JP2010231149A - Terminal using kana-kanji conversion system for voice recognition, method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a terminal, a method and a program, capable of integrating a resource of a language model of a whole system by using a kana-kanji (Japanese-Chinese character) conversion system for voice recognition. <P>SOLUTION: The terminal includes: an acoustic score calculating means for calculating an acoustic score which is given by an acoustic model for each Japanese syllabary of a plurality of candidates for a phoneme frame in order to calculate a score; an N-gram score calculating means for calculating an N-gram score according to a connection strength between words for Japanese syllabary series which is a candidate; a case grammar score calculating means for calculating a case grammar score according to the connection strength between syllables; and an integrated score calculating means for calculating an integrated score based on the plurality of scores. In order to search kana-kanji in which the integrated score is the highest based on morphological analysis, a search control means for repeating score calculation, and a case frame processing means for selecting a conversion candidate by a case structure of a sentence, are included. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a terminal, a method, and a program using a kana-kanji conversion system for speech recognition.

  2. Description of the Related Art Conventionally, there is a kana-kanji conversion system that converts a hiragana character input to a user by a key operation into a kana-kanji character string in order to input characters to an application. Usually, it operates as a Japanese input function built into word processor software or text editor software.

  On the other hand, there is a voice recognition system that acquires a voice signal uttered by a user with a microphone and converts the voice signal into a kana-kanji character string. A user can input a character string without operating a key. However, the number of vocabulary words spoken by people is enormous. For this reason, it is difficult to improve the accuracy of voice recognition.

  In order to improve the accuracy of a character string predicted by speech recognition, there is a technique for creating topic prediction data from sentences (see, for example, Patent Document 1). A character string of kana / kanji is predicted based on the topic prediction data and hiragana.

  There is also a technique for improving recognition accuracy by narrowing down a plurality of candidate character strings to be predicted to specific recognition target words (for example, see Patent Document 2). However, since the recognition target words are narrowed down, the use of the speech recognition system is also limited.

  FIG. 1 shows a system having a kana-kanji conversion function and a speech recognition function in the prior art.

  According to FIG. 1, the terminal 1 includes a display 100, a key operation unit 111, and a microphone 121 as hardware. The display 100 displays an input interface of an application such as a text editor, and also displays hiragana and kanji kanji of prediction candidates input by the user. The key operation unit 111 causes the user to input hiragana using a key. The microphone 121 acquires an audio signal uttered by the user.

  The terminal 1 includes a kana / kanji conversion engine 112, a language model storage unit 113, a history storage unit 114, and a kana / kanji selection screen control unit 115 as a kana / kanji conversion function. The kana-kanji conversion engine 112 uses the language model storage unit 113 and the history storage unit 114 to convert the hiragana input by the key operation unit 111 into a plurality of candidate kana-kanji characters. The kana / kanji selection screen control unit 115 causes the user to select one kana / kanji character from a plurality of candidate kana / kanji characters.

  The language model storage unit 113 stores a word dictionary and a language model. The language model defines the probability of the connection relationship between words. Words that are not registered in the word dictionary cannot be converted. Therefore, the language model storage unit 113 needs to store a word dictionary including all words input by key operations and their language models. The language model storage unit 113 corresponding to the kana-kanji conversion system stores the dependency grammar probability and the case grammar probability.

  The history accumulating unit 114 accumulates a learning history of kana / kanji characters (words, phrases, etc.) selected by the user in the past.

  The terminal 1 is connected to the voice recognition server 2 via a communication interface. And the terminal 1 transmits the audio | voice signal acquired by the microphone 121 to the audio | voice recognition server 2 via a network. In speech recognition, the number of vocabulary words spoken by a person is enormous, requiring a large-capacity language model and a high processing load. Therefore, it is preferable to separately provide a speech recognition engine as a speech recognition server.

  The speech recognition server includes a speech recognition engine 122, an acoustic model storage unit 123, and a language model storage unit 124. The speech recognition engine 122 uses the acoustic model storage unit 123 and the language model storage unit 124 to convert the received speech signal into a plurality of candidate kana-kanji characters. A plurality of candidate kana-kanji characters are transmitted to the terminal 1 via the network. The voice recognition selection screen control unit 125 of the terminal 1 causes the user to select one kana / kanji character from a plurality of candidate kana / kanji characters.

  The acoustic model storage unit 123 stores an acoustic model. The acoustic model stores a correspondence relationship between acoustic feature parameters and phonemes. A phoneme is roughly equivalent to an alphabetic unit when a word is written in Roman letters. The acoustic model storage unit 123 stores the probability of the phoneme with respect to the acoustic feature parameter.

