JP5762365B2 - Speech recognition apparatus, speech recognition method, and program - Google Patents

Description

  The present invention relates to a speech recognition technique for converting utterance content in input speech into text.

  Speech recognition is a media processing technology that converts speech to text. Speech recognition is processed using both an acoustic model that defines the characteristics of a phoneme sequence for an input speech and a language model that defines the likelihood of the word order as a word sequence.

  When speech recognition is performed on dialogue speech by two speakers, the speech recognition results for each speaker can be obtained by performing speech recognition independently on speech separated separately for each speaker. Can be obtained. A conventional technology for performing speech recognition of dialogue speech is disclosed in Patent Document 1, for example. The speech recognition apparatus described in Patent Document 1 performs speech recognition using a feature value obtained from a speech signal including the utterance content of one speaker A, an acoustic model, and a language model before adaptation, and the recognition result A ′. The language model after adaptation is obtained using the language model before adaptation and speech recognition is performed using the feature amount and acoustic model obtained from the speech signal including the speech content of the other speaker B and the language model after adaptation. The recognition result B ′ is obtained.

  In the conventional voice recognition technology for dialogue voice, when the dialogue voice in a specific scene is recognized, the recognition accuracy of one speaker may be high and the recognition accuracy of the other speaker may be low. For example, when recognizing dialogue voices between an operator and a customer at a call center, etc., the operator is polite and the voice recording status is good, so the recognition rate is high, but the customer is rough and the voice recording status is bad. The recognition rate tends to be low.

  A technique described in Non-Patent Document 1 is considered as a solution to this problem. Non-Patent Document 1 describes a method of improving the performance of a language model by unsupervised language model adaptation for the purpose of improving speech recognition accuracy. The unsupervised language model adaptation method described in Non-Patent Document 1 performs processing as follows. First, speech recognition is performed using a baseline language model created from a topic-independent baseline corpus. Next, a keyword is extracted from the word that appears in the recognition result, and a search query is constructed from the keyword. The search query is used to acquire a WWW page from the Internet using a search engine, and a WWW corpus is created by filtering and shaping text from the acquired WWW page. Then, an adaptive language model is newly created from the baseline corpus and the WWW corpus, and input speech is recognized using the adaptive language model.

JP 2011-107314 A

Ryo Masumura, Hitoshi Ito, Akinori Ito, Shozo Makino, “Language Model Unsupervised Adaptation Effect Using Valid Search Queries Using WWW”, Proceedings of the 2009 Autumn Meeting of the Acoustical Society of Japan, 2-1-1, 2009.

  In the method described in Non-Patent Document 1, if the recognition accuracy based on the baseline language model is low, the extraction accuracy of keywords used for adaptation of the language model also decreases, and thus the adaptation effect may not be obtained. When many recognition errors are included in the recognition result by the baseline language model, the recognition accuracy by the adaptive language model may be lowered.

  In addition, even when talking about the same topic, the tendency of words appearing between speakers may be different. For example, in the call center example described above, when the operator is talking about “address change”, the customer may only say “moving”. As a result, the recognition result of the operator has high recognition accuracy, and even when language model adaptation is performed using the recognition result, sufficient accuracy may not be obtained by customer speech recognition.

  The present invention has been made in view of the above points, and in the case of dialogue speech recognition, when the recognition rate of one speaker is high and the recognition rate of the other speaker is low, An object of the present invention is to provide a speech recognition technique capable of improving the recognition rate.

  In order to solve the above problems, the speech recognition apparatus of the present invention groups a plurality of speakers in advance, and a conversation between a first speaker belonging to one group and a second speaker belonging to the other group. Recognize speech. The speech recognition apparatus includes a dialogue corpus storage unit, a language model storage unit, an acoustic model storage unit, a first recognition processing unit, a translation topic model generation unit, a word occurrence probability calculation unit, a language model adaptation unit, and a second recognition processing unit. . The dialogue corpus storage unit stores a dialogue corpus in which conversation speech between a speaker belonging to one group and a speaker belonging to the other group is transcribed. The language model storage unit stores a language model. The acoustic model storage unit stores an acoustic model. The first recognition processing unit recognizes the voice of the first speaker using the acoustic model and the language model and obtains a first speaker recognition result. The translation topic model generation unit obtains a translation topic model for converting the utterance content of the first speaker into the utterance content of the second speaker from the dialogue corpus. The word occurrence probability calculation unit calculates the word occurrence probability of the second speaker using the first speaker recognition result and the translation topic model. The language model adaptation unit obtains an adaptive language model from the language model using the word occurrence probability. The second recognition processing unit recognizes the voice of the second speaker using the acoustic model and the adaptive language model, and obtains a second speaker recognition result.

  By giving language constraints that take advantage of the characteristics of dialogue, the performance of the language model is improved and the speech recognition accuracy is improved. In particular, when the recognition rate of one speaker is high but the recognition rate of the other speaker is low, by providing language restrictions based on the utterance content of the speaker with a high recognition rate, Recognition rate is improved.

