WO2010100853A1 - Language model adaptation device, speech recognition device, language model adaptation method, and computer-readable recording medium - Google Patents

Abstract

A language model adaptation device (100) which adapts a base language model. The language model adaptation device (100) is provided with: a segmentation unit (105) which segments an inputted text into a plurality of segments; a topic analysis unit (107) which identifies topics contained in the respective segments and creates, on the basis of each identified topic, topic models expressing the appearance tendency of words in the topics; and an adapted language model creation unit (110) which adapts the base language model on a segment-by-segment basis using the topic model corresponding to the topics contained in the segment, thereby creating an adapted language model.

Description

Language model adaptation apparatus, speech recognition apparatus, language model adaptation method, and computer-readable recording medium

The present invention relates to a language model adaptation device, a speech recognition device using the same, a language model adaptation method, and a language model adaptation method for adapting a language model used in speech recognition according to a recognition result of a recognition target speech. The present invention relates to a computer-readable recording medium on which a program is recorded.

In the speech recognition process that converts the content of utterances contained in speech data into text, a statistical language model that gives the generation probability of word strings is used. The most widely used statistical language model in speech recognition is the N-gram language model. The N-gram language model is a probability model that assumes that the word generation probability at a certain time depends only on the immediately preceding N-1 words.

That is, the generation probability of the i-th word w _i is given by P (w _i | w _{i−N + 1} ⁱ⁻¹ ). Here, w _{i−N + 1} ⁱ⁻¹ represents the (i−N + 1) to (i−1) th word string. Usually, N = 2 and N = 3 are often used, and are called bigram and trigram, respectively. When the N-gram language model is used, the generation probability P (w ₁ ⁿ ) of the word string w ₁ ⁿ = (w ₁ , w ₂ ,..., W _n ) is expressed by the following (Equation 1).

The N-gram language model is created by the maximum likelihood estimation method using large-scale learning text data. The maximum likelihood estimation method is a method of learning model parameters so that the generation probability of learning text data is maximized. However, it is usually difficult to prepare learning text data whose content completely matches the recognition target speech. Therefore, the language model created in advance does not always appropriately represent the appearance tendency of words in the recognition target speech. Therefore, a technique for adapting a language model in accordance with the characteristics of the recognition target speech is desired.

A cache model has been proposed as a technique for adapting a language model in accordance with the characteristics of speech to be recognized (see Non-Patent Document 1, for example). The cache model is a method of adapting the language model to the speech to be recognized by utilizing the property of the word that “the same word or phrase is easily used repeatedly”. In the following, the previously created N-gram language model is P _BASE (w _i | w _{i-N + 1} ^i-1 ), and this is adapted by the cache model, so that the i-th word w _{i in} the recognition target The case of obtaining a language model that gives a generation probability will be described.

First, the M word w i-M i-1 appearing immediately before the word w i as a cache interval cache probability P CACHE | calculated by (w i w i-M i-1) to the following (Equation 2) . The recognition result may be used as the immediately preceding M word w i-M i-1 .

δ (·) is a Kronecker δ function, which is 1 when the two arguments are equal, and 0 otherwise. Therefore, the cache probability P _CACHE (w _i | w _i−M ⁱ⁻¹ ) is a word distribution in the cache interval of the M word immediately before the word w _i , and this is the appearance of the word near the word w _{i to be} recognized. It is thought that it represents a trend. The cache length M is experimentally determined in advance.

Then, cache probability _P CACHE to adapt the ^{_{| | (w i-N +}} 1 i-1 w i) by using the ^{_{(w i w i-M i}} -1), a pre-created N-gram language model _P BASE Then, an adaptive language model P _ADAPT (w _i | W _{i−N + 1} ⁱ⁻¹ , w _i−M ⁱ⁻¹ ) that gives the generation probability of the i-th word w _i is obtained. Specifically, P _BASE and P _CACHE may be linearly interpolated by the following ( _Equation 3).

However, λ is a constant of 0 to 1 and is experimentally determined in advance. The language model P _ADAPT thus obtained is a model reflecting the appearance tendency of words in the recognition target speech based on the language model P _BASE created in advance.

Note that in Non-Patent Document 1, the immediately preceding M word is treated equally, but a method for calculating the cache probability by considering that the closer the word to the word w _i is, the greater the influence is proposed (for example, non-patent) Reference 2). In Non-Patent Document 2, the cache probability P _CACHE is calculated by the following ( _Equation 4) so that the influence decreases as the distance from the word w _i increases.

However, in the above (Equation 4), α is a decay rate of the influence of the word, and β is a normalization constant. In the above (Equation 4), the cache interval is set to all the words before the word w _i .

R.Kuhn and R.de Mori, "A Cache-Based Natural Language Model forSpeech Recognition," IEEE Transactions on Pattern Analysis and MachineIntelligence, pp.570-583, 1990. P.R.Clarkson and A.J.Robinson, "Language Model Adaptation UsingMixtures and an Exponentially Decaying Cache," Proceedings of The 1997 IEEEInternational Conference on Acoustics, Speech, and SignalProcessing, pp.799-802,1997.

However, in the techniques disclosed in Non-Patent Document 1 and Non-Patent Document 2 described above, when the topic suddenly changes in the recognition target speech, and when the same topic is repeated at intervals in the recognition target speech, There is a problem that the adaptation effect of the language model is insufficient.

The reason is that in the above technique, the language model is adapted using the cache interval set without considering the topic change in the recognition target speech. The topic here is, for example, the genre of news such as “politics”, “economics”, “sports” if the sound of a news program, “election of the House of Representatives”, “XX company goes bankrupt” In addition, individual agenda items such as the agenda of the conference if the audio of the conference.

In the following, this problem will be described in detail by taking as an example the case where the sound of a news program is targeted for recognition. FIG. 11 is a diagram for explaining a conventional problem, and shows that a news program is composed of news of various topics. The topics in the sections 1 to 4 are “sports”, “politics”, “economy”, and “sports”, respectively. “Sports” exist in two places, which means that after sports news was taken up at the beginning of the program, the same sports news was again taken up in detail in the sports corner in the second half of the program.

At this time, it is considered that the language model is adapted by the technique disclosed in Non-Patent Document 1 or Non-Patent Document 2 described above and an adapted language model used at time T is created. What is important here is that the appearance tendency of words is completely different for different topics. For example, in sports news, the appearance probability of a sports player name or the like is considered to be high, whereas in economic news, the appearance probability of an economic term is considered to be high. Therefore, the adaptive language model used at time T is preferably adapted to reflect the appearance tendency of words in section 1 and section 4, which are the same topic (sports) as time T, and is a topic different from time T. The appearance tendency of words in section 2 and section 3 should not be used for adaptation.

