JP2002091484A - Language model generator and voice recognition device using the generator, language model generating method and voice recognition method using the method, computer readable recording medium which records language model generating program and computer readable recording medium which records voice recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a language model having high estimation precision and a voice recognition device having high recoginition precision. SOLUTION: A learning text data tree structure clustering means 2001 conducts a tree structure clustering to hierarchically divide learning text data 1001 so as to have a linguistically similar nature and generates a tree structure learning text data cluster 2002. A language model generating means 1004 generates a tree structure cluster language model 2003 employing each learning text data which belongs to the cluster 2002.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a language model generating apparatus to be referred to when performing speech recognition, a speech recognizing apparatus using the same, a language model generating method, a speech recognizing method using the same, and a language model generating program. And a computer-readable recording medium on which a voice recognition program is recorded.

[0002]

2. Description of the Related Art In recent years, the practical use of continuous speech recognition technology that enables a user to continuously input words has been actively studied. Continuous speech recognition is decoding based on an acoustic observation sequence of speech such that a decoded sequence of words has a maximum posterior probability. This is expressed by the following equation (1).

(Equation 1) Here, O is an acoustic observation value sequence [o ₁ , o ₂ ,
o ₃ ,. . . , O _T ], and W is the word string [w ₁ ,
w ₂ , w ₃ ,. . . , W _n ]. P (O | W) is a probability for the observed value series O when the word string W is given, and is calculated by an acoustic model. P
(W) is the occurrence probability (appearance probability) of the word string W, which is calculated by a language model.

[0003] Speech recognition is described in "Speech Information Processing" by Sadahiro Furui, published by Morikita Publishing Co., Ltd.
Reference 1), "Speech Recognition by Stochastic Model" by Seichi Nakagawa published by the Institute of Electronics, Information and Communication Engineers (hereinafter referred to as Reference 2), Lawrence Rabiner, published by NTT Advanced Technology Corporation, B
"Basics of Speech Recognition (upper and lower)", translated by Iing-Hwang Jung and translated by Sadahiro Furui (hereinafter referred to as Reference 3)
Is described in detail.

[0004] P (O | W) calculated by an acoustic model
Is based on the Hidden Markov Model (HM
M) has been actively studied. The acoustic model using the hidden Markov model is described in detail in, for example, Chapter 6 of Reference 3.

[0005] In addition, P calculated by a language model
(W) often uses a statistical method, and a typical example is an N-gram model (N is 2 or more). These are described in detail in Chapter 3 of “Probabilistic Language Model” (hereinafter referred to as Reference 4) by Kenji Kita published by The University of Tokyo Press. The N-gram model statistically gives a transition probability from the immediately preceding (N-1) words to the next word. Word strings w ^L ₁ = w ₁ . . . The occurrence probability of w _L is given by the following equation (2).

(Equation 2)

In the above equation (2), the probability P (w _t | w
_{t + 1-N} ^t-1 ) is a word string w composed of (N-1) words
_The probability that the word w _t will occur after _{t + 1−N} ^t−1 , and Π indicates the product. For example, "I, ha, station, go, go"
When the occurrence probability of a word string such as (• represents a word delimiter) is determined by 2-gram (bigram), the following equation (3) is used. In equation (3), # is the beginning of a sentence,
This is a symbol indicating the end of a sentence. P (I ・ Ha ・ Station ・ Go ・ Go) = P (I│ #) P (Ha│I) P (Station│Ha) P (To│Station) P (Go│To) P (# │Go) ( 3)

[0007] The probability _{P (w t | w t +} 1-N t-1) is determined by the relative frequency of the word sequence of the learning text data.
The appearance frequency of the word string W in the learning text data is represented by C
Assuming (W), for example, the 2-gram probability P (ha | I) of “I.ha” is calculated by the following equation (4).
In the equation (4), C (I. ha) is the frequency of appearance of the word string "I. ha", and C (I) is the frequency of appearance of "I". P (ha | me) = C (me ・ ha) / C (me) (4)

However, if the probability value of the N-gram model is simply estimated based on the relative frequency, there is a major drawback that a word set that does not appear in the learning text data is set to 0 (zero frequency problem). Further, even if it appears in the learning text data, it is difficult to estimate a statistically reliable probability value for a word string having a low appearance frequency (sparseness problem). To address these problems, a technique called smoothing or smoothing is usually used. Regarding smoothing, some methods are described in Chapter 3.3 of the above-mentioned Document 4, and a specific description is omitted here.

In learning a language model, sentences in fields, scenes, and situations to be subjected to speech recognition are used as learning text data. However, in an actual application, speech recognition is performed in various fields and in various scenes.・ It is often voice of situation. Since the probability of occurrence of word strings will be different for different fields, scenes and situations, if the language model is generated by learning the learning text data collectively and ignoring the differences in fields, scenes and situations However, the accuracy of the language model is not good.

[0010] In order to improve the performance of a speech recognition apparatus for speech recognition in various fields and various scenes and situations, text data for learning a language model is clustered and divided into clusters. A method of creating a language model for each is being studied. As a conventional technique, for example, there is “Statistical Language Model Generation Apparatus and Speech Recognition Apparatus” of JP-A-2000-75886 (hereinafter referred to as Document 5). Here, the cluster can be obtained by, for example, dividing the fields into categories such as cluster 1 for politics and cluster 2 for sports, or clustering sentences by defining the distance between sentences.

When the text data for learning is divided into clusters, the text data for learning per cluster is reduced, so that the above-mentioned problem of zero frequency and sparseness is further increased. In [5 contrast, the language model LM _a and, in another cluster estimated using training text data divided into clusters language model LM _c ^k estimated using the entire training text data is not divided into clusters (K is a cluster number), and a highly accurate language model is obtained by estimating LM _map ^k by a maximum posterior probability estimation method.

FIG. 13 is a block diagram showing a configuration of a conventional language model generation device described in Reference 5. In the figure, 1001 is text data for learning a language model, 1002 is texturing clustering means for learning, 1003 is a text data cluster for learning, 100
Reference numerals 3-1 to 1003-M denote learning text data of clusters 1 to M, 1004 denotes a language model generating unit, 1005 denotes a language model for each cluster, and 1005-1 to 1005-M denote language models of clusters 1 to M.

Next, the operation will be described. The learning text data 1001 is text data for learning a language model, and is obtained by converting words or sentences to be recognized by the speech recognition device into characters. The learning text data 1001 is input to the learning text data clustering means 1002.

The learning text data clustering means 1002 clusters the learning text data 1001. In Reference 5, texts are clustered on a sentence basis using a method similar to the k-means method. The points different from the ordinary k-means method are that (1) the cluster center vector is a language model generated by a sentence belonging to the cluster, and (2) the sentence generation probability is used as a distance scale. is there. The language model has N
-The gram model is used.

Learning text data cluster 1003
Is a learning text data clustering means 1002
Thus, the learning text data 1003-1 of the cluster 1 and the learning text data 1003-M of the cluster M are clustered into M clusters.

The language model generating means 1004 includes the learning text data 1003-1 to 1003-1 for the cluster 1 obtained by the learning text data clustering means 1002.
The learning text data 1003-M of the cluster M is input, and the language model 1005-1 of the cluster 1 is input.
A language model 1005 for each cluster composed of the language models 1005-M of the cluster M is generated. The language model generation unit 1004 is provided to prevent a decrease in the estimation accuracy of the language model due to a decrease in the number of learning text data for each cluster.
By using the language model LM _a estimated using the whole learning text data that does not cluster split, the cluster-specific language model LM _c ^k estimated using the learning text data that has been divided into clusters, the maximum a posteriori probability estimation method _{Is used} to estimate a language model LM _map ^k for each cluster.