The figure which illustrates the function structure of the speech recognition apparatus which concerns on 1st Embodiment. The figure which illustrates the processing flow of the speech recognition apparatus which concerns on 1st Embodiment. The figure which illustrates the function structure of the speech recognition apparatus which concerns on 2nd Embodiment. The figure which illustrates the processing flow of the speech recognition apparatus which concerns on 2nd Embodiment. The figure which illustrates the functional structure of the speech recognition apparatus which concerns on 3rd Embodiment. The figure which illustrates the processing flow of the speech recognition apparatus which concerns on 3rd Embodiment. The figure which illustrates the function structure of the speech recognition apparatus which concerns on a modification. The figure which illustrates the processing flow of the speech recognition apparatus which concerns on a modification. The figure which illustrates the function structure of the speech recognition apparatus which concerns on 4th Embodiment. The figure which illustrates the processing flow of the speech recognition apparatus which concerns on 4th Embodiment.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the component which has the same function in drawing, and duplication description is abbreviate | omitted.

[Points of Invention]
Prior to the description of the embodiments, the points of the present invention will be described.

  In conventional speech recognition technology, language restrictions are generally given using a language model expressed in N-gram, but language restrictions that take into consideration the characteristics of dialogue are not given. It was.

  In the dialogue, there is a strong relationship between the utterance contents of each other, and it is considered that the utterance contents of the other speaker are limited by receiving the utterance contents of one speaker. In consideration of this point, in the present invention, when recognizing the dialog voice, the speech content of one speaker is used to give a language restriction on the speech content of the other speaker.

  In that case, the N-gram probability that appears in the recognition result of one speaker is not directly applied to the language model, but the translation method is used to convert the N-gram probability that appears in the recognition result of one speaker to the other. It is converted to the N-gram probability of the speaker and adapted to the language model. Thereby, it becomes possible to generate a language model of a speaker with low recognition accuracy from the recognition result of the speaker with high recognition accuracy, and an effect of improving the recognition accuracy for a speaker with low recognition accuracy can be obtained.

[First Embodiment]
<Overview>
The speech recognition apparatus 10 according to the first embodiment of the present invention recognizes dialogue speech between the first speaker A having a high recognition rate and the second speaker B having a low recognition rate, and the first speaker recognition result A ′. And the second speaker recognition result B ′ is output. When applied to the example used in the call center described above, the first speaker A corresponds to the operator and the second speaker B corresponds to the customer.

<Configuration>
With reference to FIG. 1, the structural example of the speech recognition apparatus 10 which concerns on this embodiment is demonstrated. The speech recognition apparatus 10 includes input terminals 101A and 101B, speech signal acquisition units 110A and 110B, feature amount analysis units 120A and 120B, a first recognition processing unit 130A, a second recognition processing unit 130B, a translation topic model generation unit 140, and words. An occurrence probability calculation unit 150, a language model adaptation unit 160, a dialogue corpus storage unit 210, a translation topic model storage unit 220, a language model storage unit 230, an acoustic model storage unit 240, and an adaptive language model storage unit 250 are provided.

  The dialogue corpus storage unit 210, the translation topic model storage unit 220, the language model storage unit 230, the acoustic model storage unit 240, and the adaptive language model storage unit 250 are, for example, a main storage device such as a RAM (Random Access Memory), a hard disk, or an optical disk. Alternatively, it can be constituted by an auxiliary storage device constituted by a semiconductor memory element such as a flash memory.

<Voice recognition processing>
With reference to FIG. 2, an operation example of the speech recognition apparatus 10 will be described in the order of procedures actually performed.

  The dialogue corpus storage unit 210 stores a dialogue corpus in advance. The conversation corpus is a conversation between speakers who belong to one group and speakers who belong to the other group for the speakers grouped according to their roles in the conversation in advance (in the call center example, operator group and customer group). It is a parallel corpus consisting of voice transcription pairs (in the call center example, a transcription pair of an operator-customer conversation). A parallel corpus is called a bilingual corpus in the field of translation, and is a corpus in which two sentences written in different languages are combined in the form of a bilingual translation.

  The language model storage unit 230 stores language models in advance. The language model may be a general-purpose language model, or may be a language model specially constructed for use situations such as for a call center. If a language model specialized for the use scene is used, the recognition result of the speech of the first speaker A will be higher. If the recognition rate of the speech recognition result of the first speaker A is high, the adaptation effect of the language model using this will increase, and the speech recognition rate of the second speaker B using the adaptive language model will also increase. It is done.

  The acoustic model storage unit 240 stores an acoustic model in advance. Various modeling methods have been proposed for acoustic models. Since it depends on a voice recognition processing method to be described later, an acoustic model to be used may be determined corresponding to the voice recognition processing method to be used.

  The analog voice signal of the first speaker A is input to the voice signal acquisition unit 110A via the input terminal 101A. The first speaker A is a speaker belonging to one group in which a plurality of speakers are grouped in advance according to their roles in dialogue. When applied to the usage scene in the call center described above, the first speaker A corresponds to the operator. The voice signal acquisition unit 110A converts the input analog voice signal of the first speaker A into a digital voice signal (S110A).