However, in the technique disclosed in Non-Patent Document 1 or Non-Patent Document 2 described above, as in the example illustrated in FIG. 11, a topic 3 (economic) section 3 different from a topic (sports) to be applied is a cache section. May be included. Furthermore, section 1 that is distant from time T may not be included in the cache section. As a result, the cache probability that appropriately reflects the appearance tendency of the word to be adapted cannot be obtained, and the adaptation effect becomes insufficient. In the technology disclosed in Non-Patent Document 1 or Non-Patent Document 2 described above, such a problem cannot be avoided because the cache interval is set without considering the change in the topic in the recognition target speech.

An object of the present invention is to solve the above-mentioned problem and to obtain a sufficient adaptation effect of a language model even when the topic changes suddenly and even when the same topic is repeated at intervals. An adaptive device, a speech recognition device, a language model adaptation method, and a computer-readable recording medium are provided.

In order to achieve the above object, a language model adaptation apparatus according to the present invention is a language model adaptation apparatus that performs adaptation of a base language model,
A dividing unit for dividing the input text into a plurality of sections;
Determining a topic included in each of the plurality of sections, and for each determined topic, a topic analysis unit that creates a topic model representing an appearance tendency of words in the topic;
For each of the plurality of sections, using the topic model corresponding to the topic included in the section, adapting the base language model, and creating an adapted language model, It is characterized by providing.

To achieve the above object, a speech recognition device according to the present invention is a speech recognition device that performs speech recognition by adapting a language model,
A speech recognition unit that performs speech recognition on speech data using the language model;
A dividing unit for dividing the text obtained by the speech recognition into a plurality of sections;
Determining a topic included in each of the plurality of sections, and for each determined topic, a topic analysis unit that creates a topic model representing an appearance tendency of words in the topic;
For each of the plurality of sections, using the topic model corresponding to the topics included in the section, adapting the base language model, and creating an adapted language model, ,
A re-recognition unit that performs speech recognition on the speech data in the section corresponding to the adaptation language model using the adaptation language model created by the adaptation language model creation unit. Features.

To achieve the above object, a language model adaptation method according to the present invention is a language model adaptation method for adapting a base language model,
(A) dividing the text into a plurality of sections;
(B) determining a topic included in each of the plurality of sections, and creating a topic model representing the appearance tendency of words in the topic for each determined topic;
(C) For each of the plurality of sections, using the topic model corresponding to the topic included in the section, adapting the base language model and creating an adapted language model; It is characterized by having.

In order to achieve the above object, a computer-readable recording medium according to the present invention is a computer-readable recording medium in which a program for executing adaptation of a language model as a base by a computer is recorded.
In the computer,
(A) dividing the text into a plurality of sections;
(B) determining a topic included in each of the plurality of sections, and creating a topic model representing the appearance tendency of words in the topic for each determined topic;
(C) For each of the plurality of sections, performing the step of adapting the base language model using the topic model corresponding to the topic included in the section and creating an adapted language model A program including an instruction is recorded.

With the above features, according to the language model adaptation device, the speech recognition device, the language model adaptation method, and the computer-readable recording medium according to the present invention, when a topic changes suddenly, the same topic is repeated at intervals. Even when the language model is used, a sufficient adaptation effect of the language model can be obtained.

FIG. 1 is a block diagram showing a schematic configuration of a language model adaptation apparatus according to an embodiment of the present invention. FIG. 2 is a flowchart showing the operation of the language model adaptation apparatus in the embodiment of the present invention. FIG. 3 is a block diagram illustrating a schematic configuration of the language model adaptation device and the speech recognition device according to the first embodiment of the present invention. FIG. 4 is a flowchart showing operations of the language model adaptation device and the speech recognition device according to the first exemplary embodiment of the present invention. FIG. 5 is a diagram for explaining a specific example of the recognition result in the first embodiment of the present invention. FIG. 6 is a diagram for explaining a specific example of the division result in the first embodiment of the present invention. FIG. 7 is a diagram illustrating a specific example of the topic determination result in the first embodiment of the present invention. FIG. 8 is a diagram for explaining another specific example of the division result according to the first embodiment of the present invention. FIG. 9 is a diagram illustrating a specific example of the topic determination result in the second embodiment of the present invention. FIG. 10 is a block diagram illustrating an example of a computer that can implement the language model adaptation device and the speech recognition device according to the first and second embodiments of the present invention. FIG. 11 is a diagram for explaining a conventional problem.

(Embodiment)
Hereinafter, a language model adaptation apparatus and adaptation method according to an embodiment of the present invention will be described with reference to FIGS. 1 and 2. Initially, the structure of the language model adaptation apparatus in this Embodiment is demonstrated using FIG. FIG. 1 is a block diagram showing a schematic configuration of a language model adaptation apparatus according to an embodiment of the present invention.

As shown in FIG. 1, a language model adaptation apparatus 100 according to the present embodiment is an apparatus that executes adaptation of a base language model (hereinafter referred to as “base language model”). The language model adaptation apparatus 100 includes a division unit 105, a topic analysis unit 107, and an adaptation language model creation unit 110.

In this embodiment, the language model adaptation apparatus 100 further includes a base language model storage unit 102 that stores a base language model. An example of the base language model is a known N-gram language model. The N-gram language model is a probability model that gives the word generation probability on the assumption that the word generation probability at a certain time depends only on the immediately preceding N−1 words.

The dividing unit 105 divides the input text into a plurality of sections. The topic analysis unit 107 determines topics included in each of the plurality of sections. In addition, “determining a topic” is not limited to specifying a specific topic, but also includes determining whether a topic in one section and a topic in another section are the same or similar. It is. Further, after determining the topic, the topic analysis unit 107 creates a topic model representing the appearance tendency of words in the topic for each determined topic.

Also, the adaptation language model creation unit 110 adapts the base language model for each of the plurality of sections using the topic model corresponding to the topic included in the section, and creates an adaptation language model. The created adaptive language model is then output.

Next, the operation of the language model adaptation apparatus 100 shown in FIG. 1 will be described with reference to FIG. FIG. 2 is a flowchart showing the operation of the language model adaptation apparatus in the embodiment of the present invention. In the present embodiment, the language model adaptation method is executed by operating the language model adaptation apparatus 100. For this reason, the description of the language model adaptation method in the present embodiment is replaced with the following description of the operation of the language model adaptation apparatus 100.

As shown in FIG. 2, first, when a text is input to the language model adaptation apparatus 100, the dividing unit 105 receives the input of the text (step S1). The text input in step S1 may be any text. Specifically, texts created using a base language model, for example, texts obtained by speech recognition and texts obtained by machine translation can be mentioned.