Next, a conventional speech recognition apparatus using the above-described language model generation apparatus will be described. FIG. 14 is a block diagram showing a configuration of a conventional speech recognition device disclosed in Reference 5. In the figure, reference numeral 1101 denotes a recognition target voice; 1102, a voice feature amount extraction unit; 1103, an acoustic model;
Is a language model selecting means, 1105 is a matching means, 1106
Is the speech recognition result. Cluster-based language model 1005
Are the same functional blocks as those in FIG. 13, and are denoted by the same reference numerals and description thereof is omitted.

Next, the operation will be described. The recognition target voice 1101 is a voice to be recognized, and is input to the voice feature amount extraction unit 1102. Voice feature extraction means 110
2 extracts a speech feature amount included in the recognition target speech 1101. The acoustic model 1103 is a model for performing acoustic collation of voice. As the acoustic model 1103, for example, an HMM using a phoneme that recognizes a phoneme environment before and after, which is learned using sentences or words uttered by a large number of speakers, is used as a recognition unit.

The language model selecting means 1104 outputs the language model 1 of the cluster 1 generated using the language model generating device.
A language model to be used by the matching unit 1105 is selected from the cluster-specific language models 1005 composed of the language models 1005-M of the clusters 005-1 to M. In Reference 5, collation is performed using an unspecified language model before being divided into clusters.
One cluster-specific language model having the highest probability of occurrence is selected from the language model 1005-1 of the cluster 1 to the language model 1005-M of the cluster M.

The collating unit 1105 includes a recognition target word [W (1), W (2),..., W (w) set by the language model selected by the language model selecting unit 1104.
n)] (where wn is the number of words to be recognized) is converted into a recognition unit label notation, and according to this label, H in phoneme units stored in the acoustic model 1103 is converted.
The MMs are concatenated and the standard pattern of the word to be recognized [λ _{W (1)} ,
λ _{W (2)} ,. . . , Λ _{W ( wn)} ].

The collation means 1105 uses the standard pattern of the word to be recognized and the occurrence probability of the word string represented by the selected language model, and
The collation is performed on the speech feature amount output from 102, and a speech recognition result 1106 is output. Voice recognition result 1106
Is a word number sequence Rn = [r of the word having the highest matching score among the words to be recognized with respect to the voice to be recognized 1101.
(1), r (2),. . . , R (m)], and the word Rw = [W (r (1)), W (r
(2)),. . . , W (r (m))]. Here, r (i) indicates the word number of the i-th word in the word sequence of the speech recognition result 1106. M indicates the number of words in the recognized word sequence.

[0022]

Since the conventional language model generator is configured as described above, if the number of clusters to be divided by clustering increases, the number of text data for learning per cluster decreases, and the language model There is a problem that the accuracy of speech recognition does not increase because the estimation accuracy of the speech recognition becomes worse.

When the number of clusters to be divided increases,
If one utterance has the linguistic properties of multiple clusters,
There was a problem that the recognition rate did not increase.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and has a language model generating apparatus, a language model generating method, and a computer storing a language model generating program capable of generating a language model with high estimation accuracy. It is an object to obtain a readable recording medium.

It is another object of the present invention to provide a computer-readable recording medium storing a speech recognition device, a speech recognition method, and a speech recognition program with high speech recognition accuracy using a language model with high estimation accuracy. I do.

[0026]

A language model generating apparatus according to the present invention is configured to generate a language model for inputting learning text data and calculating the occurrence probability of a word string. A learning text data tree structure clustering means for performing a tree structure clustering for hierarchically dividing so as to have a property similar to each other and generating a tree structure learning text data cluster; and belonging to the tree structure learning text data cluster. Language model generating means for generating a language model for each tree structure cluster using each learning text data.

The language model generating apparatus according to the present invention performs an interpolation process using a tree structure cluster-based language model located above a tree structure in which the tree structure cluster-based language model is located, and performs the interpolation-processed tree structure cluster. It is provided with a language model interpolation means for generating another language model.

A speech recognition apparatus according to the present invention, wherein a speech to be recognized is inputted, speech recognition is performed, and a speech recognition result is output. And a tree model clustering that hierarchically divides the training text data into linguistically similar properties by performing an acoustic model that calculates the probability of the acoustic observation sequence of speech Model selecting means for selecting a language model having the highest probability of occurrence for a word string of a speech recognition result candidate from the language model for each tree structure cluster generated using the text data for speech and the language model for each tree structure cluster And using the language model selected by the language model selection means and the acoustic model, the speech feature quantity extracted by the speech feature quantity extraction means When the comparison is that a matching means for outputting a speech recognition result.

In the speech recognition apparatus according to the present invention, the language model selecting means selects a language model from the cluster-based language model of the lowest leaf node in the tree-structure cluster-based language model.

A speech recognition apparatus according to the present invention, wherein a speech to be recognized is inputted, speech recognition is performed, and a speech recognition result is output. And a tree model clustering that hierarchically divides the training text data into linguistically similar properties by performing an acoustic model that calculates the probability of the acoustic observation sequence of speech -Language model for each tree-structured cluster generated using text data for speech, and a multi-language model for selecting a plurality of language models having a high probability of occurrence for a word string of a speech recognition result candidate from the language model for each tree-structured cluster Selecting means, and a mixed language model generating means for generating a mixed language model by inputting a plurality of language models selected by the plurality of language model selecting means Using the language model generated by the mixed language model generating unit and the acoustic model, performing matching against the voice feature amount extracted by the voice feature amount extracting unit and outputting a voice recognition result. It is a thing.

In the speech recognition apparatus according to the present invention, the plurality of language models selecting means selects a plurality of language models from the cluster-based language model of the lowest leaf node in the tree-structure cluster-based language model.

In the speech recognition apparatus according to the present invention, the language model for each tree structure cluster is obtained by performing an interpolation process using a language model for each tree structure cluster located at a higher level of the tree structure. It is a language model.

A language model generating method according to the present invention is a method for generating a language model for inputting learning text data and calculating the probability of occurrence of a word string, wherein the learning text data has a linguistically similar property. The first step of generating a tree structure learning text data cluster by performing tree structure clustering that hierarchically divides into a tree structure, and using each learning text data belonging to the tree structure learning text data cluster to generate a tree And a second step of generating a language model for each structural cluster.

In the language model generating method according to the present invention, an interpolation process is performed by using a tree structure cluster-based language model located above a tree structure in which a tree structure cluster-based language model is located, and the interpolated tree structure cluster is obtained. And a third step of generating a different language model.

A speech recognition method according to the present invention is characterized in that a speech to be recognized is inputted, speech recognition is performed, and a speech recognition result is output. , Perform tree structure clustering that hierarchically divides the training text data so as to have linguistically similar properties, and from the language model for each tree structure cluster generated using the training text data of each tree structure cluster, A second step of selecting a language model having the highest occurrence probability for the word string of the speech recognition result candidate;
Speech recognition is performed by using the acoustic model for calculating the probability of the acoustic observation sequence of the speech and the language model selected in the second step to the speech features extracted in the first step. And a third step of outputting a result.

In the speech recognition method according to the present invention, in the second step, a language model is selected from the cluster-based language model of the lowest leaf node in the tree-structured cluster-based language model.

A speech recognition method according to the present invention performs a speech recognition in which a recognition target speech is input and outputs a speech recognition result, wherein a first step of inputting the recognition target speech and extracting a speech feature amount; Tree structure clustering is performed to divide the training text data hierarchically so as to have linguistically similar properties, and speech is generated from the language model for each tree structure cluster generated using the training text data for each tree structure cluster. A second step of selecting a plurality of language models having a high occurrence probability for the word string of the recognition result candidate; and a second step of inputting the plurality of language models selected in the second step to generate a mixed language model. Step 3, using the acoustic model for obtaining the probability of the acoustic observation sequence of the speech and the language model generated in the third step, It is obtained and a fourth step of outputting a speech recognition result collates the speech features extracted by the flop.