  The digital voice signal of the first speaker A is input to the feature amount analysis unit 120A. The feature amount analysis unit 120A extracts an acoustic feature amount from the input digital voice signal of the first speaker A (S120A). As the feature quantity to be extracted, for example, 1 to 12 dimensions of MFCC (Mel-Frequent Cepstrum Coefficient) and dynamic parameters such as ΔMFCC which is the change amount, power, Δ power, and the like are used. Also, CMN (cepstrum average normalization) processing may be performed. The feature amount is not limited to MFCC or power, and a parameter used for speech recognition may be used. For details on the method for extracting the feature quantity, refer to, for example, “Sadaaki Furui,“ Acoustic / Voice Engineering ”, Modern Science, September 1992 (reference 1)”.

  The feature amount of the voice of the first speaker A is input to the first recognition processing unit 130A. 130 A of 1st recognition process parts use the language model memorize | stored in the language model memory | storage part 230, and the acoustic model memorize | stored in the acoustic model memory | storage part 240, and the feature-value of the voice of the 1st speaker A is obtained. Speech recognition is performed, and a first speaker recognition result A ′ is generated (S130A). As a specific recognition processing method, for example, the method described in Reference 1 can be applied.

The translation topic model generation unit 140 obtains a translation topic model for converting the utterance content of the first speaker A into the utterance content of the second speaker B from the dialogue corpus stored in the dialogue corpus storage unit 210 (S140). . Specifically, the translation topic model is composed of the following three probability values.
1. Occurrence probability P (w _{a, n} | z _i ) of word w _n under topic z _i in first speaker A
2. Occurrence probability P (w _{b, n} | z _i ) of word w _n under topic z _i in second speaker B
3. Occurrence probability P (z _i ) of topic z _i

  Many methods for generating a translation topic model have been proposed. For example, a Polylingual Topic Model (PLTM) can be applied.

  The generated translation topic model is stored in the translation topic model storage unit 220. However, the translation topic model may not necessarily be stored in the translation topic model storage unit 220, and the translation topic model generated by the translation topic model generation unit 140 may be directly input to the word occurrence probability calculation unit 150. .

The first speaker recognition result A ′ generated by the first recognition processing unit 130A is input to the word occurrence probability calculating unit 150. The word occurrence probability calculation unit 150 uses the first speaker recognition result A ′ and the translation topic model stored in the translation topic model storage unit 220 (or input from the translation topic model generation unit 140) to The word occurrence probability of the speaker B is calculated (S150). The word occurrence probability P (w _{b, n} | d _a ) of the second speaker B can be calculated by the following equation.

Here, P (z _i | d _a ) is the topic probability in the document, and the occurrence probability of topic z _i under the first speaker recognition result d _a . Many methods for calculating the topic probability in a document have been proposed. For example, Probabilistic Latent Semantic Analysis (PLSA) or Latent Dirichlet Allocation (LDA) can be applied.

The word occurrence probability P (z _i , d _a ) of the second speaker B is input to the language model adaptation unit 160. The language model adaptation unit 160 generates an adaptation language model from the language model stored in the language model storage unit 230 using the word occurrence probability P (z _i , d _a ) (S160). Many methods for adapting language models have been proposed. For example, Unigram Rescaling can be applied. Unigram Rescaling is a method of adapting the unigram model to the N-gram model. Unigram Rescaling makes it possible to estimate the N-gram probability from the word occurrence probability P (z _i , d _a ).

Unigram Rescaling is explained in more detail. The principle of this language model adaptation method is based on the idea that the ratio of the N-gram probability between the base language model before adaptation and the language model after adaptation can be approximated to the ratio of unigram probabilities. The N-gram probability after adaptation is calculated. Unigram Rescaling can generally estimate N-gram probabilities from unigram probabilities, but in the field of speech recognition, a language model of trigram probabilities is mainly used, so here a method for estimating trigram probabilities will be described. The trigram probability of the base language model is P (w _n | w _n-2 , w _n-1 ), the unigram probability of the base language model is P (w _n ), and the trigram probability of the adaptive language model is P_d _a (w _n | w _n-2 , w _n-1 ), the unigram probability of the adaptive language model is P (w _n | d _a ), and the Unigram Rescaling coefficient is α. The trigram probability P_d _a (w _n | w _n−2 , w _n−1 ) of the adaptive language model can be calculated by the following equation.

  The generated adaptive language model is stored in the adaptive language model storage unit 250. However, it is not always necessary to store the adaptive language model in the adaptive language model storage unit 250, and the adaptive language model generated by the language model adaptation unit 160 may be directly input to the second recognition processing unit 130B.

  The analog voice signal of the second speaker B is input to the voice signal acquisition unit 110B via the input terminal 101B. The second speaker B is a speaker belonging to a group different from the group to which the first speaker A belongs. When applied to the use scene in the call center described above, the second speaker B corresponds to the customer. The audio signal acquisition unit 110B converts the input analog audio signal of the second speaker B into a digital audio signal (S110B).

  The digital voice signal of the second speaker B is input to the feature amount analysis unit 120B. The feature amount analysis unit 120B extracts an acoustic feature amount from the input digital voice signal of the second speaker B (S120B). Since the feature amount to be extracted is the same as the feature amount extracted by the feature amount analysis unit 120A, description thereof is omitted here.

