JP2012063536A - Terminal device, speech recognition method and speech recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a terminal device and a speech recognition processing method capable of realizing speech recognition processing that achieves both large vocabulary and high-accuracy speech recognition processing and customization of a language model or acoustic model for each user. The purpose is to do.
SOLUTION: Voice data input from a microphone 102 is transmitted to a voice recognition server 200 where a communication unit 101 performs voice recognition. The speech recognition server 200 performs speech recognition processing using one acoustic model and one language model, and transmits the result to the terminal device 100 as a speech recognition result. In the terminal device 100, the speech recognition result is received from the speech recognition server 200, and the received speech recognition result is input to the hypothesis re-evaluation unit 106 by the input speech data and the sound provided in the terminal device 100. Evaluate using models and language models.
[Selection] Figure 1

Description

  The present invention relates to a terminal device that processes a speech recognition result, a speech recognition method, and a speech recognition program.

  The voice recognition process may be performed in the client terminal or on the server side. When performed in a client terminal, the terminal is often dedicated to the user. Therefore, as user-specific information in the terminal, user dictionary, user input voice and call voice, acoustic training results, etc. It is easy to personally customize language models and acoustic models. Accordingly, user-adaptive speech recognition is possible, but there are disadvantages such as a limited amount of memory and computing resources, and a restriction on the number of vocabularies and the range of hypothesis search.

  On the other hand, when it is performed on the server side, there are merits that a large amount of memory and computing resources can be used compared to in-terminal speech recognition, and speech recognition with high vocabulary and high accuracy is possible. However, since the server is shared by many users, it learns each user's language model or acoustic model, or loads a customized language model or acoustic model for each user immediately and quickly according to access. That is expensive. Therefore, there is a demerit that it is difficult to customize the language model or the acoustic model for each user in the speech recognition on the server side.

  Therefore, there is a problem of realizing a speech recognition process that combines the advantages of both, a large vocabulary and highly accurate speech recognition process, and customization of each language model or acoustic model for each user. One attempt to solve the problem is shown. In Patent Document 1, the server returns the time information of the word boundary in the recognition result to the terminal, and the terminal refers to the time information and performs re-recognition using a dictionary owned by itself. In particular, we aim to improve recognition accuracy by re-recognizing only words that are determined as unknown words or proper nouns.

JP 2010-85536 A

  However, in such a technique of Patent Document 1, if the word boundary is not correctly recognized by the server side recognition, the terminal side will refer to the incorrect time information, and the correct recognition result even after re-recognition at the terminal May not be obtained. Also, if the word to be recognized on the server side is a word outside the vocabulary, it may not be possible to correctly determine unknown words or proper nouns, and correct recognition results may not be obtained even after re-recognition at the terminal There is.

  Furthermore, the speech recognition server may not be able to recognize speech correctly due to a lack of vocabulary in the language model. In the case of an unknown word, since it is not registered in the language model as a target for normal speech recognition, the problem becomes remarkable.

  Therefore, the present invention has been made in view of the above, and a terminal device capable of realizing a speech recognition process that achieves both a large vocabulary and high-accuracy speech recognition process and customization of a language model or an acoustic model for each user. Another object is to provide a speech recognition processing method and a speech recognition program.

  In order to solve the above-described problem, a terminal device of the present invention includes an input unit that receives input of voice data, a transmission unit that transmits voice data input by the input unit to an external device that performs voice recognition, A receiving means for receiving a speech recognition result recognized using one acoustic model and one language model in the external device from the external device, and a speech recognition result received by the receiving means in the input means A hypothesis re-evaluation unit that evaluates using the input speech data, another acoustic model different from the one acoustic model, and another language model different from the one language model; and the hypothesis re-evaluation unit Presenting means for presenting one speech recognition result according to the evaluation.