Next, the dividing unit 105 divides the input text into a plurality of sections (step S2). In step S2, the dividing unit 105 can perform division based on, for example, the number of words in the text or the number of utterances (in the case of speech recognition). Further, the dividing unit 105 can perform division based on a change point of the word distribution in the text or a position where a preset topic boundary expression appears in the text, and is related to a preset topic. Division can also be performed using a model representing the appearance tendency of words.

Next, the topic analysis unit 107 determines one or more topics included in the section for each section divided by the dividing unit 105 (step S3). In step S3, the topic analysis unit 107 classifies a plurality of sections into groups based on the similarity between the texts of the sections, and determines that the topics of the sections belonging to the same group are common. In the present embodiment, as the topic determination, it is determined whether the topic is common between sections.

Further, in step S3, the topic analysis unit 107 assumes a probability model representing the appearance tendency of words in each section using the topic as a hidden variable, and further learns the parameters of the probability model using the text in each section as learning data. In this way, the topic can be determined.

Subsequently, the topic analysis unit 107 creates a topic model representing the appearance tendency of words in the topic for each determined topic (step S4). After that, the adaptation language model creation unit 110 adapts the base language model to each divided section by using one or more topic models of topics that the topic analysis unit 107 determines to be included in the section. Thus, an adaptive language model adapted to the section is created (step S5).

Thereafter, the language model adaptation device 100 outputs the adaptation language model created in step S5 (step S6), and ends the process.

As described above, in the present embodiment, language model adaptation apparatus 100 divides text into a plurality of sections, and for each topic, in other words, for each section having a common topic, Implement adaptation that reflects the appearance trend of. Therefore, according to the present embodiment, sufficient recognition effect of the language model can be obtained even when the topic suddenly changes in the recognition speech or the like that is the source of the text, and even when the same topic is repeated at intervals. Obtainable. Note that language model adaptation apparatus 100 according to the present embodiment may include a speech recognition unit as shown in Examples 1 and 2 below when text obtained by speech recognition is targeted.

Next, the present invention will be described using examples. Hereinafter, a language model adaptation device, a speech recognition device including the language model adaptation device, and a language model adaptation method according to Embodiment 1 of the present invention will be described in detail with reference to FIGS. First, the configuration of the language model adaptation device and the speech recognition device according to the first embodiment will be described with reference to FIG. FIG. 3 is a block diagram illustrating a schematic configuration of the language model adaptation device and the speech recognition device according to the first embodiment of the present invention.

The language model adaptation apparatus 100 in the first embodiment shown in FIG. 3 constitutes a part of the speech recognition apparatus 10 and adapts the language model in accordance with the characteristics of the recognition target speech. Further, the speech recognition apparatus 10 according to the first embodiment recognizes speech data with higher accuracy using the language model adapted by the language model adaptation apparatus 100.

As shown in FIG. 3, the language model adaptation apparatus 100 according to the first embodiment includes a base language model storage unit 102, a division unit 105, a topic analysis unit 107, an adaptation language, as described in the embodiment. And a model creation unit 110. Furthermore, in the first embodiment, the language model adaptation device 100 includes a recognition result storage unit 104, a division result storage unit 106, a topic determination result storage unit 108, a topic model storage unit 109, and an adaptive language model storage unit. 111. In addition to the language model adaptation device 100, the speech recognition device 10 according to the first embodiment includes a speech data storage unit 101, a speech recognition unit 103, a re-recognition unit 112, and a re-recognition result storage unit 113. Yes. Each of these units operates as follows.

The voice data storage unit 101 stores voice data to be recognized. As audio data, various audio such as conference audio, lecture audio, broadcast audio, and the like can be considered. Note that these audio data may be archive data prepared in advance or data input in real time from a microphone or the like.

The base language model storage unit 102 stores a base language model used by the speech recognition unit 103. A known N-gram language model can be used as the base language model. The base language model can be constructed in advance by performing learning using a large amount of text data. In the first embodiment, as described later, by adapting the base language model according to the characteristics of the recognition target speech, it is possible to recognize speech data with higher accuracy.

The voice recognition unit 103 reads the voice data stored in the voice data storage unit 101, and performs voice recognition of the voice data using the base language model stored in the base language model storage unit 102. Then, the speech recognition unit 103 outputs the recognition result as text (text data) to the recognition result storage unit 104. The recognition result storage unit 104 stores the recognition result (see FIG. 5).

The dividing unit 105 divides the recognition result text stored in the recognition result storage unit 104 into a plurality of sections, and outputs the division result to the division result storage unit 106. The division result storage unit 106 stores the division result (see FIG. 6).

The topic analysis unit 107 refers to the recognition result text stored in the recognition result storage unit 104 and the division result stored in the division result storage unit 106, and is included in the section for each divided section. Determine the topic. Further, the topic analysis unit 107 outputs the determination result to the topic determination result storage unit 108 and stores it therein. Furthermore, the topic analysis unit 107 creates a topic model representing the appearance tendency of words in the topic for each determined topic, and outputs the created topic model to the topic model storage unit 109 for storage therein. .

The adaptation language model creation unit 110 refers to the topic determination result stored in the topic determination result storage unit 108 and the topic model stored in the topic model storage unit 109. Then, the adaptation language model creation unit 110 creates, for each divided section, an adaptation language model adapted to the section from the base language model using the topic model of the topic included in the section. To do. Also, the adaptation language model creation unit 110 outputs the created adaptation language model to the adaptation language model storage unit 111 and stores it therein.

The re-recognition unit 112 uses the adaptation language model for each section stored in the adaptation language model storage unit 111 to perform speech recognition (repeating speech corresponding to the corresponding section of the speech data stored in the speech data storage unit 101 ( Execute (Re-recognition). Then, the re-recognition unit 112 outputs the re-recognition result to the re-recognition result storage unit 113 as text data.

Next, operations of the language model adaptation apparatus 100 and the speech recognition apparatus 10 in the first embodiment will be described with reference to FIGS. FIG. 4 is a flowchart showing operations of the language model adaptation device and the speech recognition device according to the first exemplary embodiment of the present invention. FIG. 5 is a diagram for explaining a specific example of the recognition result in the first embodiment of the present invention. FIG. 6 is a diagram for explaining a specific example of the division result in the first embodiment of the present invention. FIG. 7 is a diagram illustrating a specific example of the topic determination result in the first embodiment of the present invention. FIG. 8 is a diagram for explaining another specific example of the division result according to the first embodiment of the present invention.

In the first embodiment, the language model adaptation method is executed by operating the language model adaptation device 100. Therefore, the description of the language model adaptation method is replaced with the description of the operation of the language model adaptation apparatus 100. In the following description, FIG. 3 will be referred to as appropriate.