  The feature amount of the voice of the second speaker B is input to the second recognition processing unit 130B. The second recognition processing unit 130B uses the adaptive language model stored in the adaptive language model storage unit 250 (or input from the language model adaptation unit 160) and the acoustic model stored in the acoustic model storage unit 240. Then, the feature amount of the voice of the second speaker B is voice-recognized, and the second speaker recognition result B ′ is generated (S130B). As a specific recognition processing method, for example, the method described in Reference 1 can be applied in the same manner as the first recognition processing unit 130A.

<Effect>
As described above, the speech recognition apparatus 10 according to the first embodiment estimates a topic using the first speaker recognition result A ′ obtained by recognizing the voice of the first speaker A, and generates a translation topic model generated from the dialogue corpus. Using, the word occurrence probability of the second speaker B under the topic is calculated. The language model is adapted based on the calculated word occurrence probability, and the voice of the second speaker B is recognized using the adapted language model. That is, when the speech of the second speaker B having a low recognition rate is recognized, the speech recognition rate of the second speaker B is improved by giving the conversation topic as a language constraint.

  In particular, when recognizing the conversation voice between the operator and the customer in the call center, the recording environment of the customer voice is bad and the effect of the acoustic model cannot be expected. In such a case, the recognition rate is low with the conventional speech recognition technology, but according to the present invention, an improvement in the recognition rate can be expected.

<Others>
Note that the voice recognition device 10 may receive a digital voice signal instead of an analog voice signal from the input terminals 101A and 101B. In addition, a digital audio signal can be received from an external storage medium or a communication device. In this case, the voice recognition device 10 may not include the voice input terminals 101A and 101B and the voice signal acquisition units 110A and 110B.

  In this embodiment, the case of recognizing the conversation voice between the operator and the customer in the call center has been described. However, other conversation voices or conversation voices may be used. If there are more than two speakers, the speaker with the highest recognition rate based on the language model before adaptation (for example, a speaker with a sound collection environment, speaking speed, words, The first speaker A is a group of speakers with appropriate grammar, etc., and the speaker is expected to have a low recognition rate based on the language model before adaptation (for example, a speaker in a sound collecting environment with a lot of noise, etc.) , A group of a speaker who speaks fast, a speaker whose word or grammar is used, etc.) is set as the second speaker B, and the language model is adapted to the conversation content in the same manner as this embodiment in consideration of the characteristics of the conversation. Can be adapted.

  Note that the speech recognition apparatus 10 does not necessarily output the first speaker recognition result A ′. For example, when it is desired to record only the utterance content of the customer as text data in the call center, only the second speaker recognition result B 'may be output and stored.

[Second Embodiment]
<Overview>
In general, speech recognition results include words representing tone, speaker characteristics, and the like in addition to words related to a topic. In the first embodiment, the language model adaptation is performed using the entire speech recognition result including the words representing the tone and speaker characteristics. With this configuration, the first speaker A and the second story are configured. When the tone and speaker characteristics are different between the person B and the person B, an incorrect adaptation effect may be obtained. Therefore, in the second embodiment, adaptation is performed only for keywords determined in advance when language model adaptation is performed. By configuring in this way, it is possible to prevent erroneous adaptation, and it is possible to improve the recognition accuracy of only predetermined keywords.

<Configuration>
A configuration example of the speech recognition apparatus 20 according to this embodiment will be described with reference to FIG. Similar to the speech recognition device 10 according to the first embodiment, the speech recognition device 20 includes input terminals 101A and 101B, speech signal acquisition units 110A and 110B, feature amount analysis units 120A and 120B, a first recognition processing unit 130A, and a first recognition processing unit 130A. 2 recognition processing unit 130B, translation topic model generation unit 140, word occurrence probability calculation unit 150, language model adaptation unit 160, dialogue corpus storage unit 210, translation topic model storage unit 220, language model storage unit 230, and acoustic model storage unit 240 And an adaptive language model storage unit 250. The speech recognition apparatus 20 further includes keyword extraction units 170A and 170B and a keyword storage unit 260.

  The keyword storage unit 260 can be configured by, for example, a main storage device such as a RAM (Random Access Memory), an auxiliary storage device configured by a semiconductor memory element such as a hard disk, an optical disk, or a flash memory.

<Voice recognition processing>
With reference to FIG. 4, the operation example of the speech recognition apparatus 20 will be described in the order of procedures actually performed.

  The keyword storage unit 260 stores a predetermined keyword. The keyword is preferably a word related to the topic. For example, “noun / verb / adjective” is a keyword using the part-of-speech information, and words that are highly likely to be unrelated to the topic, such as particles and auxiliary verbs, may be excluded from the keyword.

A keyword such as tf-idf may be used for keyword selection. That is, for a word included in the first speaker recognition result A ′, the number of occurrences of the word in the first speaker recognition result is tf _{i, j} and the word is included using the dialog text included in the dialog corpus as a document. The tf-idf value is calculated by the following formula, where df _i is the number of documents and N is the total number of documents.

  A word with a high calculated tf-idf value is selected as a keyword. As a keyword selection criterion, a predetermined number of words may be selected in descending order of tf-idf values, or all words having a predetermined tf-idf value or more may be selected as keywords. Words that appear in many dialogue texts are considered common words and are therefore likely not important in relation to topics. By selecting a keyword based on the tf-idf value, there is an effect that only words related to the topic can be obtained as keywords.