  The language model accumulation unit 124 defines the probability of the connection relationship between words. The language model storage unit 124 of the speech recognition system employs an N-gram that uses the probability of the connection relationship between adjacent N words. On the other hand, the language model storage unit 113 of the kana-kanji conversion system employs a dependency grammar that uses only the probability of the connection relationship between two adjacent words. For this reason, the language model storage unit 124 of the speech recognition system has more resources than the language model storage unit 113 of the kana / kanji system. Of course, the terminal 1 may further include a speech recognition engine 122, an acoustic model storage unit 123, and a language model storage unit 124.

  There is also a technique for converting a voice signal uttered by a person into a hiragana by a voice recognition system, and then converting the hiragana into a kana / kanji by a kana-kanji conversion system (see, for example, Non-Patent Documents 1 and 2).

JP 2000-285112 A Japanese Patent No. 40112143

Voice recognition engine "VORERO (VOice REcognition RObustR)", Asahi Kasei, [online], [March 19, 2009 search], Internet <URL: http://www.asahi-kasei.co.jp/vorero/jp /vorero/index.html> Japanese Kana-Kanji conversion system "Mobile Wnn", OMRON, [online], [March 19, 2009 search], Internet <URL: http://www.omronsoft.co.jp/SP/mobile/>

  According to the above-described prior art, the speech recognition system and the kana-kanji conversion system are provided separately and use resources independent of each other. The acoustic model and language model in the speech recognition system need to correspond to all spoken words such as buzzwords, and the resources are enormous. On the other hand, since the speech recognition system does not execute processing such as case grammar, morphological analysis, and case frame processing incorporated in the kana-kanji conversion system, erroneous conversion in kana-kanji conversion is likely to occur.

  Therefore, the present invention provides a terminal, a method, and a program capable of integrating the language model resources of the entire system and reducing erroneous conversion of speech recognition by using a kana-kanji conversion system for speech recognition. With the goal.

According to the present invention,
Acoustic model storage means for storing an acoustic model used in speech recognition;
Language model storage means for storing a word dictionary and a language model corresponding to N-gram and case grammar;
Microphone means for acquiring a voice signal uttered by a user;
Acoustic analysis means for extracting phoneme frames from the speech signal;
Acoustic score calculation means for calculating an acoustic score given by the acoustic model for each of a plurality of candidate reading pseudonyms for the phoneme frame;
N-gram score calculating means for calculating an N-gram score according to the connection strength between words given by the language model for candidate reading kana series;
Case grammar score calculating means for calculating a case grammar score according to the connection strength between the clauses for the candidate kana series,
An integrated score calculation means for calculating an integrated score based on the acoustic score, the N-gram score, and the case grammar score for the candidate kana series,
In order to select a plurality of candidate kana-kanji characters based on morphological analysis and to search for kana-kanji characters having the highest integrated score, acoustic score calculation means, N-gram score calculation means, case grammar score calculation means, and integrated score calculation means Search control means for repeating
And a case frame processing means for selecting a conversion candidate according to a case structure of a sentence including the searched kana / kanji, and realizing a speech recognition function.

According to another embodiment of the terminal of the present invention,
History accumulating means for storing kana-kanji characters converted in the past;
A history score calculating means for calculating a history score according to a past learning history with respect to candidate reading kana series using the history storage means,
The integrated score calculation means preferably further includes the history score in the integrated score.

According to another embodiment of the terminal of the present invention,
It further has a key operation means for acquiring a hiragana sequence input by a user using a key,
It is also preferable to realize the kana-kanji conversion function by the N-gram score calculation means, the case grammar score calculation means, the integrated score calculation means, the search control means, and the case frame processing means.

According to another embodiment of the terminal of the present invention,
A candidate selection screen control means for allowing the user to select any one kana / kanji character from a plurality of conversion candidates;
It is also preferable that the candidate selection screen control means allows the user to visually recognize the same candidate selection screen for the voice recognition function and the kana-kanji conversion function.

According to the present invention, there is provided a speech recognition method in a terminal,
The terminal
An acoustic model storage unit that stores an acoustic model used in speech recognition;
A language model storage unit that stores a word dictionary and a language model corresponding to N-gram and case grammar;
A first step of acquiring an audio signal uttered by a user with a microphone unit;
A second step of extracting phoneme frames from the speech signal;
A third step of calculating an acoustic score given by the acoustic model for each of a plurality of candidate reading pseudonyms for the phoneme frame;
A fourth step of calculating an N-gram score according to the connection strength between words given by the language model for the candidate kana sequence;
A fifth step of calculating a case grammar score according to the connection strength between clauses for the candidate kana series;
A sixth step of calculating an integrated score based on the acoustic score, the N-gram score, and the case grammar score for the candidate kana sequence;
A seventh step of repeating the third to sixth steps to select a plurality of candidate kana-kanji characters based on morphological analysis and to search for a kana-kanji character having the highest integrated score;
And an eighth step of selecting a conversion candidate according to the case structure of the sentence including the searched kana / kanji, and realizing a speech recognition function.