In the speech recognition method according to the present invention, in the second step, a plurality of language models are selected from the cluster-based language model of the lowest leaf node in the tree-structured cluster-based language model.

A computer-readable recording medium on which a language model generating program according to the present invention is recorded is for inputting learning text data and generating a language model for obtaining a word string occurrence probability. Tree clustering procedure for performing tree structure clustering that hierarchically divides text data for learning so as to have linguistically similar properties to generate text data clusters for tree structure learning, A language model generation procedure for generating a language model for each tree structure cluster using each learning text data belonging to the text data cluster.

A computer-readable recording medium having recorded thereon a language model generation program according to the present invention performs interpolation processing using a tree-structure cluster language model located above a tree structure in which a tree-structure cluster language model is located. This implements a language model interpolation procedure for generating a language model for each tree structure cluster subjected to interpolation processing.

A computer-readable recording medium on which a speech recognition program according to the present invention is recorded is for inputting a speech to be recognized and performing speech recognition and outputting a speech recognition result. Speech feature extraction procedure to extract the amount, and tree structure clustering that hierarchically divides the training text data so that it has linguistically similar properties, and generates using the training text data of each tree structure cluster Language model selection procedure to select the language model with the highest probability of occurrence for the word sequence of the speech recognition result candidate from the extracted tree model cluster-based language models, and an acoustic model to obtain the probability of the acoustic observation value sequence of the speech And using the language model selected in the language model selection procedure, the speech feature quantity extracted in the speech feature quantity extraction procedure Verification was carried out is used for realizing a matching procedure for outputting the result of speech recognition.

In the computer-readable recording medium storing the speech recognition program according to the present invention, the language model selecting step selects a language model from the cluster-based language model of the lowest leaf node in the tree-structure cluster-based language model. Things.

A computer-readable recording medium on which a speech recognition program according to the present invention is recorded is for inputting a speech to be recognized, performing speech recognition and outputting a speech recognition result. A speech feature extraction procedure for extracting speech features and a tree structure clustering that hierarchically divides the learning text data into linguistically similar properties are performed, and the training text data of each tree structure cluster is used. A multi-language model selecting step of selecting a plurality of language models having a high probability of occurrence for a word string of a speech recognition result candidate from the tree model cluster-based language model generated by A mixed-language model generation procedure for generating a mixed-language model by inputting multiple language models Using a language model generated by the mixed language model generation procedure, and using a language model generated by the mixed language model generation procedure to perform a verification on the voice features extracted by the voice feature extraction procedure and output a voice recognition result. Is realized.

In a computer-readable recording medium storing a speech recognition program according to the present invention, a plurality of language models are selected by selecting a plurality of language models from a cluster-based language model of the lowest leaf node in a tree-structured cluster-based language model. Is to select.

[0045]

DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a block diagram showing a configuration of a language model generating device according to Embodiment 1 of the present invention. In the figure, 2001 is a learning text data tree structure clustering means, 2002 is a tree structure learning text data cluster, 2002-1 to 2002-M are learning text data of tree structure clusters 1 to M, and 2003 is a tree structure cluster by cluster. Language models 2003-1 to 2003-M are language models of the tree-structure clusters 1 to M. The same functional blocks as those in FIG. 13 showing the configuration of the conventional language model generation device are denoted by the same reference numerals and description thereof will be omitted.

The language model learning text data 1001 is obtained by converting words and sentences used in a field, a scene, and a situation to be recognized by speech recognition into characters. For example, when a political news uttered by an announcer is targeted for speech recognition, it is text data in which an utterance content of a politics section of a newspaper or a politics news is transcribed as characters.

Next, the operation will be described. FIG. 2 is a flowchart showing a language model generation method in the language model generation device according to the first embodiment of the present invention. The text data tree structure clustering means 2001 for learning includes:
In step ST101, the learning text data 1
001 is input, and in step ST102, the hierarchical level I of clustering is set to 0. In step ST103, clustering is first performed on all the learning text data 1001. This text data for learning 1
The clustering for all 001 is referred to as the hierarchical 0 clustering.

Here, clustering refers to dividing manually into two or more fields, or using km-m shown in Reference 5.
This is to divide the learning text data into two or more sets using a method similar to the eans algorithm. The learning text data belonging to the cluster obtained by the clustering has linguistically similar properties.

FIG. 3 is an explanatory diagram of the learning text data tree structure clustering performed by the learning text data tree structure clustering means 2001, and shows a state in which sentences are hierarchically clustered in units. In FIG. 3, all of the learning text data 1001 is changed to 2 by clustering the tree structure cluster 00 of the hierarchy 0.
Divided into two clusters. The set of the divided learning text data is a tree-structure cluster 10 and a tree-structure cluster 11 of the hierarchy 1.

In step ST104 of FIG. 2, the learning text data tree structure clustering means 2001 is used.
Increments the hierarchy I, and in step ST105, the learning text data tree structure clustering means 2001 assigns the learning text data tree structure clustering unit 2001 to the learning text data belonging to each cluster clustered at the hierarchy I-1 (here, I = 0). Perform clustering. In FIG.
Of the tree structure cluster 10 of the hierarchy 1
Tree structure cluster 20 and tree structure cluster 21
Are generated, and a tree structure cluster 22 and a tree structure cluster 23 of the hierarchy 2 are generated from the tree structure cluster 11.

In step ST106, it is checked whether or not the number of clusters has reached a predetermined number M.
Otherwise, the process returns to step ST104, increments the hierarchy I, and repeats the clustering process of step ST105. The above processing is repeated until the number of clusters reaches a predetermined number M, and the tree structure cluster 1
Of the tree structure cluster M is generated.

After performing tree structure clustering of the text data for learning up to a predetermined number of clusters, in step ST107, the language model generation means 1004
A language model is generated for each of the clustered tree-structured clusters using learning text data belonging to each cluster, and a language model 2003-1 of the tree-structured cluster 1 is generated.
A tree model cluster-specific language model 2003 composed of the language model 2003-M of the tree structure cluster M is generated.

In step ST106, the hierarchical clustering of the learning text data is repeated until the number of clusters reaches a predetermined number M. Here, the number of clusters is used as a criterion for terminating clustering. However, if the number of text data for learning in a cluster is equal to or less than a certain value, clustering may be terminated based on the number of layers. In the cluster obtained by the hierarchical clustering, the properties of the learning text data belonging to the cluster at the lower level express the difference in the field, scene, and situation.

FIG. 4 is an explanatory diagram of language model generation for each tree structure cluster. In FIG. 4, the nodes of the tree structure represent the tree structure clusters of the learning text data, and a language model is generated for each tree structure cluster using the learning text data belonging to the cluster. The tree structure cluster of the parent node of the tree structure includes all the learning text data belonging to the tree structure cluster of the child node. In FIG.
For example, the language model generated using the learning text data belonging to the tree structure cluster 00 is the tree structure cluster 0
The language model generated using the language model LM00 of 0 and the text data for learning belonging to the tree structure cluster 10 is
It corresponds to the language model LM10 of the tree structure cluster 10, respectively.

The nature of the generated language model is such that the closer to the language model of the lower-level tree-structured cluster, the more the language model expresses differences in language properties due to differences in fields, scenes, and situations. In addition, the language model of the tree structure cluster in the upper layer does not finely express the difference in language properties due to differences in fields, scenes, and situations. It is a valid language model when it includes multiple fields, scenes and situations. Further, since the language model of the upper-layer tree-structured cluster has a large amount of learning text data, the parameter estimation accuracy is higher than that in a case where the number of clusters is equal to the number of tree-structured clusters at once.

A specific method of generating a language model is described in Reference 4.
, An N-gram model, a hidden Markov model, a stochastic context-free grammar, and the like described in Chapters 3 to 5.