  The keyword extraction unit 170A extracts a predetermined keyword stored in the keyword storage unit 260 from the dialogue corpus stored in the dialogue corpus storage unit 210, and outputs a keyword string (S170A). The extracted keyword string is output to the translation topic model generation unit 140. The translation topic model generation unit 140 obtains a translation topic model using the keyword string output by the keyword extraction unit 170A instead of the dialogue corpus (S140).

  The keyword extraction unit 170B extracts a predetermined keyword stored in the keyword storage unit 260 from the first speaker recognition result A ′ output from the first recognition processing unit 130A, and outputs a keyword string ( S170B). The extracted keyword string is output to the word occurrence probability calculation unit 150. The word occurrence probability calculation unit 150 calculates the word occurrence probability of the second speaker B using the keyword string output by the keyword extraction unit 170B instead of the first speaker recognition result A ′ (S150).

<Effect>
As described above, the speech recognition apparatus 20 according to the second embodiment extracts a predetermined keyword from the first speaker recognition result A ′ obtained by recognizing the voice of the first speaker A, and extracts only the extracted keyword. The word occurrence probability is calculated for. By configuring in this way, it is possible to prevent erroneous adaptation from being performed by a word representing a tone or speaker characteristics not related to the topic included in the first speaker recognition result A ′. . As a result, it is possible to improve the recognition accuracy of only predetermined keywords.

[Third Embodiment]
<Overview>
A speech recognition result always includes a recognition error. In the first embodiment, language model adaptation is performed using the entire first speaker recognition result, so it is conceivable that words due to recognition errors will adversely affect language model adaptation. Therefore, in the third embodiment, when performing speech recognition, together with the recognition result, a recognition reliability indicating the certainty of the recognition result is output, and only a word having a high recognition reliability is set as a word to be applied. With this configuration, the influence of recognition errors on language model adaptation can be suppressed, and the recognition accuracy can be further improved.

<Configuration>
With reference to FIG. 5, a configuration example of the speech recognition apparatus 30 according to this embodiment will be described. Similar to the speech recognition device 10 according to the first embodiment, the speech recognition device 30 includes input terminals 101A and 101B, speech signal acquisition units 110A and 110B, feature amount analysis units 120A and 120B, and a second recognition processing unit 130B. Topic model generation unit 140, word occurrence probability calculation unit 150, language model adaptation unit 160, dialogue corpus storage unit 210, translation topic model storage unit 220, language model storage unit 230, acoustic model storage unit 240, and adaptive language model storage unit 250 With. The voice recognition device 30 further includes a high reliability word extraction unit 180. Further, a first recognition processing unit 131A is provided instead of the first recognition processing unit 130A.

<Voice recognition processing>
With reference to FIG. 6, the operation example of the speech recognition apparatus 30 will be described in the order of procedures actually performed.

  The first recognition processing unit 131A uses the language model stored in the language model storage unit 230 and the acoustic model stored in the acoustic model storage unit 240 to output the first episode from the feature amount analysis unit 120A. The feature amount of the voice of the person A is recognized as a voice, and the first speaker recognition result A ′ and the corresponding recognition reliability are generated (S131A). The recognition reliability is a measure representing how much the recognition result output by the speech recognition process can be trusted, and is generally calculated based on the posterior probability. The value ranges from 0.0 to 1.0. A value closer to 1.0 indicates that there were almost no other competitive candidates for the final recognition result, and a value closer to 0.0 indicated that many other candidates with similar scores appeared. Show.

  The recognition reliability is input to the high reliability word extraction unit 180. The high reliability word extraction unit 180 extracts only words having a higher recognition reliability than a predetermined threshold from the first speaker recognition result A ′ (S180). The threshold used here is set to a value at which the recognition rate of the second speaker's voice is highest using the evaluation set. As described above, since the recognition reliability is in the range of 0.0 to 1.0, the recognition accuracy of the second speaker included in the evaluation set is tried for each of a plurality of different values in this range, and the recognition accuracy is the best. May be adopted as a threshold value. The evaluation set is composed of dialogue voices between the speakers of the group to which the first speaker belongs and the speakers of the group to which the second speaker belongs, and a correct sentence that transcribes the dialogue voice. The difference between the evaluation set and the dialogue corpus is that the evaluation set consists of a dialogue voice and a correct sentence, but the dialogue corpus consists only of the correct sentence, and the dialogue corpus requires a large number of cases, but the evaluation set It is a point that does not need as many cases as the dialogue corpus. Therefore, it is only necessary to collect a large amount of dialogue speech, use only the correct sentence that transcribes the dialogue speech as a dialogue corpus, and set the subset of the dialogue speech and the corresponding correct sentence as an evaluation set. The extracted high reliability word is input to the word occurrence probability calculation unit 150. The word occurrence probability calculation unit 150 calculates the word occurrence probability of the second speaker B using the high reliability word output from the high reliability word extraction unit 180 instead of the first speaker recognition result A ′ ( S150).