  The speech recognition method of the present invention includes an input step for receiving input of speech data, a transmission step for transmitting the speech data input by the input step to an external device that performs speech recognition, A reception step of receiving a speech recognition result recognized using an acoustic model and one language model from the external device, and a speech recognition result received in the reception step, the speech data input in the input step; And a hypothesis re-evaluation step that uses another acoustic model that is different from the one acoustic model and another language model that differs from the one language model, and an evaluation according to the hypothesis re-evaluation step. A presentation step for presenting a speech recognition result.

  The speech recognition program of the present invention includes an input module that receives input of speech data, a transmission module that transmits speech data input by the input module to an external device that performs speech recognition, A voice recognition result recognized using an acoustic model and one language model from the external device; a voice recognition result received by the reception module; a voice data input in the input module; And a hypothesis re-evaluation module that evaluates using another acoustic model different from the one acoustic model and another language model different from the one language model, and an evaluation according to the evaluation by the hypothesis re-evaluation module. A presentation module for presenting a speech recognition result.

  According to these inventions, the input speech data is transmitted to an external device that performs speech recognition, and the speech recognition result recognized by the external device using one acoustic model and one language model is received from the external device. The received speech recognition result is evaluated using the input speech data and another acoustic model different from the one acoustic model or another language model different from the one language model. Therefore, one speech recognition result is presented. As a result, an appropriate speech recognition result can be selected within a range of results recognized by the external device, and a large vocabulary and highly accurate speech recognition process can be performed. Furthermore, the amount of computation of the CPU, the amount of memory used, and the like can be reduced as compared to performing speech recognition processing in a normal terminal.

  The terminal device 100 according to the present invention includes a user dictionary storage unit that stores a user dictionary in which a word determined by a user and pronunciation information indicating pronunciation of the word are associated with each other, and voice recognition received by the reception unit. By searching the user dictionary stored in the user dictionary storage unit for a word that approximates the word of the predetermined condition among the words included in the result, and adding the approximate word to the speech recognition result And a recognition result expansion means for extending the speech recognition result, wherein the hypothesis re-evaluation means evaluates the speech recognition result expanded by the recognition result expansion means.

  According to the present invention, a user dictionary in which a word defined by a user and pronunciation information indicating pronunciation of the word is associated is stored, and a predetermined word among words included in the received speech recognition result is stored. A word that approximates the conditional word is searched from the user dictionary, and the approximated word is added / expanded to the speech recognition result, and the expanded speech recognition result is evaluated. As a result, an accurate speech recognition result can be obtained even for speech data including a user-specific word by adding, expanding, and evaluating a word specific to the user and evaluating it. . In particular, even if the speech data includes unknown words such as uncommon proper nouns, a speech recognition result can be obtained in consideration thereof.

  In the terminal device 100 of the present invention, the other acoustic model and the other language model are determined based on a usage history of the terminal device or a user operation.

  According to the present invention, the other acoustic model and the other language model are determined based on the usage history of the terminal device or the user operation, whereby the user-specific acoustic model and language model of the terminal device are obtained. It can be used, and voice recognition processing can be realized more in line with user identification.

  Further, in the terminal device 100 of the present invention, the hypothesis re-evaluation means calculates a language likelihood of the speech recognition result received by the receiving means using the other language model, and uses the calculated language likelihood. The speech recognition result is evaluated based on the re-evaluation score and the recognition score accompanying the speech recognition result received by the receiving means.

  According to the present invention, the language likelihood of the speech recognition result received by the receiving unit is calculated using another language model, and the re-evaluation score including the calculated language likelihood and the received speech recognition result are calculated. The speech recognition result can be evaluated based on the accompanying recognition score. Therefore, the recognition accuracy can be improved by the language model adapted to the individual.

  Further, in the terminal device 100 of the present invention, the hypothesis re-evaluation means is received by the receiving means when the voice data input by the input means is given using the other acoustic model. The speech likelihood is calculated by calculating the likelihood of the speech recognition result, and the speech recognition is performed based on the re-evaluation score composed of the calculated acoustic likelihood and the recognition score attached to the speech recognition result received by the receiving unit. Evaluate the results.