Further, the language model generating method according to the first embodiment can be recorded on a recording medium as a language model generating program. In this case, it is composed of a learning text data tree structure clustering procedure for realizing the same processing as the learning text data tree structure clustering means 2001, and a language model generation procedure for realizing the same processing as the language model generating means 1004. The generated language model generation program is recorded on a recording medium.

As described above, according to the language model generating apparatus and the language model generating method of the first embodiment, the learning text data is hierarchically clustered into a tree structure, and the learning text data belonging to each tree structure cluster is clustered. Generates language models for each tree-structured cluster using, thus reducing the problem of zero frequency and sparseness of language models caused by a small amount of text data for learning, and the effect of generating language models with high recognition rates Is obtained. Even when one utterance to be recognized includes a plurality of fields, scenes, and situations, there is a language model that has learned the language features of the plurality of fields, scenes, and situations. An effect that can be generated is obtained.

Embodiment 2 FIG. 5 is a block diagram showing a configuration of a language model generating device according to Embodiment 2 of the present invention. In the figure, reference numeral 3001 denotes a language model interpolation unit;
Reference numeral 3002 denotes a language model for each tree structure cluster subjected to interpolation processing, and reference numerals 3002-1 to 3002-M denote language models of the tree structure clusters 1 to M subjected to interpolation processing. The same functional blocks as in FIG. 1 of the first embodiment are denoted by the same reference numerals, and description thereof is omitted.

Next, the operation will be described. FIG. 6 is a flowchart showing a language model generation method in the language model generation device according to the second embodiment of the present invention. The processing from step ST201 to step ST207 is the same as the processing from step ST101 to step ST107 in FIG. 2 of the first embodiment.

In step ST208, the language model interpolation means 3001 generates the language model 2003-1 of the tree structure cluster 1, which is the language model for each tree structure cluster generated by the language model generation means 1004, and the language model 2003 of the tree structure cluster M. -M is input to generate a language model 3002-1 of the tree structure cluster 1 subjected to the interpolation processing to a language model 3002-M of the tree structure cluster M subjected to the interpolation processing. In the interpolation processing at this time, the interpolation processing is performed using the language model of the tree structure cluster of the parent node of the tree structure node where the cluster language model to be interpolated is located.

In the example of FIG. 4, when the language model LM20 of the tree structure cluster 20 is interpolated, the language model LM10 of the tree structure cluster 10 which is the parent node and the language model LM20 of the tree structure cluster 00 which is a further upper parent node are interpolated. Model LM00
And interpolate using In this interpolation process, for example, when the language model is N-gram model, the probability of following the word sequence ^{_{w n + 1-N n-}} 1 occurring is w _n is a parameter, the following expression (5) Ask by.

(Equation 3)

In the above equation (5), P '_s(W_n| W
_{n + 1-N} ^n-1) Is the language model of the tree structure cluster S subjected to the interpolation processing.
The word sequence w in Dell_{n + 1-N} ^n-1Followed by w_nOccurs
Probability, Ω is the cluster of the tree structure cluster S and its parent node
Set of numbers, P_i(W_n| W _{n + 1-N} ^n-1) Is the tree structure class
Word sequence w in the language model of i_{n + 1-N} ^n-1Followed by w
_nProbability of occurrence of α_iIs a weight coefficient. This α
_iIs, for example, the deletion interpolation described in Chapter 3 of Reference 4.
It can be estimated by the method.

In this description, P _i (w _n | w _{n + 1−N} ⁿ⁻¹ )
Is the occurrence probability before interpolation, but the occurrence probability P ′ _i (w _n | w _{n + 1−N) obtained} by performing interpolation from the upper layer of the tree structure and performing interpolation processing
^n-1 ) may be used. In the tree-structured cluster, since the lower-level cluster has a small amount of text data for learning, a zero frequency problem and a problem of sparseness tend to occur in language model generation. using cluster language model, since the interpolation process of the probability of occurrence is w _n Following parameters: the word sequence ^{_{w n + 1-n n-}} 1, the language model estimation accuracy is increased.

Further, the language model generating method according to the second embodiment can be recorded on a recording medium as a language model generating program. In this case, a learning text data tree structure clustering procedure for realizing the same processing as the learning text data tree structure clustering means 2001, a language model generation procedure for realizing the same processing as the language model generating means 1004, Model interpolation means 300
A language model generation program including a language model interpolation procedure for realizing the same processing as in step 1 is recorded on a recording medium.

As described above, according to the language model generating apparatus and the language model generating method of the second embodiment, the learning text data is hierarchically clustered into a tree structure, and the learning text data belonging to each tree structure cluster is obtained. Is used to generate a language model for each tree structure cluster, and the generated cluster language model is interpolated using the cluster language model of the parent node of the tree structure. The effect of reducing the problem of zero frequency and the problem of sparseness can be obtained, and a language model with a high recognition rate can be generated.

Even if one utterance to be recognized includes a plurality of fields, scenes, and situations, there is a language model that has learned the language features of the plurality of fields, scenes, and situations, so that the recognition rate is high. The advantage is that a language model can be generated.

Embodiment 3 FIG. 7 is a block diagram showing a configuration of a voice recognition device according to Embodiment 3 of the present invention. In the figure, the same functional blocks as those in FIG. 1 of Embodiment 1 and FIG. 14 of the conventional speech recognition apparatus are denoted by the same reference numerals, and description thereof will be omitted.

Next, the operation will be described. FIG. 8 is a flowchart showing a voice recognition method in the voice recognition device according to the third embodiment of the present invention. The voice feature amount extraction unit 1102 inputs the recognition target voice 1101 in step ST301, and extracts a voice feature amount in step ST302. Here, the speech feature amount represents a feature of speech with a small amount of information, and is, for example, a cepstrum described in Chapter 5 of Document 1 and a feature vector composed of dynamic features of the cepstrum.

In step ST303, the language model selecting means 1104 converts the language model used by the matching means 1105 into the language model 2003-1 of the tree-structure cluster 1 and the language model 2003 of the tree-structure cluster M by the tree-structure cluster-based language model 2003. -Select one from M. The language model is selected by using, for example, the method described in Reference 5, and the language model of the tree-structure cluster having the highest occurrence probability is selected.

In step ST304, the matching means 1
Reference numeral 105 denotes a tree structure cluster language model selected by the language model selecting unit 1104 and an acoustic model 1103.
Is input to perform matching on the speech feature amount of the recognition target speech 1101, and a word string having the highest likelihood (matching score) is output as the speech recognition result 1106.

The collation processing in this case will be specifically described.
The matching unit 1105 outputs the recognition target words [W (1), W (2),...) Set by the tree-structure cluster language model selected by the language model selecting unit 1104. . . , W
(Wn)] (where wn is the number of words to be recognized) is converted into a recognition unit label notation, and HMMs of phoneme units stored in the acoustic model 1103 are connected according to the label, and Standard patterns of words [λ _{W (1)} , λ _{W (2)} ,. . . , Λ _{W (wn)} ].

The collation unit 1105 uses the recognition target word standard pattern and the occurrence probability of the word string represented by the selected tree-structured cluster language model to generate the speech feature amount output from the speech feature amount analysis unit 1102. And outputs a speech recognition result 1106. The speech recognition result 1106 is a word number sequence Rn = [r (1),
r (2),. . . , R (m)], and the word Rw = [W (r (1)), W (r
(2)),. . . , W (r (m))]. Here, r (i) indicates the word number of the i-th word in the word sequence of the speech recognition result, and m indicates the number of words in the recognized word sequence.

The language model for each tree structure cluster to be selected has been described above as the language model 2003-1 of the tree structure cluster 1 generated in the first embodiment and the language model 2003-M of the tree structure cluster M. Language Model 30 of Interpolated Tree Structure Cluster 1 Generated in Embodiment 2
The language model 3002-M of the tree structure cluster M subjected to the interpolation processing 02-1 to 02 may be used.