<Effect>
As described above, the speech recognition apparatus 30 according to the third embodiment extracts words with high recognition reliability from the first speaker recognition result A ′ that recognizes the voice of the first speaker A, and extracts only the extracted words. The word occurrence probability is calculated for. By comprising in this way, the influence which the recognition error contained in 1st speaker recognition result A 'has on language model adaptation can be suppressed. As a result, it can be expected that the recognition accuracy is further improved as compared with the speech recognition apparatus 10 according to the first embodiment.

[Modification]
<Overview>
It is also possible to apply the idea of the third embodiment to the voice recognition device 20 of the second embodiment. That is, the language model is applied only to a predetermined keyword among words having high recognition reliability. By configuring in this way, the effects of words that represent a tone or speaker characteristics that have high recognition reliability but are not related to the topic, and the effects of words that coincide with a predetermined keyword due to recognition errors, etc. Can be eliminated.

<Configuration>
With reference to FIG. 7, the structural example of speech recognition apparatus 30 'which concerns on this modification is demonstrated. Similar to the speech recognition device 20 according to the second embodiment, the speech recognition device 30 ′ includes input terminals 101A and 101B, speech signal acquisition units 110A and 110B, feature amount analysis units 120A and 120B, and a second recognition processing unit 130B. Translation topic model generation unit 140, word occurrence probability calculation unit 150, language model adaptation unit 160, keyword extraction units 170A and 170B, dialogue corpus storage unit 210, translation topic model storage unit 220, language model storage unit 230, and acoustic model storage unit 240, an adaptive language model storage unit 250, and a keyword storage unit 260. The speech recognition device 30 ′ further includes a high reliability word extraction unit 180. Further, a first recognition processing unit 131A is provided instead of the first recognition processing unit 130A.

<Voice recognition processing>
With reference to FIG. 8, an operation example of the speech recognition apparatus 30 ′ will be described in the order of procedures actually performed.

  The first recognition processing unit 131A uses the language model and the acoustic model to recognize the feature amount of the voice of the first speaker A, and generates the first speaker recognition result A ′ and the corresponding recognition reliability. (S131A). The recognition reliability is input to the high reliability word extraction unit 180. The high reliability word extraction unit 180 extracts only words having a higher recognition reliability than a predetermined threshold from the first speaker recognition result A ′ (S180). The extracted high reliability word is input to the keyword extraction unit 170B. The keyword extraction unit 170B extracts a predetermined keyword stored in the keyword storage unit 260 from the high reliability word, and outputs a keyword string (S170B). The extracted keyword string is output to the word occurrence probability calculation unit 150. The word occurrence probability calculation unit 150 calculates the word occurrence probability of the second speaker using the keyword string extracted from the high reliability word and the translation topic model (S150).

<Effect>
As described above, the speech recognition device 30 ′ according to the modification extracts a predetermined keyword from only words having high recognition reliability from the first speaker recognition result A ′ that recognizes the voice of the first speaker A. The word occurrence probability is calculated only for the extracted keywords. By configuring in this way, the influence of words that express tone and speaker characteristics that are high in recognition reliability but are not related to the topic, or the influence of words that accidentally match a predetermined keyword due to recognition errors, etc. Can be eliminated. As a result, the recognition accuracy can be further improved as compared with the speech recognition apparatus 20 according to the second embodiment.

[Fourth Embodiment]
<Overview>
There may be a case where it is not known in advance which voice of the two parties included in the conversation voice has high recognition accuracy. Also, there are cases where the recognition rate of any speech is sufficiently high and language model adaptation is not required, or the recognition rate of any speech is comparable and the language model adaptation is not required. Therefore, in the fourth embodiment, each recognition accuracy is estimated from a provisional recognition result obtained by recognizing both voices in advance, whether or not language model adaptation is necessary, or which voice is the voice of the first speaker A, Is determined to be the voice of the second speaker B.

<Configuration>
With reference to FIG. 9, the structural example of the speech recognition apparatus 40 which concerns on this embodiment is demonstrated. The voice recognition device 40 includes provisional recognition processing units 135A and 135B and reliability estimation units 190A and 190B in addition to the configurations of the voice recognition devices 10, 20, 30, and 30 ′ shown in the first to third embodiments and the modified examples. And an adaptive processing determination unit 195.

<Voice recognition processing>
With reference to FIG. 10, the operation example of the speech recognition apparatus 40 will be described in the order of procedures actually performed.

  An analog voice signal of the speaker X is input to the voice signal acquisition unit 110A via the input terminal 101A. The speaker X is a speaker belonging to any group in which a plurality of speakers are grouped in advance according to the role of dialogue. The voice signal acquisition unit 110A converts the input analog voice signal of the speaker X into a digital voice signal (S110A).

  An analog voice signal of the speaker Y is input to the voice signal acquisition unit 110B via the input terminal 101B. The speaker Y is a speaker belonging to a group different from the group to which the speaker X belongs. The voice signal acquisition unit 110B converts the input analog voice signal of the speaker Y into a digital voice signal (S110B).

  The digital audio signal of the speaker X is input to the feature amount analysis unit 120A. The feature amount analysis unit 120A extracts an acoustic feature amount from the input digital audio signal of the speaker X (S120A). Since the feature quantity to be extracted is the same as in the other embodiments described above, description thereof is omitted here.