  According to this invention, the sound is obtained by obtaining the likelihood of the speech recognition result received by the receiving unit when the voice data input by the input unit is given using the other acoustic model. The likelihood is calculated, and the speech recognition result can be evaluated based on the re-evaluation score composed of the calculated acoustic likelihood and the recognition score attached to the received speech recognition result. Therefore, the recognition accuracy can be improved by the acoustic model adapted to the individual.

  According to the present invention, an appropriate speech recognition result can be selected within a range of results recognized by an external device, and a large vocabulary and highly accurate speech recognition process can be performed. Furthermore, the amount of computation of the CPU, the amount of memory used, and the like can be reduced as compared to performing speech recognition processing in a normal terminal.

It is a block diagram which shows the function of the speech recognition system of this embodiment. 2 is a block diagram showing a hardware configuration of a terminal device 100. FIG. It is explanatory drawing which shows the candidate list | wrist of the speech recognition result in a N-best system. It is explanatory drawing which showed the recognition result candidate which is one of the speech recognition results. It is explanatory drawing which shows the candidate list | wrist including the candidate extended in the recognition result expansion part 105, and each score corresponding to it. 3 is a flowchart showing processing of the voice recognition system 10.

  Embodiments of the present invention will be described with reference to the accompanying drawings. Where possible, the same parts are denoted by the same reference numerals, and redundant description is omitted.

  FIG. 1 is a block diagram showing functions of the voice recognition system of the present embodiment. As shown in FIG. 1, the voice recognition system 10 according to the present embodiment includes a terminal device 100 and a voice recognition server 200. The voice input in the terminal device 100 is transmitted to the voice recognition server 200, and voice recognition processing is performed in the voice recognition server 200 which is an external device. Then, the voice recognition result recognized by the voice recognition server 200 is returned to the terminal device 100. The returned speech recognition result includes a plurality of recognition candidates, and the terminal device 100 evaluates which recognition candidate is valid, and uses the recognition candidate having a high evaluation value as the speech recognition result. It is displayed on the display unit 107.

  Hereinafter, functions of the terminal device 100 and the voice recognition server 200 will be described. As shown in FIG. 1, the terminal device 100 includes a communication unit 101 (transmission unit, reception unit), a microphone 102 (input unit), a voice processing unit 103, a user dictionary holding unit 104 (user dictionary storage unit), and a recognition result extension. A unit 105 (recognition result expansion unit), a hypothesis re-evaluation unit 106 (hypothesis re-evaluation unit), and a display unit 107 (presentation unit) are included. The terminal device 100 is a mobile terminal, for example, and is configured by hardware shown in FIG.

  FIG. 2 is a block diagram illustrating a hardware configuration of the terminal device 100. As shown in FIG. 2, the terminal device 100 shown in FIG. 1 physically includes a CPU 11, a RAM 12 and a ROM 13 that are main storage devices, an input device 14 such as a keyboard and a mouse that are input devices, and an output such as a display. The computer 15 includes a device 15, a communication module 16 that is a data transmission / reception device such as a network card, an auxiliary storage device 17 such as a hard disk, and the like. Each function described in FIG. 1 has an input device 14, an output device 15, and a communication module 16 under the control of the CPU 11 by reading predetermined computer software on hardware such as the CPU 11 and the RAM 12 shown in FIG. 2. This is realized by reading and writing data in the RAM 12 and the auxiliary storage device 17. Hereinafter, each functional block will be described based on the functional blocks shown in FIG.

  The microphone 102 is a part that receives a voice uttered by the user of the terminal device 100.

  The sound processing unit 103 is a part that processes the sound received by the microphone 102, and is a part that converts the sound into, for example, digital data.

  The communication unit 101 is a part that transmits the voice data processed by the voice processing unit 103 to the voice recognition server 200 that is an external device, and receives a voice recognition result from the voice recognition server 200. When the communication unit 101 receives the speech recognition result from the speech recognition server 200, the communication unit 101 outputs the speech recognition result to the recognition result extension unit 105. This speech recognition result includes a plurality of recognition candidates.