The voice recognition method according to the third embodiment can be recorded on a recording medium as a voice recognition program. In this case, in addition to the language model generation program of the first embodiment, a speech feature extraction procedure for realizing the same processing as that of the speech feature extraction unit 1102 and a process similar to the language model selection unit 1104 are realized. A speech recognition program including a language model selection procedure and a collation procedure for realizing the same processing as the collation unit 1105 is recorded on a recording medium.

As described above, according to the speech recognition apparatus and the speech recognition method in the third embodiment, the learning text data 1001 is hierarchically clustered in a tree structure.
Learning text data 200 belonging to each tree structure cluster
Since the language model for each tree structure cluster 2003-1 to 2003-M is generated using 2-1 to 2002-M, the problem of zero frequency and sparseness of the language model caused by a small amount of text data for learning is generated. Can be reduced, and a language model is selected from the tree model cluster-based language model 2003 to perform speech recognition. Therefore, an effect that speech recognition with high recognition accuracy can be performed can be obtained.

Even when the speech to be recognized includes a plurality of fields, scenes, and situations, a tree-structure cluster language model that has learned the language features of the plurality of fields, scenes, and situations is selected and speech recognition is performed. Therefore, an effect of enabling speech recognition with high recognition performance can be obtained.

Embodiment 4 FIG. 9 is a block diagram showing a configuration of a speech recognition apparatus according to Embodiment 4 of the present invention. In the figure, reference numeral 5001 denotes a multiple language model selecting unit;
Reference numeral 5002 denotes a mixed language model generation unit. The same functional blocks as in FIG. 7 of the third embodiment are denoted by the same reference numerals, and description thereof is omitted.

Next, the operation will be described. FIG. 10 is a flowchart showing a voice recognition method in the voice recognition device according to Embodiment 4 of the present invention. Step ST40
1 and step ST402 correspond to steps ST301 and ST302 in FIG.
Is the same as the processing of

In step ST403, the multiple language model selecting means 5001 selects the language model 2003-1 of the tree structure cluster 1 to the language model 200 of the tree structure cluster M.
Select language models of two or more (K or less) tree-structure clusters from 3-M. The selection of the language model uses, for example, a method that extends the method described in Document 5 and selects K language models in descending order of occurrence probability.

In step ST404, the mixed language model generating means 5002 sets the plural language model selecting means 50
01, a plurality of tree-structure cluster language models selected as input, and one mixed language model is generated. If the mixed model is, for example, an N-gram model, the occurrence probability is calculated by the following equation (6).

(Equation 4)

In the above equation (6), P _m (w _n | w
_{n + 1-N} ^n-1 ) is the occurrence probability of the mixed language model, Ψ is the set of numbers of the tree-structured cluster language model selected by the multiple language model selecting means 5001, and P _i (w _n | w)
_{n + 1-N} ^n-1 ) is the probability of occurrence of the selected language model,
β _i is a weight coefficient. Here, with respect to β _i , for example, a language model with a high occurrence probability is set to have a large β _{i in} accordance with the occurrence probability at the time of selecting a language model shown in Document 5.

In step ST405, the matching means 1
105, the mixed language model generated by the mixed language model generating means 5002 and the acoustic model 1103 are input, and a collation is performed on the speech feature amount of the recognition target speech 1101, and a word string having the highest likelihood is subjected to speech recognition. The result is output as 1106.

In the above, the tree structure cluster language model to be selected has been described as the language model 2003-1 of the tree structure cluster 1 generated in the first embodiment to the language model 2003-M of the tree structure cluster M. Language Model 300 of Interpolated Tree Structure Cluster 1 Generated in Form 2
2-1 may be a language model 3002-M of the tree-structured cluster M subjected to the interpolation processing.

Further, the voice recognition method according to the fourth embodiment can be recorded on a recording medium as a voice recognition program. In this case, in addition to the language model generation program of the first embodiment, a speech feature amount extraction procedure for realizing the same processing as the speech feature amount extraction unit 1102, and a matching unit 1
105, a multilingual model selecting procedure for realizing the same processing as the multilingual model selecting means 5001, and a mixed language model generating procedure for realizing the same processing as the mixed language model generating means 5002 Is recorded on a recording medium.

As described above, according to the speech recognition apparatus and the speech recognition method in the fourth embodiment, the learning text data 1001 is hierarchically clustered in a tree structure.
Learning text data 2002-1 for each tree structure cluster
-2002-M, the tree-structure cluster-based language models 2003-1 to 2003-M are generated, and the zero frequency problem and the sparseness problem of the language model caused by a small amount of text data for learning can be reduced. Since a mixed language model is generated by using a plurality of tree-structure cluster language models selected from the tree-structure cluster language model 2003 and is used for speech recognition, an effect that speech recognition with higher recognition accuracy can be performed can be obtained.

Even when one utterance to be recognized includes a plurality of fields, scenes, and situations, a language model that has learned language features of a plurality of fields, scenes, and situations is selected to generate a mixed language model. Therefore, the present invention has the effect of enabling speech recognition with high recognition performance.

Embodiment 5 FIG. 11 is a block diagram showing a configuration of a speech recognition apparatus according to Embodiment 5 of the present invention. In the figure, 6001 is a language model for each leaf node cluster, and 6001-1 to 6001-L are language models for leaf node clusters 1 to L. The same functional blocks as in FIG. 7 of the third embodiment are denoted by the same reference numerals, and description thereof is omitted.

Next, the operation will be described. FIG. 12 is a flowchart showing a voice recognition method in the voice recognition device according to the fifth embodiment of the present invention. Step ST50
1 and step ST502 correspond to steps ST301 and ST302 in FIG.
Is the same as the processing of

In step ST503, the language model selecting means 1104 converts the language model used by the matching means 1105 from the language model of the leaf node cluster of the tree structure cluster into the language model 6001-1 of the leaf node cluster 1.
~ 1 from language model 6001-L of leaf node cluster L
Choose one. Here, the language model of the leaf node cluster is the language model of the lowest tree structure cluster of the tree structure. In the example of FIG. 4, the language model LM20 of the tree structure cluster 20, the language model LM21 of the tree structure cluster 21,
The language model LM22 of the tree structure cluster 22 and the language model LM23 of the tree structure cluster 23 correspond to the language model of the leaf node cluster.

The language model of such a leaf node cluster is a model that expresses in detail the differences in the language properties due to the differences in fields, scenes, and situations.
This is effective when the recognition target speech is such that the situation is clearly separated. Further, the number of cluster language models to be selected is smaller than in the case where language models for all tree-structured clusters are used, so that there is an effect of saving memory and reducing the amount of calculation. The language model of the leaf node cluster is selected by using, for example, the method described in Reference 5, and the language model of the leaf node cluster having the highest occurrence probability is selected.

In step ST504, the matching means 1
Reference numeral 105 denotes a language model of the leaf node cluster selected by the language model selecting unit 1104 and the acoustic model 11
03 is input, the speech feature amount of the recognition target speech 1101 is collated, and the word string having the highest likelihood is output as the speech recognition result 1106.

In the above description, the language model of the leaf node cluster to be selected is the language model of the leaf node cluster generated by the tree structure cluster 2003 generated in the first embodiment. A language model of the leaf node cluster of the processed tree structure cluster language model 3002 may be used. Language model selection means 11
04 may be a multiple language model selecting means 5001, and a mixed language model generating means 5002 may be connected at a later stage to perform the collation processing using the mixed language model.

Further, the voice recognition method according to Embodiment 5 can be recorded on a recording medium as a voice recognition program. In this case, in addition to the language model generation program of the first embodiment, a speech feature extraction procedure for realizing the same processing as that of the speech feature extraction unit 1102 and a process similar to the language model selection unit 1104 are realized. A speech recognition program including a language model selection procedure and a collation procedure for realizing the same processing as the collation unit 1105 is recorded on a recording medium.