According to another embodiment of the speech recognition method of the present invention,
The terminal further includes a history storage unit that stores kana-kanji characters converted in the past,
Using the history accumulating unit, the method further includes a step of calculating a history score according to the past learning history for candidate kana series,
In the sixth step, it is also preferable to add the history score to the integrated score.

According to another embodiment of the speech recognition method of the present invention,
The terminal further includes a key operation unit that acquires a hiragana sequence input by a user using a key,
It is also preferable to realize the kana-kanji conversion function from the fourth step to the eighth step.

According to another embodiment of the speech recognition method of the present invention,
By further comprising the step of allowing the user to select any one kana / kanji character from among a plurality of conversion candidates, the voice recognition function and kana / kanji conversion function can be displayed on the same candidate selection screen for the user. Visual recognition is also preferable.

According to the present invention, a speech recognition program for causing a computer mounted on a terminal to function,
Acoustic model storage means for storing an acoustic model used in speech recognition;
Language model storage means for storing a word dictionary and a language model corresponding to N-gram and case grammar;
Microphone means for acquiring a voice signal uttered by a user;
Acoustic analysis means for extracting phoneme frames from the speech signal;
Acoustic score calculation means for calculating an acoustic score given by the acoustic model for each of a plurality of candidate reading pseudonyms for the phoneme frame;
N-gram score calculating means for calculating an N-gram score according to the connection strength between words given by the language model for candidate reading kana series;
Case grammar score calculating means for calculating a case grammar score according to the connection strength between the clauses for the candidate kana series,
An integrated score calculation means for calculating an integrated score based on the acoustic score, the N-gram score, and the case grammar score for the candidate kana series,
In order to select a plurality of candidate kana-kanji characters based on morphological analysis and to search for kana-kanji characters having the highest integrated score, acoustic score calculation means, N-gram score calculation means, case grammar score calculation means, and integrated score calculation means Search control means for repeating
The computer is caused to function as a case frame processing means for selecting a conversion candidate according to the case structure of the sentence including the searched kana / kanji.

  According to the terminal, method, and program of the present invention, by using the kana-kanji conversion system for speech recognition, it is possible to integrate the language model resources of the entire system and reduce erroneous conversion of speech recognition.

This is a system having a kana-kanji conversion function and a speech recognition function in the prior art. It is a functional block diagram of a kana / kanji conversion system and a speech recognition system. It is a function block diagram of the terminal in this invention. It is a flowchart in this invention. It is a predictive conversion sequence of Kana-Kanji comparing the prior art and the present invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 2 is a functional configuration diagram of the kana-kanji conversion system and the speech recognition system.

  FIG. 2A shows the functional configuration of only the kana-kanji conversion system. The kana-kanji conversion engine 112 calculates the probability of kana-kanji candidate by score (scoring) for each word concatenation. The kana-kanji conversion engine 112 includes a dependency grammar score calculating unit 1121, a case grammar score calculating unit 1122, a history score calculating unit 1123, an integrated score calculating unit 1124, a morpheme analyzing unit 1125, and a case frame processing unit 1126. including. The dependency grammar score calculation unit 1121 represents the connection strength between words as a co-occurrence probability between words. Incidentally, according to the kana-kanji conversion system, the character's reading kana is fixed because it is inputted by the key. Therefore, it is only necessary to detect a plurality of synonymous meaning words for the reading kana.

  FIG. 2B shows the functional configuration of only the voice recognition system. The speech recognition engine 122 includes an acoustic analysis unit 1221, an acoustic score calculation unit 1222, an N-gram score calculation unit 1223, an integrated score calculation unit 1224, and a candidate range control unit 1225. In contrast to the dependency grammar score calculation unit 1121 of the kana-kanji conversion system, the speech recognition system uses an N-gram score calculation unit 1223 that calculates the probabilities of the concatenation relationship of N adjacent words. N-gram is an extension of “dependency grammar” that represents the probability of the connection between two adjacent words. In the case of a voice recognition system, the reading pseudonym itself is indeterminate. Therefore, the candidate range control unit 1225 needs to search for kana kanji corresponding to each different reading kana.