Specifically, the communication unit 101 receives the voice recognition result shown in FIG. FIG. 3A is an explanatory diagram illustrating a candidate list including a recognizer of a speech recognition result in the N-best method. In FIG. 3A, it is assumed that the following utterance is made as a pronunciation.
(Speech contents) Today, when Yokoshima was met, if a candidate list is represented by the N-best method, a plurality of recognition candidates are output as a speech recognition result as shown in FIG. In the N-Best method, a recognition score is calculated for each sentence obtained by combining candidates obtained by recognizing words that are classified into phrases by speech.

  In addition, the candidate list is not limited to the N-best method, and a confusion network (CN) method may be used as shown in FIG. In the CN method, management is performed so as to indicate word connections classified for each phrase, and a recognition score is calculated for each connection.

  In any of the methods, a plurality of recognition candidates can be obtained by combining words segmented for each phrase.

  The user dictionary holding unit 104 is a part that stores a user dictionary set by the user of the terminal device 100. This user dictionary is dictionary data set by the user for speech recognition, or an extended dictionary for phonetic kanji conversion, phonebook data, etc., and pronunciation (phonetic) and kanji (or katakana) , Hiragana may be associated with the dictionary data.

  The recognition result extension unit 105 is a part that performs processing for extending the voice recognition result received from the voice recognition server 200 in the communication unit 101. Specifically, the recognition result expansion unit 105 is a pronunciation word close to a word that satisfies a predetermined condition (for example, a word with a low recognition score (below a predetermined value)) among words included in each of a plurality of recognition candidates. The speech recognition result can be expanded by searching the user dictionary in the user dictionary holding unit 104 and adding the searched word. For the evaluation of the closeness of pronunciation, the edit distance of the character string is used, but other methods may be used. In addition, here, processing is performed to expand a word with a low recognition score as a predetermined condition, but in addition, it is determined that the word has a predetermined part of speech or a proper noun, and the word is expanded. Also good.

  FIG. 4 shows a specific example of recognition candidate expansion in the speech recognition result. FIG. 4A is an explanatory diagram showing recognition candidates that are one of speech recognition results. In FIG. 4A, for example, it is assumed that the recognition score of “Yokohama / Yokohama” is low (below a predetermined value). Therefore, the recognition result expansion unit 105 searches the user dictionary in the user dictionary holding unit 104 for a word that is pronounced like “Yokohama”.

  As shown in FIG. 4B, “Yokoshima / Yokoshima” is searched here, and “Yokoshima / Yokoshima” is added as another candidate for “Yokohama / Yokohama”.

  It seems reasonable that the recognition score of the expanded recognition candidate is the same as that of the original recognition candidate, but a penalty may be given to the score of the original candidate based on the closeness of pronunciation. . For example, the recognition score of “I met Yokohama today” is 0.32. And if you give a penalty of 0.05 for each editing distance of pronunciation, the editing distance of “Yokohama” and “Yokoshima” is 1, so the recognition score of the expanded candidate “I met Yokoshima today” is 0.27.

Returning to FIG. 1, the description will be continued. The hypothesis re-evaluation unit 106 is a part that evaluates a plurality of recognition candidates including the recognition candidates expanded by the recognition result expansion unit 105, and at least one of a language model and an acoustic model provided in the terminal device 100 is used. Use and evaluate. Specifically, the hypothesis reevaluation unit 106 uses a language model to obtain a language likelihood based on the word N-gram probability for each of a plurality of recognition candidates of the speech recognition result. Further, the hypothesis re-evaluation unit 106 uses the acoustic model to obtain the acoustic likelihood for each of a plurality of speech recognition result candidates. That is, each of a plurality of speech recognition result candidates is treated as a recognition hypothesis, and the phoneme of the recognition hypothesis that is the speech recognition result when the speech data processed by the speech processing unit 103 is given using an acoustic model. The acoustic likelihood can be obtained by obtaining the likelihood of the column. The acoustic likelihood calculation method is widely known as described in the following reference. For example, in the case of an acoustic model based on a hidden Markov model typically used at present, the phoneme HMM included in the acoustic model is used. (Hidden Markov Model) is concatenated to generate a sentence HMM corresponding to the phoneme sequence constituting the recognition hypothesis, and the output probability on the sentence HMM of the feature amount sequence obtained by analyzing the speech data is DP matching, etc. Find it in The hypothesis re-evaluation unit 106 calculates a re-evaluation score for evaluation by obtaining at least one of language likelihood and acoustic likelihood for each recognition candidate.
[References] ITText Speech Recognition System Edited by Information Processing Society of Japan, Kiyohiro Shikano, Katsunobu Ito, Tatsuya Kawahara, Kazuya Takeda, Mikio Yamamoto, 2001/05 published