First, as shown in FIG. 4, the speech recognition unit 103 performs speech recognition on speech data using a base language model, and converts this into text (step S201). Specifically, the voice recognition unit 103 reads the voice data stored in the voice data storage unit 101 and detects a plurality of utterances by a known voice detection technique. Furthermore, the speech recognition unit 103 converts the speech data into text by applying a known large vocabulary continuous speech recognition technology to each utterance, and outputs this to the recognition result storage unit 104. .

In step S201, the base language model stored in the base language model storage unit 102 is used as the language model for speech recognition. As the acoustic model, for example, a known HMM acoustic model with phonemes as units can be used.

As a result of step S201, for example, as shown in FIG. 5, an utterance of 10 to 13 seconds of voice data is recognized as “Good evening is news”, and an utterance of 16 to 20 seconds is “the first news today is baseball. Is a Japanese series. "

Next, the dividing unit 105 divides the recognition result text stored in the recognition result storage unit 104 into a plurality of sections, and outputs the division result to the division result storage unit 106 (step S202). In the first embodiment, the recognition result is divided for each predetermined number of utterances. For example, the recognition result may be divided for each predetermined number of words.

In the example of FIG. 6, the recognition result text is divided into five utterances. As shown in FIG. 6, the recognition results are divided into a section 1 composed of utterances 1 to 5 and a section 2 composed of utterances 6 to 10. Naturally, each section is associated with a recognition result text of an utterance constituting the section. Note that the number of utterances included in one section can be determined by conducting an experiment in advance.

Next, the topic analysis unit 107 determines, for each divided section, the topics included in the section and outputs the determination result to the topic determination result storage unit 108 (step S203). Subsequently, the topic analysis unit 107 creates a topic model indicating the appearance tendency of words in the topic for each determined topic, and outputs this to the topic model storage unit 109 (step S204).

Here, first, an example of a method for determining a topic included in each divided section will be described in step S203. In the first embodiment, the topic analysis unit 107 performs clustering on each section based on the similarity of semantic content, and the sections assigned to the same cluster have the same topic. The topic is judged.

As shown in FIG. 7, as a result of clustering each section, section 1, section 2, section 9, and section 10 are assigned to the same cluster. This indicates that these sections are determined to be sections related to the same topic (topic 1). Similarly, the sections 3 and 4 are determined to be sections related to the topic 2, and the sections 5, 6, 7, and 8 are determined to be sections related to the topic 3.

Thus, by clustering each section based on the similarity of semantic content, it can be determined which section contains the same topic as each other section. At this time, it is not necessary to determine the specific content of each topic (topic 1, topic 2, topic 3), which is a problem in the present invention. Don't be.

By the way, when clustering each section, it is necessary to obtain the similarity between the two sections. Since the purpose is to group sections of the same topic into the same cluster, it is desirable that the similarity between the two sections reflects the similarity of the semantic content between the two sections. As such similarity, the text of each section obtained by the division is regarded as a document, and the cosine similarity between document vectors can be used. A document vector is a vector obtained by assigning the weight of each word appearing in a document to each dimension. In the first embodiment, the weight of each word appearing in the text of the section may be used. Examples of the word weight include the appearance frequency of the word in the document and the word weight calculated by a known TF / IDF method. For example, d ₁ the document vector of the section _1, when the document vector of the section 2 was set to d _2, cosine similarity cosθ of the section 1 and the section 2 is calculated by the following equation (5).

The cosine similarity cos θ is a real number from 0 to 1, and the closer the value is to 1, the greater the similarity between both sections. As each dimension constituting the document vector, it is not necessary to use all the words appearing in the section, and only words having predetermined parts of speech such as nouns and verbs may be used. By doing so, the similarity of the document vector becomes a scale reflecting the similarity of the contents.

In the first embodiment, a known k-means method or hierarchical clustering method may be used as a clustering algorithm. The procedure when the k-means method is used when the number of clusters is K is as follows. (1) Document vectors of K sections selected at random are determined as representative points of K clusters. (2) Document vectors in all sections are assigned to representative points that are the closest, that is, the highest similarity. (3) The centroid point of the document vector is calculated for each cluster and used as a representative point. (2) and (3) are repeatedly executed. When the section (document vector) assigned to each cluster does not change, the clustering is terminated.

The simplest hierarchical clustering procedure is as follows. When the total number of sections is N, (1) N clusters including only one section are created. (2) Merge two clusters having the highest similarity. If this merging is repeated until the number of clusters reaches K, K clusters can be obtained. Here, as the similarity between two clusters, a known shortest distance method, longest distance method, Ward method, or the like may be used.

Whichever clustering method is used, the cosine similarity of the document vector can be used as the similarity between the two sections. Of course, methods other than those listed here may be used as the clustering method. In addition, the similarity between two sections is not limited to the cosine similarity, and any scale can be used as long as it can measure the similarity of the semantic content between the two sections.

Note that the number of clusters K for clustering may be, for example, K = α · N, where N is the total number of sections to be clustered. Here, α is a proportionality constant. For example, if α = 0.2, one cluster includes five sections on average. In addition, the number of clusters K may be proportional to the total number of utterances, not the total number of sections, and may of course be determined by other methods.

Thus, in step S203, the sections divided by the dividing unit 105 are clustered, and the result is output to the topic determination result storage unit 108. By such processing, even when a plurality of topics are included in the recognition target speech, it is possible to determine the topics in each section of the recognition result.

Next, a method for creating a topic model for each topic determined by clustering will be described. A topic model of a certain topic is a model representing the appearance tendency of words in the topic. The topic model can be created as follows. (1) The texts of a plurality of sections determined as the same topic (cluster) by clustering are collected. (2) A probability model representing the word generation probability is learned based on the text compiled for each topic (cluster). Here, an N-gram language model may be used as the probability model.

In the example of FIG. 7, since the sections determined as the topic 1 are the sections 1, 2, 2, and 10, first, the texts of these four sections are collected as one group. Next, the topic model of topic 1 is obtained by learning the N-gram language model by the maximum likelihood estimation method based on the collected text. Specifically, an N = 1 unigram model, an N = 2 bigram model, an N = 3 trigram model, and the like may be learned. Similarly, the topic model of topic 2 may be learned based on the texts of sections 3 and 4, and the topic model of topic 3 may be learned based on the texts of sections 5, 6, 7, and 8. Just do it.

Thus, in step S204, a topic model representing the appearance tendency of words is created for each topic obtained by clustering, and the topic model is output to the topic model storage unit 109. Therefore, even when a plurality of topics are included in the recognition target speech, it is possible to create a topic model representing the appearance tendency of words separately for each topic. Moreover, even if the same topic appears several times at intervals, the same topic model can be created by collecting the appearance tendency of words in the entire section of the same topic.

