text data is based on the text weight w ₁
And the corrected recognition result is the text weight w ₂ ′
, And is added to the past large amount of text data to obtain a new vocabulary.

[0046] In other words, the frequency f of a word, the frequency of the past of the frequency f ₀ of a mass text data G _0, frequency f ₁ of the most recent text data G _1, modified recognition result G ₂ 'f ₂ ′, and is represented by the product of the text weights w ₁ and w ₂ ′ (see Equation 5). Then, the words are registered in the vocabulary V in the descending order of the frequency f. However, an upper limit (V _max ) is set in advance for the number of registered vocabularies, and the vocabulary is registered so as not to exceed the upper limit.

That is, in the language model calculating means 11A of the statistical language model generating apparatus 1A, the appearance of a new word (not included in the past mass text data) in the latest text data (the latest news manuscript etc.) Frequency is increased. Moreover, since the recognition result by the speech recognition means 7 is corrected by the recognition result correcting means 13 and a new vocabulary is determined based on the corrected recognition result, the vocabulary is generated by the statistical language model generating apparatus 1A. The use of the language model by a speech recognition device (not shown) improves the recognition performance at the time of speech recognition. In this embodiment, the generated language model is fed back to the speech recognition means 7 and is used again for speech recognition.

In the statistical language model generating apparatus 1A, the latest reporter document accumulating means 5 acquires and accumulates the latest news program or the like for which speech is to be recognized. The output of the recognition is modified to create a correct string corresponding to the recognized speech. Therefore, this statistical language model generation device 1A
If the language model generated by the above is used, the information of a correct character string for a broadcast program (target of voice recognition) at a very close time is used, and the voice recognition performance can be improved. .

Further, by referring to the information of the correct character string for the broadcast program (target of voice recognition) at a time very close in time, a recognition error included in the voice recognition output of the past mass database is detected. It can be corrected by the recognition result correcting means 13.

(Operation of Statistical Language Model Generating Apparatus) Next, the operation of the statistical language model generating apparatus 1 will be described with reference to the flowchart shown in FIG. First, the past news manuscript accumulation means 3 accumulates (accumulates) a large amount of past text data, and determines an initial vocabulary according to the appearance frequency of each word contained in the past large amount of text data. (S1). Typically, the initial vocabulary is formed from hundreds of thousands of words. Generally, the vocabulary in the language model is set in advance according to the storage capacity of a storage unit (not shown) or the processing capacity of a main control unit (not shown). ,
The words are registered in the vocabulary and determined in the order of words having a high frequency of appearance in the words to be accumulated or learned.

On the other hand, the latest reporter manuscript stacking means 5
Text data (the latest text data) provided for the latest broadcast program or the like is accumulated, and the past large-volume text data and the latest text data are voice-recognized by the voice recognition unit 7. Recognition results obtained by voice recognition are accumulated in the recognition result accumulation means 9.

Then, the language model calculating means 11 first sets each language model (bigram P ₀ , P ₁ ,
P ₂ ) are created and these bigrams P ₀ , P ₁ , P ₂
Is linearly interpolated (see Equation 1) (S2). The probability weights λ ₀ , λ ₁ and λ ₂ of these language models are calculated (calculated) by the EM algorithm (see Equation 3) (S3), and the text weights are calculated based on these probability weights λ ₀ , λ ₁ and λ _2. w ₁ and w ₂ are calculated (calculated) (see Equation 4) (S
4).

Further, the language model calculating means 11 calculates (calculates) the appearance frequency f of the word based on the text weights w ₁ and w ₂ (see Equation 5) (S5). A new vocabulary is determined in the order of the words having the higher appearance frequency f so as to be included in the number of registered words (S6). Then, a language model is generated based on the new vocabulary (S7).

Although the present invention has been described based on the embodiments, the present invention is not limited to these embodiments.

For example, the statistical language model generators 1, 1
Each configuration realized in A can be handled as a program stored in a specific storage medium. Furthermore, n-gram (trigram or more) larger than bigram
m, 4-gram), as in biggram, each probability weight λ is calculated, the probability weight λ is converted into a text weight, a new vocabulary V is created, and a statistical language is calculated from this vocabulary. It is possible to generate a model.