  The digital voice signal of the speaker Y is input to the feature amount analysis unit 120B. The feature amount analysis unit 120B extracts an acoustic feature amount from the input digital voice signal of the speaker Y (S120B). Since the feature quantity to be extracted is the same as in the other embodiments described above, description thereof is omitted here.

  The feature amount of the voice of the speaker X is input to the provisional recognition processing unit 135A. The provisional recognition processing unit 135 </ b> A recognizes the feature amount of the speaker X using the language model stored in the language model storage unit 230 and the acoustic model stored in the acoustic model storage unit 240. The provisional recognition result and the corresponding recognition reliability are generated (S135A). Since the specific recognition processing method and the recognition reliability estimation method are the same as those of the first recognition processing unit 131A in the above-described embodiment, description thereof is omitted here.

  The feature amount of the voice of the speaker Y is input to the provisional recognition processing unit 135B. The provisional recognition processing unit 135 </ b> B recognizes the feature amount of the voice of the speaker Y using the language model stored in the language model storage unit 230 and the acoustic model stored in the acoustic model storage unit 240. The provisional recognition result and the corresponding recognition reliability are generated (S135B). Since the specific recognition processing method and the recognition reliability estimation method are the same as those of the provisional recognition processing unit 135A, description thereof is omitted here.

  The recognition reliability of the speaker X is input to the reliability estimation unit 190A. The reliability estimation unit 190A estimates the document reliability of the provisional recognition result of the speaker X using the recognition reliability of the speaker X (S190A). The document reliability is the recognition reliability in units of documents. In the speech recognition of dialogue voice, the document is one dialogue. For example, Taichi Asami, Satoshi Kobashikawa, Yoshikazu Yamaguchi, Hirokazu Masatoshi, Satoshi Takahashi, “Using word context consistency and acoustic likelihood” "Establishment of speech document recognition reliability", IEICE Technical Report, vol. 110, no. 143, SP2010-42, pp. 43-48, July 2010.

  The recognition reliability of the speaker Y is input to the reliability estimation unit 190B. The reliability estimation unit 190B estimates the document reliability of the provisional recognition result of the speaker Y using the recognition reliability of the speaker Y (S190B). Since the document reliability estimation method is the same as that of the reliability estimation unit 190A, description thereof is omitted here.

  The document reliability of the speaker X and the document reliability of the speaker Y are input to the adaptive processing determination unit 195. The adaptive processing determination unit 195 uses the document reliability of the speaker X and the document reliability of the speaker Y as a voice of the first speaker A from the voice of the speaker X and the voice of the speaker Y. , Which of the voices of the second speaker B is determined is determined (S195). At this time, if it is determined that it is not necessary to perform the language model adaptation process, the temporary recognition result is output as the first speaker recognition result A ′ or the second speaker recognition result B ′.

  The rule of the determination method which the adaptive process determination part 195 performs is illustrated. First, when both the document reliability of speaker X and the document reliability of speaker Y are larger than a predetermined threshold, the language model adaptation process is not performed. This is because the recognition rate is already high and the adaptation effect of the language model cannot be expected, and the processing cost is considered to be wasted. Next, if both the document reliability of speaker X and the document reliability of speaker Y are smaller than the threshold, the language model adaptation process is not performed. This is because if the recognition rate is low, language model adaptation is performed based on an erroneous recognition result, so that the adaptation effect cannot be expected, and the recognition accuracy may even deteriorate. Further, when the document reliability of the speaker X is higher than the threshold and the document reliability of the speaker Y is lower than the threshold, the speaker X is the first speaker A and the speaker Y is the second speaker B. Perform model adaptation processing. If the document reliability of speaker X is smaller than the threshold and the document reliability of speaker Y is greater than the threshold, speaker Y is first speaker A and speaker X is second speaker B. Perform model adaptation processing.

  The threshold used for the determination is set to a value at which the recognition rate of the voice of the speaker X and the voice of the speaker Y is highest using the evaluation set. Since the document reliability is a recognition reliability in document units, and the recognition reliability is in the range of 0.0 to 1.0 as described above, the speaker X included in the evaluation set for each of a plurality of different values in this range. May be used as the threshold value, and a value that gives the best recognition accuracy may be used. The same evaluation set as that used when determining the threshold value of the highly reliable word extraction unit 180 in the third embodiment may be used as the evaluation set. An evaluation set different from the evaluation set used when determining the threshold value of the high reliability word extraction unit 180 may be separately prepared.

  When it is determined that the language model adaptation process needs to be performed in the adaptation process determination unit 195 and the voice of the first speaker A and the voice of the second speaker B are determined, the process described in the above embodiment is performed. . In this case, the processing cost can be reduced by configuring the temporary recognition result to be reused as the first speaker recognition result A ′.

<Effect>
As described above, the voice recognition device 40 according to the fourth embodiment recognizes each of the two voices included in the dialogue voice even if the voice recognition accuracy is not known in advance. From the recognition result, it is determined whether language model adaptation is necessary, and if necessary, which voice is designated as the first speaker A and which voice is designated as the second speaker B. By configuring in this way, unnecessary language model adaptation can be omitted and the processing cost can be reduced. In addition, it is possible to prevent the recognition accuracy from deteriorating by applying the language model adaptation by mistakenly setting the speaker with a low recognition rate as the first speaker A and the speaker with a high recognition rate as the second speaker B.