As described above, according to the speech recognition apparatus and the speech recognition method in the fifth embodiment, the learning text data 1001 is hierarchically clustered in a tree structure.
Learning text data 2002-1 for each tree structure cluster
-2002-M, the tree structure cluster language model 2
Since 003 is generated, the zero frequency problem and the sparseness problem of the language model caused by a small amount of learning text data can be reduced, and the language model 600 of the leaf node cluster of the tree structure cluster language model 2003 can be reduced.
Since the language model selected from No. 1 is used for speech recognition, speech recognition with high recognition accuracy can be performed, the memory capacity of the language model can be reduced, and the amount of calculation when selecting the language model can be reduced.

Even if one utterance to be recognized includes a plurality of fields, scenes, and situations, if a language model of a plurality of leaf node clusters is selected and a mixed language model is generated,
Since a language model that has learned language features of a plurality of fields, scenes, and situations is used for speech recognition, an effect of enabling speech recognition with high recognition performance can be obtained.

[0097]

As described above, according to the present invention, the language model generating apparatus performs tree structure clustering for hierarchically dividing the learning text data so as to have linguistically similar properties. A learning text data tree structure clustering means for generating a learning text data cluster; and a language model generating means for generating a language model for each tree structure cluster using each learning text data belonging to the tree structure learning text data cluster. , The problem of zero frequency and sparseness of the language model caused by a small amount of text data for learning can be reduced, a language model with a high recognition rate can be generated, and one utterance of the recognition target Even if it includes fields, scenes, and situations, language models that have learned the language features of multiple fields, scenes, and situations Since Le is present, the effect of high recognition rate language model can be generated.

According to the present invention, the language model generating apparatus performs an interpolation process using the tree-structure-cluster-based language model located above the tree structure in which the tree-structure-cluster-based language model is located. By providing a language model interpolation means for generating a language model for each structural cluster,
It can reduce the problem of zero frequency and sparseness of language models caused by a small amount of text data for learning, can generate a language model with high recognition rate, and can recognize one utterance in multiple fields, scenes and situations. Is included, there is a language model that has learned the language features of a plurality of fields and scenes / situations, so that there is an effect that a language model with a high recognition rate can be generated.

According to the present invention, the speech recognition apparatus performs tree structure clustering for hierarchically dividing the speech feature amount extracting means, the acoustic model, and the learning text data so as to have linguistically similar properties. From the language model for each tree structure cluster generated using the training text data of each tree structure cluster and the language model for each tree structure cluster, the language model with the highest probability of occurrence for the word string of the speech recognition result candidate is A language model selecting means for selecting, and a matching means for performing matching with respect to the speech feature quantity extracted by the speech feature quantity extracting means using the selected language model and acoustic model and outputting a speech recognition result. Can reduce the problem of zero frequency and sparseness of the language model caused by a small amount of text data for training. Since speech recognition is performed by selecting a language model from Dell, high-accuracy speech recognition can be performed, and even when the recognition target speech includes multiple fields, scenes, and situations, multiple fields, scenes, and situations are recognized. Since the tree structure cluster language model that has learned the language features of the situation is selected and speech recognition is performed, there is an effect that speech recognition with high recognition performance can be performed.

According to the present invention, the language model selecting means of the speech recognition apparatus selects the language model from the cluster-based language model of the lowest leaf node in the tree-structure cluster-based language model, so that the memory capacity of the language model is And the amount of calculation when selecting a language model can be reduced.

According to the present invention, the speech recognition device performs tree structure clustering for hierarchically dividing the speech feature amount extracting means, the acoustic model, and the learning text data into linguistically similar properties. From the language model for each tree structure cluster generated using the training text data of each tree structure cluster and the language model for each tree structure cluster, a plurality of language models with a high probability of occurrence for the word sequence of the speech recognition result candidate Multi-language model selecting means for selecting a language model, a mixed language model generating means for generating a mixed language model by inputting a plurality of selected language models, and speech feature extraction using the generated language model and acoustic model Means for collating the speech features extracted by the means and outputting a speech recognition result, so that the amount of text data for learning is small. It can reduce the problem of zero frequency and sparseness of the language model caused by the language model, and generate a mixed language model by using a tree structure cluster language model selected from multiple language models for each tree structure cluster, and use it for speech recognition. In addition to high speech recognition, even if one utterance to be recognized includes multiple fields, scenes, and situations, a language model that has learned the language features of multiple fields, scenes, and situations is selected and a mixed language model is created. Since it is generated and used for speech recognition, there is an effect that speech recognition with high recognition performance can be performed.

According to the present invention, the plurality of language models selecting means of the speech recognition apparatus selects a plurality of language models from the cluster-based language models of the lowest leaf node in the tree-structure cluster-based language model, whereby the language model is selected. This has the effect of reducing the memory capacity of the device and the amount of calculation when selecting a language model.

According to the present invention, the language model for each tree structure cluster in the speech recognition apparatus is an interpolated tree structure cluster obtained by performing an interpolation process using the language model for each tree structure cluster positioned at the top of the tree structure. By using a different language model, the problem of zero frequency and sparseness of the language model caused by a small amount of training text data can be reduced, a language model with a higher recognition rate can be generated, and one utterance of the recognition target can be generated. However, even if includes a plurality of fields, scenes, and situations, there is a language model that has learned the linguistic features of the plurality of fields, scenes, and situations, so that a language model with a high recognition rate can be generated.

According to the present invention, as a language model generation method, tree structure clustering for hierarchically dividing learning text data so as to have linguistically similar properties is performed to generate a tree structure learning text data cluster. And a second step of generating a language model for each tree structure cluster using each text data for learning belonging to the text data cluster for tree structure learning. It is possible to reduce the problem of zero frequency and sparseness of the language model caused by a small amount, to generate a language model with a high recognition rate, and to have one utterance to be recognized includes multiple fields, scenes and situations. However, there are language models that have learned the language features of multiple fields and scenes / situations. There can be effectively.

According to the present invention, as a language model generation method, an interpolation process is performed by using a tree structure cluster-based language model located above a tree structure in which a tree structure cluster-based language model is located, and the interpolated tree The provision of the third step of generating the language model for each structural cluster can reduce the zero frequency problem and the sparseness problem of the language model caused by a small amount of text data for learning, and can further reduce the language model having a high recognition rate. Can be generated, and even when one utterance to be recognized includes a plurality of fields, scenes, and situations, there is a language model that has learned the language features of the plurality of fields, scenes, and situations. There is an effect that a language model can be generated.

According to the present invention, as a voice recognition method,
A first step of extracting speech features and a tree structure clustering that hierarchically divides the learning text data so as to have linguistically similar properties are performed, and using the learning text data of each tree structure cluster. A second step of selecting a language model having the highest occurrence probability for the word string of the speech recognition result candidate from the generated language model for each tree structure cluster, and using the acoustic model and the selected language model, A third step of collating feature amounts and outputting a speech recognition result can reduce a zero frequency problem and a sparseness problem of a language model caused by a small amount of learning text data, Speech recognition is performed by selecting a language model from the language model for each tree-structured cluster. Even if the data includes multiple fields, scenes, and situations, it selects a tree-structured cluster language model that has learned the language features of multiple fields, scenes, and situations, and performs speech recognition. There is an effect that can be done.

According to the present invention, in the second step of the speech recognition method, the language model is selected by selecting the language model from the cluster-based language model of the lowest leaf node in the tree-structure cluster-based language model. This has the effect of reducing the amount of computation at the time.