  When the hypothesis re-evaluation unit 106 calculates a re-evaluation score for each candidate, the hypothesis re-evaluation unit 106 calculates a final score based on the recognition score obtained from the speech recognition server 200 and the re-evaluation score. For example, the hypothesis re-evaluation unit 106 may calculate the sum of the recognition score and the re-evaluation score, thereby making it the final score, weighting either or both of the recognition score and the re-evaluation score, The final score may be obtained by calculating the sum. Further, only the re-evaluation score may be used as the final score. Furthermore, the reevaluation score may be obtained using only one of the acoustic model and the language model.

  FIG. 5 is an explanatory diagram showing a candidate list including candidates expanded in the recognition result expansion unit 105 and scores corresponding thereto.

  For example, the candidate “I went to Yokohama today” is associated with recognition score: 0.5, re-evaluation score: 0.3, and final score: 0.8. These numerical values are temporarily stored in the hypothesis re-evaluation unit 106. In FIG. 5, “I met Yokoshima today”, but the final score is 0.87, which is the highest value among the recognition candidates, and this is output as a recognition result.

  The display unit 107 displays the recognition result evaluated by the hypothesis re-evaluation unit 106.

  Next, the voice recognition server 200 will be described. As shown in FIG. 1, the voice recognition server 200 includes a communication unit 201 and a voice recognition unit 202. The communication unit 201 is a part that receives voice data transmitted from the terminal device 100 and returns a voice recognition result recognized by a voice recognition unit 202 described later.

  The voice recognition unit 202 is a part that performs voice recognition processing on the voice data received by the communication unit 201. The speech recognition unit 202 performs speech recognition using an acoustic model and a language model, and the acoustic model and the language model are models that can be applied to everyone.

  In this way, the voice data received in the terminal device 100 is subjected to voice recognition processing in the voice recognition server 200 and returned to the terminal device 100. The terminal device 100 is uniquely configured based on the returned voice recognition result. By further utilizing the dictionary data, it is possible to obtain a more accurate speech recognition result.

  Next, processing of the speech recognition system 10 configured as described above will be described. FIG. 6 is a flowchart showing a speech recognition method processed in the speech recognition system 10. First, in the terminal device 100, voice is received through the microphone 102 (S101). Then, the voice data is processed by the voice processing unit 103 and transmitted to the voice recognition server 200 by the communication unit 101 (S102).

  In the voice recognition server 200, voice data is received by the communication unit 201 (S201), voice recognition processing is performed by the voice recognition unit 202, and a voice recognition result is obtained (S202). Then, the voice recognition result is transmitted to the terminal device 100 by the communication unit 201 (S203).

  In the terminal device 100, the speech recognition result is received by the communication unit 101 (S 103), and the recognition result expansion unit 105 searches the user dictionary in the user dictionary holding unit 104 for words whose pronunciation is similar based on the speech recognition result. (S104). Then, the retrieved word is added to the speech recognition result by the recognition result expansion unit 105 (S105).

  The hypothesis reevaluation unit 106 evaluates each candidate speech recognition result expanded by the recognition result expansion unit 105, and calculates a reevaluation score (S106). Then, the final score is calculated from the recognition score obtained from the speech recognition server 200 and the re-evaluation score by the hypothesis re-evaluation unit 106 (S107), and based on the final score, a plurality of candidates as speech recognition results are calculated. A candidate is selected and displayed on the display unit 107 (S107).