Next, the adaptation language model creation unit 110 refers to the determination result stored in the topic determination result storage unit 108 and the topic model stored in the topic model storage unit 109, and is obtained by division. For each section, an adaptive language model adapted to the section is created (step S205). The created adaptation language model is output to the adaptation language model storage unit 111.

Specifically, the adaptation language model creation unit 110 adapts the base language model stored in the base language model storage unit 102 using the topic model of the topic determined in each interval, so that each interval Create different adaptive language models for. For example, in the example shown in FIG. 7, an adaptive language model is created by adapting the base language model with the topic model of Topic 1 for Section 1, Section 2, Section 9, and Section 10. The Similarly, for section 3 and section 4, an adaptive language model is created by adapting the base language model with the topic model of topic 2.

Examples of the adaptation method include linear interpolation between a base language model and a corresponding topic model. Specifically, when the base language model P _BASE and the topic model P _{TOPIC_1} corresponding to the topic 1 are both trigram models, the adaptive language model P _{ADAPT_1} corresponding to the section of the topic 1 is 6).

However, in the above (Equation 6), λ is a constant of 0 to 1, and can be determined by conducting an experiment in advance. Here, the base language model, the topic model, and the adaptation language model are all trigram examples, but other N-gram language models such as bigrams and unigrams may be used. When the topic model P _{TOPIC_1} is a unigram model, the base language model may be adapted by a known unigram rescaling method. In this case, the adaptive language model P _{ADAPT_1} is obtained by the following ( _Equation 7).

The adaptive language model finally created for each section by the processing of steps S201 to S205 described above reflects the appearance tendency of words in the same topic section as the section, but has a topic section different from the section. It is not influenced by the appearance tendency of words. For example, in the example shown in FIG. 7, the adaptive language model created for the section 9 learns the appearance tendency of words in the texts of the sections 1, 2, and 10 that are the same topic as the section 9. On the other hand, the texts of the sections 3 to 8 that are different from the section 9 are not used for learning.

Therefore, the adaptive language model created for each section according to the present invention is not limited to words in the section even when the topic suddenly changes in the recognition target speech or when the same topic is repeated in the recognition target speech. It can be said that it is a model that can accurately predict the appearance tendency of. In other words, since the appearance tendency of words differs depending on the topic, it is generally desirable to adapt the language model using only the appearance tendency of words in the same topic in adaptation of the language model, but such adaptation is realized by the present invention. It becomes possible.

Note that a cache model (see Non-Patent Document 1) is a typical method for adapting a language model using the appearance tendency of words in recognition target speech. In the cache model, for example, when an adapted language model is created for the section 9 (topic 1), the adaptation is performed based on the appearance tendency of words in the immediately preceding section 8 (topic 3). In this case, since the topic of section 8 is different from that of section 9, the appearance tendency of both words is considered to be different, and as a result, there arises a problem that an appropriate adaptive language model cannot be obtained. On the other hand, as can be seen from the first embodiment, in the present invention, it is possible to adapt the base language model without causing such topic inconsistency, so various topics appear in the recognition target speech. Even in this case, a sufficient adaptation effect can be obtained.

Finally, after the completion of step S205, the re-recognition unit 112 uses the adaptation language model for each section stored in the adaptation language model storage unit 111, and uses the corresponding section of the speech data stored in the speech data storage unit 101. Again, voice recognition (re-recognition) is executed (step S206). The re-recognition result is output to the re-recognition result storage unit 113 as text data.

That is, for example, in the example shown in FIG. 7, the utterances 1 to 5 are re-recognized by the adaptation language model created for the section 1 and the utterance 6 is created by the adaptation language model created for the section 2. Re-recognizing ~ 10 is repeated for all intervals.

As described above, the language model adapted to each section according to the present invention is a model that can accurately predict the appearance tendency of words in the section even when there is a topic change or repetition in the recognition target speech. . For this reason, the re-recognition unit 112 can recognize the voice data with high accuracy.

In the first embodiment, the dividing unit 105 divides the recognition result text into a plurality of sections by dividing the recognition result for each predetermined number of utterances. However, there is an overlap in these sections. Also good. FIG. 8 shows such an example. FIG. 8 shows a division result when 10 utterances are grouped into one section and divided into a plurality of sections by shifting each utterance by two utterances, and a topic determination result in each section. Thus, even when there is an overlap in each section, the topic analysis unit 107 can operate in exactly the same way.

In the first embodiment, by overlapping each section in this way, even when the number of utterances included in one section is increased, it is possible to grasp changes in the topic in the recognition target speech in terms of time. It is. That is, when the re-recognition unit 112 re-recognizes using the adaptive language model corresponding to each section, for example, only the utterance at the center of the section may be re-recognized.

Specifically, the utterance 13 and the utterance 14 (two utterances in the middle of the utterances 9 to 18) may be recognized using the adaptive language model corresponding to the section 5. Further, the utterance 15 and the utterance 16 (two utterances at the center of the utterances 11 to 20) may be recognized using the adaptive language model corresponding to the section 6. By doing so, it is possible to prevent the utterances to be re-recognized from overlapping.

In the first embodiment, the dividing unit 105 can also detect a boundary of meaning content in the recognition result, and can divide the text of the recognition result based on the detection result. For example, according to the method described in Reference Document 1, the dividing unit 105 may detect a change point of the word distribution in the recognition result text and divide the recognition result text based on the detected change point.

(Reference 1)
Marti A. Hearst, "MULTI-PARAGRAPH
Segmentation of Expository Text, "32nd Annual Meeting of The Association
for Computational Linguistics, pp.9-16, 1994.

Further, the dividing unit 105 can also divide the text of the recognition result based on the position where the topic boundary expression appears when the topic boundary expression representing the topic boundary is determined in advance. Examples of topic boundary expressions include expressions that change topics such as “Now”, “Now” and “Next”.

Further, when topic models relating to various topics are prepared in advance, the dividing unit 105 obtains a topic model sequence that best matches the recognition result text by the method described in Reference Document 2, and changes in topics And the text of the recognition result can be divided based on the change point. Note that the topic model here is not the topic model created by the topic analysis unit 107. The topic model here is a topic model prepared in advance regardless of the recognition result text, and represents the appearance tendency of words related to a predetermined topic. Such a topic model can be obtained, for example, by learning in advance a large amount of text data to which topic information is assigned.

(Reference 2)
JPYamron, I.Carp, L.Gillick, S.Lowe, and P.van Mulbregt, “A hidden Markov Model
Approach to Text Segmentation and Event Tracking, "IEEE International
Conference on Acoustics, Speech and Signal Processing, pp.333-336, 1998.