According to the present invention, as a speech recognition method,
A first step of extracting speech features and a tree structure clustering that hierarchically divides the learning text data so as to have linguistically similar properties are performed, and using the learning text data of each tree structure cluster. A second method of selecting a plurality of language models having a high probability of occurrence with respect to the word string of the speech recognition result candidate from the generated language model for each tree structure cluster;
And a third step of generating a mixed language model by inputting a plurality of selected language models, and performing matching on the extracted speech feature using the acoustic model and the generated language model. And the fourth step of outputting a speech recognition result can reduce a zero frequency problem and a sparseness problem of the language model caused by a small amount of text data for learning. Since a mixed language model is generated using the selected tree-structure cluster language model and used for speech recognition, speech recognition with higher recognition accuracy can be performed.
Even if one utterance to be recognized includes a plurality of fields, scenes, and situations, a language model that has learned language features of the plurality of fields, scenes, and situations is selected, and a mixed language model is generated and used for speech recognition. Therefore, there is an effect that voice recognition with high recognition performance can be performed.

According to the present invention, in the second step of the speech recognition method, the language model is selected by selecting a plurality of language models from the cluster-based language model of the lowest leaf node in the tree-structure cluster-based language model. This has the effect of reducing the amount of calculation when selecting.

According to the present invention, tree structure clustering for hierarchically dividing learning text data so as to have linguistically similar properties is performed on a recording medium storing a language model generation program. A learning text data tree structure clustering procedure for generating text data clusters and a language model generation procedure for generating a tree model cluster-specific language model using each learning text data belonging to the tree structure learning text data cluster are realized. By doing so, the problem of zero frequency and sparseness of the language model caused by a small amount of text data for learning can be reduced, and a language model with a high recognition rate can be generated. Learn the linguistic features of multiple disciplines and scenes / situations, even if they include scenes / situations. Since the language model is present, the effect of high recognition rate language model can be generated.

According to the present invention, an interpolation process is performed on a recording medium on which a language model generation program is recorded, using a tree structure cluster language model located above a tree structure where the tree structure cluster language model is located. By implementing a language model interpolation procedure that generates an interpolated tree-structured cluster-based language model, it is possible to reduce the zero frequency problem and sparseness problem of the language model caused by a small amount of text data for learning, and further recognize In addition to generating a language model with a high rate, even if one utterance to be recognized includes multiple fields, scenes, and situations, there is a language model that has learned the language features of multiple fields, scenes, and situations. This has the effect that a language model with a high recognition rate can be generated.

According to the present invention, on the recording medium on which the speech recognition program is recorded, the speech feature amount extracting procedure for extracting the speech feature amount and the learning text data are hierarchically arranged so as to have linguistically similar properties. Perform a tree structure clustering to divide, and select the language model with the highest probability of occurrence for the word string of the speech recognition result candidate from the language model for each tree structure cluster generated using the training text data of each tree structure cluster The learning text data is realized by realizing a language model selection procedure to perform matching, and a matching procedure of performing matching on the extracted speech features and outputting a speech recognition result using the acoustic model and the selected language model. Can reduce the problem of zero frequency and sparseness of the language model caused by a small amount of language. Selects Dell and performs speech recognition, enabling high-accuracy speech recognition.Also, even when the speech to be recognized includes multiple fields, scenes, and situations, the language of multiple fields, scenes, and situations is used. Since speech recognition is performed by selecting a tree-structured cluster language model whose features have been learned, there is an effect that speech recognition with high recognition performance can be performed.

According to the present invention, the language model is selected by selecting the language model from the cluster-based language model of the lowest leaf node in the tree-structure cluster-based language model in the language model selection procedure of the speech recognition program. In this case, the effect of reducing the amount of calculation can be obtained.

According to the present invention, on a recording medium on which a speech recognition program is recorded, a speech feature amount extraction procedure for extracting a speech feature amount and a learning text data are hierarchically arranged so as to have linguistically similar properties. Performing tree structure clustering to divide, and from the language model for each tree structure cluster generated using the text data for learning of each tree structure cluster, a plurality of language models with high probability of occurrence for the word sequence of the speech recognition result candidate A multi-language model selection procedure to select, a mixed-language model generation procedure to generate a mixed-language model by inputting a plurality of selected language models, an acoustic model, and a speech extracted using the generated language model By implementing the matching procedure that performs matching on the feature amount and outputs the speech recognition result, the small amount of training text data The problem of zero frequency and sparseness of language models can be reduced, and a mixed language model is generated using a tree-structured cluster language model selected from multiple language models for each tree-structured cluster and used for speech recognition. Generates a mixed language model by selecting a language model that has learned the language features of multiple fields, scenes, and situations, even if one utterance to be recognized includes multiple fields, scenes, and situations Since it is used for speech recognition, there is an effect that speech recognition with high recognition performance can be performed.

According to the present invention, the step of selecting a plurality of language models in the speech recognition program includes the step of selecting a plurality of language models from the cluster-based language models of the lowest leaf node in the tree-structure cluster-based language model. The effect of reducing the amount of computation when selecting is obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a language model generation device according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing a language model generation method in the language model generation device according to the first embodiment of the present invention.

FIG. 3 is an explanatory diagram of learning text data tree structure clustering according to the first embodiment of the present invention;

FIG. 4 is an explanatory diagram of generation of a language model for each tree structure cluster according to the first embodiment of the present invention;

FIG. 5 is a block diagram showing a configuration of a language model generation device according to a second embodiment of the present invention.

FIG. 6 is a flowchart showing a language model generation method in the language model generation device according to the second embodiment of the present invention.

FIG. 7 is a block diagram showing a configuration of a voice recognition device according to a third embodiment of the present invention.

FIG. 8 is a flowchart showing a voice recognition method in a voice recognition device according to Embodiment 3 of the present invention.

FIG. 9 is a block diagram showing a configuration of a voice recognition device according to a fourth embodiment of the present invention.

FIG. 10 is a flowchart showing a voice recognition method in a voice recognition device according to Embodiment 4 of the present invention.

FIG. 11 is a block diagram showing a configuration of a voice recognition device according to a fifth embodiment of the present invention.

FIG. 12 is a flowchart showing a voice recognition method in a voice recognition device according to Embodiment 5 of the present invention.

FIG. 13 is a block diagram illustrating a configuration of a conventional language model generation device.

FIG. 14 is a block diagram illustrating a configuration of a conventional voice recognition device.

[Explanation of symbols]

1001 text data for learning, 1004 language model generating means, 1101 recognition target speech, 1102 voice feature quantity extracting means, 1103 acoustic model, 1104 language model selecting means, 1105 collating means, 1106 voice recognition result, 2001 text data tree structure for learning Clustering means, 2002 tree structure learning text data cluster, 2002-1 tree structure cluster 1 learning text data, 2002-2 tree structure cluster 2 learning text data, 2002-M tree structure cluster M
Learning text data of 2003, a language model by tree structure cluster 2003, a language model of 2003-1 tree structure cluster 1, a language model of 2003-2 tree structure cluster 2,
2003-M Language model of tree-structured cluster M, 300
1 Language Model Interpolating Means, 3002 Interpolated Tree Model by Tree Structure Cluster, 3002-1 Interpolated Tree Model of Tree Structure Cluster 1, 3002-2 Interpolated Tree Model of Tree Structure Cluster 2, 3002
-M a language model of the interpolated tree-structured cluster M,
5001 plural language model selection means, 5002 mixed language model generation means, 6001 language model for each leaf node cluster, 6001-1 language model for leaf node cluster 1, 6001-2 language model for leaf node cluster 2, 6001-L leaf node Language model of cluster L.