  In this way, within the range recognized by the speech recognition server 200, the terminal device 100 can evaluate the speech recognition result, and can perform speech recognition processing with a large vocabulary and high accuracy.

  In the present embodiment, the terminal device 100 is exemplified as the device. However, the present invention is not limited to this, and the device may be configured as a program module for executing each function in the terminal device 100. That is, an input process from the microphone 102, an input module corresponding to the voice processing unit 103, a transmission module and reception module corresponding to the communication unit 101, a hypothesis re-evaluation module corresponding to the hypothesis re-evaluation unit 106, and a display unit 107 such as a display A program having a presentation module for display on a computer, and by causing the computer system such as a portable terminal to read the program, a function equivalent to that of the terminal device 100 can be realized. Note that this program may be stored in a recording medium such as a CDROM, a DVD, or the like so as to be readable by a computer system.

  Below, the effect of the terminal device 100 of this embodiment is demonstrated. The microphone 102 provided in the terminal device 100 receives the voice uttered by the user, and the voice processing unit 103 converts it into voice data. Then, the voice data input in this way is transmitted to the voice recognition server 200 in which the communication unit 101 performs voice recognition.

  The speech recognition server 200 performs speech recognition processing using one acoustic model and one language model, and transmits the result to the terminal device 100 as a speech recognition result. In the terminal device 100, the speech recognition result is received from the speech recognition server 200, and the received speech recognition result is input to the hypothesis re-evaluation unit 106 by the input speech data and the sound provided in the terminal device 100. Evaluate using models and language models. The acoustic model and the language model are different from the acoustic model and the language model in the speech recognition server 200. The user of the terminal device 100 performs in-terminal information, user dictionary, past input speech, call speech, acoustic training. It is customized by mode.

  Then, one speech recognition result is determined according to the evaluation by the hypothesis re-evaluation unit 106, and the display unit 107 displays the one speech recognition result.

  Thus, an appropriate speech recognition result can be evaluated and selected by the terminal device 100 within the range of the result recognized by the speech recognition server 200, and a large vocabulary and highly accurate speech recognition process can be performed. Furthermore, the amount of computation of the CPU, the amount of memory used, and the like can be reduced as compared to performing speech recognition processing in a normal terminal.

  In addition, the terminal device 100 according to the present embodiment includes a user dictionary holding unit 104 that stores a user dictionary in which a word determined by a user and pronunciation information indicating pronunciation of the word are associated with each other, and is included in the received speech recognition result. Among words that are approximated to a word of a predetermined condition from the user dictionary stored in the user dictionary holding unit, and adding the approximated word to the speech recognition result, the speech recognition result Is further included. Then, the hypothesis reevaluation unit 105 performs evaluation on the speech recognition result expanded by the recognition result expansion unit 105.

  As a result, an accurate speech recognition result can be obtained even for speech data including a user-specific word by adding, expanding, and evaluating a word specific to the user and evaluating it. . In particular, even if the speech data includes unknown words such as uncommon proper nouns, a speech recognition result can be obtained in consideration thereof.

  In addition, the acoustic model and language model provided in the terminal device 100 of the present embodiment are determined based on the usage history of the terminal device 100 or user operations. As described above, these acoustic models and language models are customized with in-terminal information, user dictionary, past input voice, call voice, acoustic training mode, and the like.

  As a result, the acoustic model and language model specific to the user of the terminal device can be used, and voice recognition processing can be realized more in line with the user's specification.

  Further, in the terminal device 100 according to the present embodiment, the hypothesis re-evaluation unit 106 calculates a language likelihood using a language model provided in the terminal device 100, and a re-evaluation score including the calculated language likelihood. The voice recognition result is evaluated based on the recognition score attached to the voice recognition result received by the communication unit 101. Therefore, the recognition accuracy can be improved by the language model adapted to the individual.