[Program, recording medium]
The present invention is not limited to the above-described embodiment, and it goes without saying that modifications can be made as appropriate without departing from the spirit of the present invention. The various processes described in the above-described embodiments are not only executed in time series according to the order described, but may be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes.

  When various processing functions in each device described in the above embodiment are realized by a computer, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on a computer, various processing functions in each of the above devices are realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the present apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

10, 20, 30, 30 ', 40 Voice recognition devices 101A, 101B Input terminals 110A, 110B Audio signal acquisition units 120A, 120B Feature amount extraction unit 130A First recognition processing unit 130B Second recognition processing unit 135A, 135B Provisional recognition processing Unit 140 translation topic model generation unit 150 word occurrence probability calculation unit 160 language model adaptation unit 170A, 170B keyword extraction unit 180 high reliability word extraction unit 190A, 190B reliability estimation unit 195 adaptation processing determination unit 210 dialogue corpus storage unit 220 translation Topic model storage unit 230 Language model storage unit 240 Acoustic model storage unit 250 Adaptive language model storage unit 260 Keyword storage unit

Claims (8)

A speech recognition device for grouping a plurality of speakers in advance and recognizing conversational speech between a first speaker belonging to one group and a second speaker belonging to the other group,
A dialogue corpus storage unit for storing a dialogue corpus that transcribes a conversation voice between a speaker belonging to the one group and a speaker belonging to the other group;
A language model storage unit for storing a language model;
An acoustic model storage unit for storing an acoustic model;
A first recognition processing unit that recognizes the voice of the first speaker and obtains a first speaker recognition result using the acoustic model and the language model;
A translation topic model generating unit for obtaining a translation topic model for converting the utterance content of the first speaker from the dialogue corpus into the utterance content of the second speaker;
Using the first speaker recognition result and the translation topic model, a word occurrence probability calculating unit that calculates a word occurrence probability of the second speaker;
A language model adaptation unit for obtaining an adaptive language model from the language model using the word occurrence probability;
A second recognition processing unit for recognizing the second speaker's voice and obtaining a second speaker recognition result using the acoustic model and the adaptive language model;
A speech recognition apparatus comprising: The speech recognition device according to claim 1,
P (z _i ) is the probability of occurrence of topic z _i , P (w _{a, n} | z _i ) is the probability of occurrence of word w _n under topic z _i in the first speaker, and P (w _{b, n} | z _i ) is the occurrence probability of the word w _n under the topic z _i in the second speaker, and P (z _i | d _a ) is the first speaker recognition result under the d _a The probability of occurrence of topic z _i
The translation topic model includes the occurrence probability P (z _i ), the occurrence probability P (w _{a, n} | z _i ), and the occurrence probability P (w _{b, n} | z _i ),
The word occurrence probability calculating unit calculates the word occurrence probability P (w _{b, n} | d _a ) according to the following equation:

A speech recognition apparatus characterized by that. The speech recognition device according to claim 1 or 2,
A keyword storage unit for storing predetermined keywords;
A keyword extraction unit that receives the first speaker recognition result or the dialogue corpus as input and outputs a keyword string obtained by extracting the keyword included in the input text;
A speech recognition apparatus comprising: The speech recognition device according to claim 3,
The speech recognition apparatus, wherein the keyword is a word selected with a high tf-idf value in the first speaker recognition result. The speech recognition device according to claim 1,
A high-reliability word extraction unit that extracts only words having higher recognition reliability than a predetermined threshold from the first speaker recognition result,
The speech recognition apparatus, wherein the first recognition processing unit outputs the recognition reliability corresponding to the first speaker recognition result. The speech recognition device according to any one of claims 1 to 5,
Using a plurality of voices as input, using the language model and the acoustic model, a provisional recognition processing unit that recognizes each of the plurality of input voices and outputs a provisional recognition result and provisional recognition reliability;
A reliability estimation unit that estimates the document reliability for each input speech using the provisional recognition reliability;
An adaptive processing determination unit that determines the voice of the first speaker and the voice of the second speaker from the input voice using the document reliability;
A speech recognition apparatus comprising: A speech recognition method for grouping a plurality of speakers in advance and recognizing conversation speech between a first speaker belonging to one group and a second speaker belonging to the other group,
A first recognition processing step of recognizing the voice of the first speaker and obtaining a first speaker recognition result using an acoustic model and a language model;
A translation topic for converting the utterance content of the first speaker into the utterance content of the second speaker from a dialogue corpus in which conversation speech between the speaker belonging to the one group and the speaker belonging to the other group is transcribed A translation topic model generation step for obtaining a model;
Using the first speaker recognition result and the translation topic model, a word occurrence probability calculating step for calculating a word occurrence probability of the second speaker;
A language model adaptation step for obtaining an adaptive language model from the language model using the word occurrence probability;
A second recognition processing step of recognizing the voice of the second speaker and obtaining a second speaker recognition result using the acoustic model and the adaptive language model;
A speech recognition method comprising:   A program for causing a computer to function as the speech recognition apparatus according to claim 1.

Images