 ──────────────────────────────────────────────────続 き Continuation of the front page (54) [Title of the Invention] A language model generation device and a speech recognition device using the same, a language model generation method and a speech recognition method using the same, and a computer recording a language model generation program Readable recording medium and computer readable recording medium on which voice recognition program is recorded

Claims (19)

[Claims] 1. A language model generating apparatus for inputting learning text data and generating a language model for obtaining an occurrence probability of a word string, wherein the learning text data is hierarchically structured so as to have linguistically similar properties. A learning text data tree structure clustering means for performing tree structure clustering for generating a tree structure learning text data cluster, and a tree structure using each learning text data belonging to the tree structure learning text data cluster. A language model generating apparatus, comprising: a language model generating means for generating a cluster-specific language model. 2. A language model for performing an interpolation process using a tree structure cluster-based language model located above a tree structure in which a tree structure cluster-based language model is located, and generating an interpolated tree structure cluster-based language model. 2. The language model generating apparatus according to claim 1, further comprising an interpolation unit. 3. A speech recognition apparatus for inputting a speech to be recognized, performing speech recognition and outputting a speech recognition result, comprising: a speech feature amount extraction unit for inputting the speech to be recognized and extracting a speech feature amount; Model that calculates the probability of a series of observed values and tree structure clustering that hierarchically divides the training text data into linguistically similar properties, and uses the training text data of each tree structure cluster. A language model having a highest probability of occurrence for a word string of a speech recognition result candidate from the language model having a tree structure cluster generated by the above-mentioned language model; Using the language model selected by the selection means and the acoustic model, the speech feature quantity extracted by the speech feature quantity extraction means is collated to perform speech recognition. Speech recognition apparatus characterized by comprising a verification means for outputting a result. 4. The speech recognition apparatus according to claim 3, wherein the language model selecting means selects a language model from a cluster-based language model of a leaf node at the lowest layer in the tree-structure cluster-based language model. 5. A voice recognition device for inputting a voice to be recognized, performing voice recognition and outputting a voice recognition result, wherein: a voice feature amount extraction means for inputting the voice to be recognized and extracting a voice feature amount; Model that calculates the probability of a series of observed values and tree structure clustering that hierarchically divides the training text data into linguistically similar properties, and uses the training text data of each tree structure cluster. A multi-language model selecting means for selecting a plurality of language models having a high probability of occurrence for a word sequence of a speech recognition result candidate from the tree model cluster-based language model generated by the above-mentioned language model; A mixed language model generating means for generating a mixed language model by inputting a plurality of language models selected by the plurality of language model selecting means; Using the language model generated by the Dell generating unit and the acoustic model, performing matching against the voice feature amount extracted by the voice feature amount extracting unit and outputting a voice recognition result. Speech recognition device. 6. The speech recognition apparatus according to claim 5, wherein said plurality of language model selecting means selects a plurality of language models from cluster-based language models of the lowest leaf node in the tree-structure cluster-based language model. . 7. The tree-structure-cluster language model is an interpolated tree-structure-cluster language model obtained by performing an interpolation process using a tree-structure-cluster language model located at a higher level of the tree structure. Claim 3 or Claim 5
The speech recognition device according to the above. 8. A language model generating method for inputting learning text data and generating a language model for obtaining an occurrence probability of a word string, wherein the learning text data is hierarchically structured so as to have linguistically similar properties. A first step of generating a tree structure learning text data cluster by performing tree structure clustering for dividing into tree structures, and using each of the learning text data belonging to the tree structure learning text data cluster, by using a tree structure cluster-specific language model. And a second step of generating a language model. 9. A third step of performing an interpolation process using a tree structure cluster-based language model located above the tree structure in which the tree structure cluster-based language model is located, and generating an interpolated tree structure cluster-based language model. 9. The language model generating method according to claim 8, further comprising the step of: 10. A voice recognition method for inputting a voice to be recognized, performing voice recognition and outputting a voice recognition result, wherein: a first step of inputting the voice to be recognized and extracting a voice feature amount; Tree-structure clustering, which hierarchically divides the tree structure so that it has linguistically similar properties, and uses the tree-structure-cluster-based language model generated using the training text data of each tree-structure cluster A second step of selecting a language model having the highest probability of occurrence for the word string; an acoustic model for obtaining a probability of an acoustic observation value sequence of speech; and a language model selected in the second step. And a third step of comparing the voice feature quantity extracted in the first step and outputting a voice recognition result. 11. The speech recognition method according to claim 10, wherein in the second step, a language model is selected from the cluster-based language model of the lowest leaf node in the tree-structure cluster-based language model. 12. A voice recognition method for performing voice recognition by inputting a voice to be recognized and outputting a voice recognition result, comprising: a first step of inputting the voice to be recognized and extracting a voice feature amount; Tree structure clustering is performed by hierarchically dividing the tree so as to have linguistically similar properties. From the language model for each tree structure cluster generated using the training text data of each tree structure cluster, the words of speech recognition result candidates are A second step of selecting a plurality of language models having a high probability of occurrence for the column, a third step of inputting the plurality of language models selected in the second step and generating a mixed language model, The sound extracted in the first step using the acoustic model for obtaining the probability of the acoustic observation sequence of the speech and the language model generated in the third step. Speech recognition method characterized by comprising a fourth step of outputting a speech recognition result collates the feature quantity. 13. The speech recognition method according to claim 12, wherein in the second step, a plurality of language models are selected from the cluster-based language models of the lowest leaf node in the tree-structure cluster-based language model. 14. A recording medium recording a language model generating program for inputting learning text data and generating a language model for obtaining an occurrence probability of a word string, wherein the learning text data is linguistically similar. A learning text data tree structure clustering procedure for performing tree structure clustering that hierarchically divides to have a property and generating a tree structure learning text data cluster; and a learning text belonging to the tree structure learning text data cluster. A computer-readable storage medium storing a language model generation program for realizing a language model generation procedure for generating a tree model cluster-based language model using data. 15. A language model for performing an interpolation process using a tree structure cluster-based language model positioned above a tree structure in which a tree structure cluster-based language model is located, and generating an interpolated tree structure cluster-based language model. 15. A computer-readable recording medium on which the language model generating program according to claim 14 for realizing an interpolation procedure. 16. A recording medium storing a speech recognition program for inputting a speech to be recognized and performing speech recognition and outputting a speech recognition result, wherein the speech feature quantity for inputting the speech to be recognized and extracting a speech feature quantity. An extraction procedure and tree-structure clustering that hierarchically divides the learning text data into linguistically similar properties, and a tree-structure cluster language generated using the learning text data of each tree-structure cluster A language model selection procedure for selecting a language model having the highest probability of occurrence for a word sequence of a speech recognition result candidate from a model; an acoustic model for obtaining a probability of an acoustic observation sequence of speech; Collating using the language model selected by the above with respect to the speech feature extracted by the above speech feature extracting procedure and outputting a speech recognition result. And a computer-readable recording medium on which a voice recognition program for realizing the above procedure is recorded. 17. The recorded speech recognition program according to claim 16, wherein the language model selecting step selects a language model from a cluster language model of a leaf node at the lowest layer in the tree structure cluster language model. Computer readable recording medium. 18. A recording medium on which a speech recognition program for inputting a speech to be recognized and performing speech recognition and outputting a speech recognition result is recorded, wherein a speech feature quantity for inputting the speech to be recognized and extracting a speech feature quantity. An extraction procedure and tree-structure clustering that hierarchically divides the learning text data into linguistically similar properties, and a tree-structure cluster language generated using the learning text data of each tree-structure cluster A multi-language model selection procedure for selecting a plurality of language models having a high probability of occurrence for a word string of a speech recognition result candidate from a model; A mixed language model generation procedure for generating a language model, an acoustic model for determining the probability of an acoustic observation sequence of speech, and the mixed language model generation Computer-readable recording of a speech recognition program for realizing a collation procedure of outputting a speech recognition result by performing collation on speech features extracted by the speech feature extraction procedure using the language model generated in order. Possible recording medium. 19. The speech recognition program according to claim 18, wherein the step of selecting a plurality of language models selects a plurality of language models from cluster-based language models of the lowest leaf node in the tree-structure cluster-based language model. A computer-readable recording medium on which is recorded.