[0056]

According to the first aspect of the present invention, the text data obtaining means obtains the most recent text data including a word whose appearance frequency is predicted to be high,
The latest text data and the previous mass text data are recognized as speech by the voice recognition unit, and the latest text data, the past large text data, and the recognized recognition result are stored by the storage unit. Then, each of the n
The probability weight in -gram is calculated, and the language model is generated by the language model generating means based on the calculation result. If this language model is used in the voice recognition device, the recognition performance at the time of voice recognition is improved. be able to.

Also, since a certain probability weight is added to the most recent text data and included in the vocabulary for generating the language model, the appearance probability of words that are likely to be included in the most recent utterance content can be increased.

According to the second aspect of the present invention, the result of the recognition of the text data as speech is corrected by the recognition result correcting means, and based on the corrected recognition result, n- Since the third probability weight of gram is calculated, if a language model obtained based on the corrected recognition result is used in the speech recognition device, the recognition performance at the time of speech recognition can be further improved.

According to the third aspect of the present invention, the text data obtaining means of the statistical language model generation program obtains the most recent text data including the word whose appearance frequency is predicted to be high, and the voice recognition means As a result, the most recent text data and the previous mass text data are recognized as speech, and the accumulation means accumulates the most recent text data, the past mass text data, and the recognized recognition result. Then, the probability weights in the respective n-grams are calculated by the probability weight calculating means, and the language model is generated by the language model generating means based on the calculation results. This language model is used in the speech recognition device. Thus, the recognition performance at the time of voice recognition can be improved.

It is also possible to distribute the statistical language model generation program in the market as a storage medium storing the program.

[Brief description of the drawings]

FIG. 1 is a functional explanatory diagram of a statistical language model generation device according to a first embodiment of the present invention.

FIG. 2 is a functional explanatory diagram of a statistical language model generation device according to a second embodiment of the present invention.

FIG. 3 is a flowchart illustrating an operation of the statistical language model generation device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1, 1A Statistical language model generation device 3 Past news manuscript accumulation means 5 Latest report manuscript accumulation means 7 Speech recognition means 9 Recognition result accumulation means 11, 11A Language model calculation means 13 Recognition result correction means 15 Correction recognition result accumulation means

 ──────────────────────────────────────────────────続 き Continuation of the front page (72) Inventor Akio Ando 1-10-11 Kinuta, Setagaya-ku, Tokyo Japan Broadcasting Research Institute F-Term (Reference) 5D015 HH00

Claims (3)

[Claims] 1. A statistical language model generating apparatus for generating a statistical language model, comprising: text data obtaining means for obtaining latest text data including a word whose appearance frequency is predicted to be high; Voice recognition means for recognizing, as speech, past large-volume text data having a larger data amount than that of the latest text data and the latest text data; and the latest text data, the past large-volume text data, and the voice recognition means. A storage unit for storing a recognition result obtained by the above-described method, and n-gr based on the past large amount of text data.
am, a second probability weight of n-gram is calculated based on the latest text data, and a third probability weight of n-gram is calculated based on the recognition result. Probability weight calculating means for calculating, and a language model generating means for generating a statistical language model based on the first probability weight, the second probability weight, and the third probability weight. A statistical language model generation device characterized by the following. 2. The apparatus according to claim 1, further comprising a recognition result correcting unit for correcting a recognition result recognized by the voice recognition unit, wherein the probability weight calculating unit calculates a third probability weight of n-gram based on the corrected recognition result. The statistical language model generator according to claim 1, wherein the statistical language model is calculated. 3. A computer, comprising: text data obtaining means for obtaining latest text data including a word whose appearance frequency is predicted to be high; a data amount larger than the latest text data and the latest text data; Voice recognition means for recognizing a large amount of past text data as voice, a storage means for storing the latest text data and the large volume of text data in the past, and a recognition result recognized by the voice recognition means; N-gr based on a large amount of text data
am, a second probability weight of n-gram is calculated based on the latest text data, and a third probability weight of n-gram is calculated based on the recognition result. Calculating a probability weight calculating means, based on the first probability weight, the second probability weight, and the third probability weight, a language model generating means for generating a statistical language model, A statistical language model generation program featuring the feature.