  Further, in the terminal device 100, the HMM is applied as another acoustic model, and the hypothesis re-evaluation unit 106 generates a sentence HMM corresponding to the phoneme sequence constituting the speech recognition result received by the communication unit 101, and the speech The acoustic likelihood is calculated by obtaining the output probability on the generated sentence HMM of the feature amount series obtained based on the speech data processed in the processing unit 103, and from the calculated acoustic likelihood. Re-evaluate. Therefore, the recognition accuracy can be improved by the acoustic model adapted to the individual.

DESCRIPTION OF SYMBOLS 100 ... Terminal device 101 ... Communication part 102 ... Microphone 103 ... Speech processing part 104 ... User dictionary holding part 105 ... Recognition result expansion part 106 ... Hypothesis re-evaluation part 107 ... Display part 200 ... Speech recognition Server 201, communication unit 202, voice recognition unit.

Claims (7)

Input means for receiving input of voice data;
Transmitting means for transmitting voice data input by the input means to an external device that performs voice recognition;
Receiving means for receiving, from the external device, a speech recognition result recognized using one acoustic model and one language model in the external device;
The speech recognition result received by the receiving means is converted into at least one of the speech data input in the input means and another acoustic model different from the one acoustic model or another language model different from the one language model. Hypothesis re-evaluation means to evaluate using one,
Presenting means for presenting one speech recognition result according to the evaluation by the hypothesis re-evaluating means;
A terminal device comprising: User dictionary storage means for storing a user dictionary in which a word defined by a user and pronunciation information indicating pronunciation of the word are associated;
Among words included in the speech recognition result received by the receiving unit, a word that approximates a word of a predetermined condition is searched from the user dictionary stored in the user dictionary storage unit, and the approximate word A recognition result expansion means for extending the speech recognition result by adding to the speech recognition result;
Further comprising
The terminal apparatus according to claim 1, wherein the hypothesis re-evaluation unit performs evaluation on the voice recognition result expanded by the recognition result expansion unit.   The terminal device according to claim 1, wherein the other acoustic model and the other language model are determined based on a usage history of the terminal device or a user operation. The hypothesis reevaluation means includes:
The language likelihood of the speech recognition result received by the receiving unit is calculated using the other language model, and the re-evaluation score composed of the calculated language likelihood and the speech recognition result received by the receiving unit are calculated. The terminal device according to any one of claims 1 to 3, wherein a voice recognition result is evaluated based on an accompanying recognition score. The hypothesis reevaluation means includes:
Applying HMM ((Hidden Markov Model)) as the other acoustic model, generating a sentence HMM corresponding to a phoneme sequence that constitutes a speech recognition result received by the receiving means,
By determining the output probability on the sentence HMM of the feature amount series obtained based on the speech data input by the input means,
An acoustic likelihood is calculated, and the speech recognition result is evaluated based on a re-evaluation score composed of the calculated acoustic likelihood and a recognition score attached to the speech recognition result received by the receiving unit. Item 5. The terminal device according to any one of Items 1 to 4. An input step for receiving input of voice data;
A transmission step of transmitting the voice data input in the input step to an external device that performs voice recognition;
Receiving a speech recognition result recognized using one acoustic model and one language model in the external device from the external device;
The speech recognition result received in the receiving step is converted into at least one of the speech data input in the input step and another acoustic model different from the one acoustic model or another language model different from the one language model. Hypothesis re-evaluation step to evaluate using one,
A presenting step of presenting one speech recognition result according to the evaluation by the hypothesis re-evaluation step;
A speech recognition method comprising: An input module that accepts voice data input;
A transmission module for transmitting audio data input by the input module to an external device for performing speech recognition;
A receiving module for receiving a speech recognition result recognized by the external device using one acoustic model and one language model from the external device;
The speech recognition result received by the receiving module is converted into at least one of speech data input in the input module and another acoustic model different from the one acoustic model or another language model different from the one language model. Hypothesis re-evaluation module to evaluate using one,
A presentation module for presenting one speech recognition result according to the evaluation by the hypothesis re-evaluation module;
A speech recognition program comprising:

Images