When the recognition result text is divided into a predetermined number of utterances or a predetermined number of words, each section does not necessarily represent a group of topics, and a plurality of topics are mixed in the section near the topic boundary. There is a possibility. However, when the dividing unit 105 divides the text of the recognition result into a plurality of sections based on the semantic content boundary, each section may represent a group of topics even if they may be mixed. become. For this reason, the clustering accuracy and topic model estimation accuracy in the topic analysis unit 107 are improved, and the accuracy of the adaptive language model finally obtained is further improved.

As described above, according to the first embodiment, a sufficient adaptation effect of the language model can be obtained even when the topic suddenly changes in the recognition target speech and when the same topic is repeated in the recognition target speech at intervals. It is done. Further, if the recognition target speech is re-recognized using the adaptive language model generated by the language model adaptation device 100 of the first embodiment, a recognition result with higher accuracy can be obtained.

Next, a language model adaptation device, a speech recognition device including the language model adaptation device, and a language model adaptation method according to Embodiment 2 of the present invention will be described in detail with reference to the drawings. The second embodiment is different from the first embodiment in terms of operations in the topic analysis unit 107 and the adaptive language model creation unit 110. In other respects, the second embodiment is the same as the first embodiment. Therefore, the description about a common part is abbreviate | omitted. Different parts will be described with reference to FIG. FIG. 9 is a diagram illustrating a specific example of the topic determination result in the second embodiment of the present invention.

In the second embodiment, the topic analysis unit 107 uses the topic z as a hidden variable as a probability variable P (w | d) that gives the appearance probability of the word w in each section d divided by the division unit 105 (the following ( Assume the model defined by equation (8). Then, the topic analysis unit 107 learns the parameters of the probability model using the text of each section as learning data.

Here, topic z is a hidden variable that cannot be observed from the outside, and the number of topics is set to an appropriate value in advance. The number of topics z may be set to be proportional to the total number of sections or the total number of utterances as in the case where the number of clusters is determined in the first embodiment, or may be set by another method.

The parameter P (w | z) in the above (Equation 8) represents the word appearance probability in the topic z, and the parameter P (z | d) represents the ratio of each topic z included in the section d. In other words, the probability model of the above (Equation 8) can be considered as a model in which the appearance probability P (w | d) of the word w in the section d is expressed as a weighted sum of the appearance probabilities of words in a plurality of topics. it can.

Therefore, in the second embodiment, topics included in each section are determined by learning the parameters of the probability model of the above (Equation 8) using the text of each section as learning data (see S203 in FIG. 4). At the same time, each topic model is created (see S204 in FIG. 4). That is, in the second embodiment, P (z | d) may be output to the topic determination result storage unit 108 as the determination result of the topic included in the section d, and P (w | z) What is necessary is just to output to the topic model memory | storage part 109 as a topic model showing the appearance tendency of the word in.

The parameters of the above probability model (Equation 8) can be estimated by maximum likelihood estimation for learning data using the EM algorithm, as described in Reference Document 3 below. That is, calculation of P (z | w, d) by the following (Equation 9), estimation of the parameter P (w | z) by the following (Equation 10), and P (z by the following (Equation 11) It is sufficient to repeat the estimation of | d).

(Reference 3)
D. Gildea and T. Hofmann, "Topic-based Language Models Using
em, "Proceedings of
The 6th European Conference on Speech Communication and Technology, 1999.

Here, in the above (Equation 10) and (Equation 11), n (w, d) represents the number of appearances of the word w in the section d. In the above (Equation 9) to (Equation 11), r represents the number of repetitions.

As shown in FIG. 9, the topic determination result according to the second embodiment is expressed as a combination of a plurality of topics. That is, for example, section 1 indicates that topic 1 is a topic combined with weight 0.8, topic 2 is weight 0.1, and topic 5 is combined with weight 0.1. 1 is a value calculated as P (z | d) in 1.

Then, also in the second embodiment, the adaptive language model creation unit 110 refers to the topic determination result stored in the topic determination result storage unit 108 and the topic model stored in the topic model storage unit 109. . Based on the reference result, the adaptation language model creation unit 110 creates an adaptation language model adapted to the section for each section obtained by the division, and outputs this to the adaptation language model storage unit 111. (See S205 in FIG. 4).

In the second embodiment, the topic analysis unit 107 has already obtained a topic ratio P (z | d) for each section d and a word appearance probability P (w | z) for each topic z. . Therefore, the word appearance probability P (w | d) can be obtained for each section d by the above (Equation 8). Therefore, an adaptive language model for each section can be created by adapting the base language model stored in the base language model storage unit 102 using P (w | d). Specifically, linear interpolation shown in the following (Equation 12) or unigram rescaling shown in the following (Equation 13) may be used.

In the second embodiment, the language model can be adapted after expressing the topic in the recognition target speech as a combination of a plurality of basic topics and extracting the appearance tendency of words in the basic topics. For example, even when “Section 1 includes Topic 1, Section 2 includes Topic 1 and Topic 2, and Section 3 includes Topic 2”, according to the second embodiment, it is included in each section. Each topic model of Topic 1 and Topic 2 can be created after determining the percentage of topics that are to be viewed.

Thus, even if it is difficult to cluster sections because each section includes a plurality of topics, according to the second embodiment, a topic model can be created for each topic. When the language model is adapted to each section, weighting of a plurality of topic models is performed according to the ratio of topics included in the section, so that the adaptation effect of the language model is improved. To do.

The language model adaptation apparatus and the speech recognition apparatus in the first and second embodiments described above are realized by using, for example, a program (language model adaptation program) that causes a computer to execute steps S201 to S206 shown in FIG. Can do. That is, if a language model adaptation program is installed in a computer and executed, a language model adaptation device and a speech recognition device are realized.

Here, the language model adaptation apparatus and the speech recognition apparatus according to the first embodiment or the second embodiment, which are realized by a program, will be described with reference to FIG. FIG. 10 is a block diagram illustrating an example of a computer that can implement the language model adaptation device and the speech recognition device according to the first and second embodiments of the present invention.

As shown in FIG. 10, the computer 300 includes a data processing device 320 including a CPU and a storage device 330 including a magnetic disk, a semiconductor memory, and the like.

The storage area of the storage device 330 includes a voice data storage unit 331, a base language model storage unit 332, a recognition result storage unit 333, a division result storage unit 334, a topic determination result storage unit 335, a topic model storage unit 336, an adaptation A structured language model storage unit 337 and a re-recognition result storage unit 338 are constructed. The storage device 330 functions as these storage units 331 to 338.

The language model adaptation program 310 is read into the data processing device 320 by reading from a computer-readable recording medium or storage device, or by transmission via a network, and controls the operation of the data processing device 320. Thereby, the speech recognition unit 103, the division unit 105, the topic analysis unit 107, the adaptive language model creation unit 110, and the re-recognition unit 112 shown in FIG. The data processing device 320 functions as a speech recognition unit 103, a division unit 105, a topic analysis unit 107, an adaptive language model creation unit 110, and a re-recognition unit 112 (see FIG. 3) under the control of the language model adaptation program 310. And execute the process. Examples of the computer-readable recording medium include an optical disk, a magnetic disk, a magneto-optical disk, a semiconductor memory, and a floppy disk.

As described above, the present invention has been described with reference to the embodiments and examples, but the present invention is not limited to the above embodiments and examples. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

This application claims priority based on Japanese Patent Application No. 2009-50151 filed on Mar. 4, 2009, the entire disclosure of which is incorporated herein.

The language model adaptation apparatus, speech recognition apparatus, language model adaptation method, and computer-readable recording medium according to the present invention have the following characteristics.

(1) A language model adaptation device for adapting a language model as a base,
A dividing unit for dividing the input text into a plurality of sections;
Determining a topic included in each of the plurality of sections, and for each determined topic, a topic analysis unit that creates a topic model representing an appearance tendency of words in the topic;
For each of the plurality of sections, using the topic model corresponding to the topic included in the section, adapting the base language model, and creating an adapted language model, A language model adaptation device comprising:

(2) The method further includes: a speech recognition unit that performs speech recognition on speech data using the base language model and inputs text obtained by the speech recognition to the dividing unit. The language model adaptation device according to 1).

(3) The topic analysis unit classifies the plurality of sections into groups based on the similarity between the texts in the sections, and determines that the topics in the sections belonging to the same group are common, and For each group, create the topic model based on the appearance tendency of words in the text in the section belonging to the group,
The language model adaptation device according to (1) above.

(4) The topic analysis unit assumes a probability model representing the appearance tendency of words in each of the plurality of sections using the topic as a hidden variable, and further uses the text of each of the plurality of sections as learning data. Learning the parameters of the probability model;
The learning determines the topic and creates the topic model.
The language model adaptation device according to (1) above.

(5) The language model adaptation device according to (1), wherein the division unit divides the text based on the number of words or the number of utterances.

(6) The language model adaptation device according to (1), wherein the dividing unit detects a change point of the word distribution in the text and divides the text based on the detected change point.

(7) The language model adaptation device according to (1), wherein the division unit divides the text based on a position where a preset topic boundary expression appears in the text.

(8) The language model adaptation device according to (1), wherein the division unit divides the text using a model representing an appearance tendency of words related to a preset topic.

(9) A speech recognition device that performs speech recognition by adapting a language model,
A speech recognition unit that performs speech recognition on speech data using the language model;
A dividing unit for dividing the text obtained by the speech recognition into a plurality of sections;
Determining a topic included in each of the plurality of sections, and for each determined topic, a topic analysis unit that creates a topic model representing an appearance tendency of words in the topic;
For each of the plurality of sections, using the topic model corresponding to the topics included in the section, adapting the base language model, and creating an adapted language model, ,
Using the adaptation language model created by the adaptation language model creation unit, a re-recognition unit that performs speech recognition on the speech data in the section corresponding to the adaptation language model;
A speech recognition apparatus comprising:

(10) A language model adaptation method for adapting a base language model,
(A) dividing the text into a plurality of sections;
(B) determining a topic included in each of the plurality of sections, and creating a topic model representing the appearance tendency of words in the topic for each determined topic;
(C) For each of the plurality of sections, using the topic model corresponding to the topic included in the section, adapting the base language model and creating an adapted language model; A language model adaptation method characterized by comprising:

(11) (d) The method further includes the step of performing speech recognition on the speech data using the base language model, thereby generating the text to be divided in the step (a). The language model adaptation method according to (10) above.

(12) In the step of (b), classifying the plurality of sections into groups based on the similarity between the texts of the sections, and determining that the topics of the sections belonging to the same group are common, For each group, create the topic model based on the appearance tendency of words in the text in the section belonging to the group,
The language model adaptation method according to (10) above.

(13) In the step of (b), using the topic as a hidden variable, a probability model representing the appearance tendency of words in each of the plurality of sections is assumed, and further, the text of each of the plurality of sections is learned data. And learning the parameters of the probability model,
The learning determines the topic and creates the topic model.
The language model adaptation method according to (10) above.

(14) The language model adaptation method according to (10), wherein in the step (a), the text is divided based on the number of words or the number of utterances.

(15) The language model adaptation method according to (10), wherein in the step (a), a change point of the word distribution in the text is detected, and the text is divided based on the detected change point.

(16) The language model adaptation method according to (10), wherein in the step (a), the text is divided based on a position where a preset topic boundary expression appears in the text.

(17) The language model adaptation method according to (10), wherein in the step (a), the text is divided by using a model representing an appearance tendency of words related to a preset topic.

(18) A computer-readable recording medium recording a program for executing adaptation of a language model as a base by a computer,
In the computer,
(A) dividing the text into a plurality of sections;
(B) determining a topic included in each of the plurality of sections, and creating a topic model representing the appearance tendency of words in the topic for each determined topic;
(C) For each of the plurality of sections, performing the step of adapting the base language model using the topic model corresponding to the topic included in the section and creating an adapted language model A computer-readable recording medium on which a program including instructions is recorded.

(19) (d) Performing speech recognition on speech data by using the base language model, thereby further generating the text to be divided in the step (a). The computer-readable recording medium according to (18), which is executed by the computer.

(20) In the step of (b), classifying the plurality of sections into groups based on the similarity between the texts of the sections, and determining that the topics of the sections belonging to the same group are common, For each group, create the topic model based on the appearance tendency of words in the text in the section belonging to the group,
The computer-readable recording medium according to (18) above.

(21) In the step (b), using the topic as a hidden variable, a probability model representing the appearance tendency of words in each of the plurality of sections is assumed, and further, text in each of the plurality of sections is learned data. And learning the parameters of the probability model,
The learning determines the topic and creates the topic model.
The computer-readable recording medium according to (18) above.

(22) The computer-readable recording medium according to (18), wherein in the step (a), the text is divided based on the number of words or the number of utterances.

(23) The computer-readable recording according to (18), wherein in the step (a), a change point of the word distribution in the text is detected, and the text is divided based on the detected change point. Medium.

(24) The computer-readable recording medium according to (18), wherein in the step (a), the text is divided based on a position where a preset topic boundary expression appears in the text.

(25) The computer-readable recording medium according to (18), wherein in the step (a), the text is divided using a model representing an appearance tendency of words related to a preset topic.

The present invention realizes an automatic speech recognition system that recognizes speech data including various topics such as conference speech, lecture speech, broadcast speech, etc., and outputs text information, and an automatic speech recognition system for a computer. It can be applied to uses such as programs. Further, the present invention can be applied to an application such as an information search system for searching for these voice data using text information of a recognition result.

DESCRIPTION OF SYMBOLS 10 Speech recognition apparatus 100 Language model adaptation apparatus 101 Speech data storage part 102 Base language model storage part 103 Speech recognition part 104 Recognition result storage part 105 Division part 106 Division result storage part 107 Topic analysis part 108 Topic determination result storage part 109 Topic model Storage unit 110 Adaptive language model creation unit 111 Adaptive language model storage unit 112 Re-recognition unit 113 Re-recognition result storage unit 300 Computer 310 Language model adaptation program 320 Data processing device 330 Storage device 331 Speech data storage unit 332 Base language model Storage unit 333 Recognition result storage unit 334 Division result storage unit 335 Topic determination result storage unit 336 Topic model storage unit 337 Adaptive language model storage unit 338 Re-recognition result storage unit

Claims (25)

A language model adaptation device for adapting a language model as a base,
A dividing unit for dividing the input text into a plurality of sections;
Determining a topic included in each of the plurality of sections, and for each determined topic, a topic analysis unit that creates a topic model representing an appearance tendency of words in the topic;
For each of the plurality of sections, using the topic model corresponding to the topic included in the section, adapting the base language model, and creating an adapted language model, A language model adaptation device comprising: The speech recognition unit according to claim 1, further comprising: a speech recognition unit that performs speech recognition on speech data using the base language model and inputs text obtained by the speech recognition to the dividing unit. Language model adaptation device. The topic analysis unit classifies the plurality of sections into groups based on the similarity between the texts of the sections, and determines that the topics of the sections belonging to the same group are common, and further, for each group , Creating the topic model based on the appearance tendency of words in the text in the section belonging to the group,
The language model adaptation apparatus according to claim 1 or 2. The topic analysis unit assumes a probability model representing the appearance tendency of words in each of the plurality of sections using the topic as a hidden variable, and further uses the text of each of the plurality of sections as learning data to generate the probability model. Learn the parameters of
The learning determines the topic and creates the topic model.
The language model adaptation apparatus according to claim 1 or 2. 5. The language model adaptation device according to claim 1, wherein the dividing unit divides the text based on the number of words or the number of utterances. The language model adaptation device according to any one of claims 1 to 4, wherein the dividing unit detects a change point of a word distribution in the text and divides the text based on the detected change point. The language model adaptation device according to any one of claims 1 to 4, wherein the dividing unit divides the text based on a position where a preset topic boundary expression appears in the text. The language model adaptation device according to any one of claims 1 to 4, wherein the dividing unit divides the text using a model representing an appearance tendency of words related to a preset topic. A speech recognition device that performs speech recognition by adapting a language model,
A speech recognition unit that performs speech recognition on speech data using the language model;
A dividing unit for dividing the text obtained by the speech recognition into a plurality of sections;
Determining a topic included in each of the plurality of sections, and for each determined topic, a topic analysis unit that creates a topic model representing an appearance tendency of words in the topic;
For each of the plurality of sections, using the topic model corresponding to the topic included in the section, adapting the base language model, and creating an adapted language model, ,
A re-recognition unit that performs speech recognition on the speech data in the section corresponding to the adaptation language model using the adaptation language model created by the adaptation language model creation unit. A featured voice recognition device. A language model adaptation method for adapting a base language model,
(A) dividing the text into a plurality of sections;
(B) determining a topic included in each of the plurality of sections, and creating a topic model representing the appearance tendency of words in the topic for each determined topic;
(C) For each of the plurality of sections, using the topic model corresponding to the topic included in the section, adapting the base language model and creating an adapted language model; A language model adaptation method characterized by comprising: (D) further comprising performing speech recognition on speech data using the base language model, thereby generating a text to be divided in the step (a). The language model adaptation method according to claim 10. In the step (b), the plurality of sections are classified into groups based on the similarity between the texts of the sections, and it is determined that the topics of the sections belonging to the same group are common. Every time, based on the appearance tendency of the words in the text in the section belonging to the group, create the topic model,
The language model adaptation method according to claim 10 or 11. In the step (b), using the topic as a hidden variable, a probability model representing the appearance tendency of words in each of the plurality of sections is assumed, and further, the text of each of the plurality of sections is used as learning data, Learn the parameters of the probabilistic model,
The learning determines the topic and creates the topic model.
The language model adaptation method according to claim 10 or 11. 14. The language model adaptation method according to claim 10, wherein in the step (a), the text is divided based on the number of words or the number of utterances. The language model adaptation method according to any one of claims 10 to 13, wherein, in the step (a), a change point of a word distribution in the text is detected, and the text is divided based on the detected change point. . 14. The language model adaptation method according to claim 10, wherein in the step (a), the text is divided based on a position where a preset topic boundary expression appears in the text. The language model adaptation method according to any one of claims 10 to 13, wherein, in the step (a), the text is divided using a model representing an appearance tendency of words related to a preset topic. A computer-readable recording medium recording a program for executing adaptation of a language model as a base by a computer,
In the computer,
(A) dividing the text into a plurality of sections;
(B) determining a topic included in each of the plurality of sections, and creating a topic model representing the appearance tendency of words in the topic for each determined topic;
(C) For each of the plurality of sections, performing the step of adapting the base language model using the topic model corresponding to the topic included in the section and creating an adapted language model A computer-readable recording medium on which a program including instructions is recorded. (D) Performing speech recognition on speech data using the base language model, thereby causing the computer to further execute a step of generating text to be divided in the step (a). The computer-readable recording medium according to claim 18. In the step (b), the plurality of sections are classified into groups based on the similarity between the texts of the sections, and it is determined that the topics of the sections belonging to the same group are common. Every time, based on the appearance tendency of the words in the text in the section belonging to the group, create the topic model,
The computer-readable recording medium according to claim 18 or 19. In the step (b), using the topic as a hidden variable, a probability model representing the appearance tendency of words in each of the plurality of sections is assumed, and further, the text of each of the plurality of sections is used as learning data, Learn the parameters of the probabilistic model,
The learning determines the topic and creates the topic model.
The computer-readable recording medium according to claim 18 or 19. The computer-readable recording medium according to any one of claims 18 to 21, wherein in the step (a), the text is divided based on the number of words or the number of utterances. The computer-readable program according to any one of claims 18 to 21, wherein in the step (a), a change point of a word distribution in the text is detected, and the text is divided based on the detected change point. recoding media. The computer-readable recording medium according to any one of claims 18 to 21, wherein in the step (a), the text is divided based on a position where a preset topic boundary expression appears in the text. . The computer-readable recording medium according to any one of claims 18 to 21, wherein in the step (a), the text is divided using a model representing an appearance tendency of a word related to a preset topic. .

Images