  According to the speech recognition system of the present invention, the integrated score calculation further includes the case grammar score calculation unit 1122 and the history score calculation unit 1123 of the kana-kanji conversion system. Further, the morphological analysis unit 1125 and the case frame processing unit 1126 in the kana-kanji conversion system and the candidate range control unit 1225 in the speech recognition system are integrated. Furthermore, according to the kana-kanji conversion system of the present invention, the N-gram score is used instead of the dependency grammar score.

  FIG. 3 is a functional configuration diagram of the terminal in the present invention.

  According to FIG. 3, the terminal 1 includes an acoustic model storage unit 123, a language model storage unit 133, a history storage unit 114, an acoustic analysis unit 1221, an acoustic score calculation unit 1222, and an N-gram score calculation unit 1223. And a case grammar score calculation unit 1122, a history score calculation unit 1123, an integrated score calculation unit 130, a search control unit 131, a case frame processing unit 1126, and a candidate selection screen control unit 132. These functional components are realized by executing a program that causes a computer installed in the terminal to function.

  The N-gram score calculation unit 1223 calculates an N-gram score according to the connection strength between words given by the language model with respect to the candidate kana sequence. Here, according to the voice recognition, since the reading kana is indeterminate, it is necessary to calculate scores for a plurality of different reading kana. For example, when "Ki ..." is connected after "Koutsu", "Regulation" is the best word conversion candidate after "Transport" among "Regulation", "Pretty", "Season", etc. High score.

The case grammar score calculation unit 1122 calculates a case grammar score according to the joint strength between clauses given by the language model for the candidate kana sequence. Case grammar analysis focuses on particles and analyzes the structure of sentences (single sentence, compound sentence, compound sentence, subject, object, modifier, predicate, complement, etc.). For example, in “I will stop you because I ’m included in the crab”, the sentence structure is judged as follows.
“I ’ll stop you because I ’m the one who ’s gone.”
(Subject) (Predicate)
Subject object modifier predicate
The phrase “Zeikinga” – the phrase “Puppeted” has high connection strength, and the importance of sentence composition is low. Judgment is “because Zeikin has been subjugated” as a modifier, and “I”, “Kunii” and “Osamu” are judged to have a high degree of importance. As described above, according to the conventional speech recognition system, only N-gram is considered as a language model, but it is also possible to deal with grammatical utterance with a high degree of freedom.

  The language model storage unit 133 stores a word dictionary and a language model. In particular, the connection relations between words in the word dictionary are scored and stored. Here, the language model storage unit 133 according to the present invention has an N-gram probability and a case grammar probability, and corresponds to the N-gram score calculation unit 1223 and the case grammar score calculation unit 1122. In that respect, it differs from the existing language model storage unit in the kana-kanji conversion system and the speech recognition system.

  The history score calculation unit 1123 sets the kana kanji (word, phrase, etc.) selected by the user in the past to a high score for the candidate reading kana series using the history storage unit 114. That is, the past learning result is reflected in the language score. Basically, the kana-kanji selected immediately before that reading kana is used as a prediction candidate. In addition, for the reading kana, kana kanji is used as a prediction candidate in accordance with the contents of sentences used in the past.

  According to the speech recognition system in the prior art, the candidate to be presented cannot be changed according to the user. On the other hand, according to the present invention, it is possible to improve the predictive conversion accuracy by specifying words that are frequently used according to the user. Also, history information is shared between the kana-kanji conversion function and the voice recognition function.

  The acoustic analysis unit 1221 converts the audio signal input from the microphone 121 into an acoustic feature parameter. The acoustic feature parameter is output to the acoustic score calculation unit 1222. The acoustic feature parameter is a parameter series such as an LPC cepstrum or MFCC obtained by analyzing an input speech signal in units of phoneme frames (phoneme frames) of several tens of msec.

  The acoustic score calculation unit 1222 uses the acoustic model storage unit 123 to calculate an acoustic score given by the acoustic model for each of a plurality of candidate reading pseudonyms for the phoneme frame.

  The integrated score calculation unit 130 integrates the language score (N-gram score, case grammar score, and history score) and the acoustic score with respect to the candidate kana series, and recognizes the word string having the highest score. Output as a result. It is possible to score the acoustic similarity and linguistic validity of the input speech signal.

  The search control unit 131 selects a search candidate based on morphological analysis, and in order to search for a kana / kanji string having the highest integrated score, an acoustic score calculation unit 1222, an N-gram score calculation unit 1223, and a case grammar score calculation The unit 1122, the history score calculation unit 1123, and the integrated score calculation unit 130 are repeated.

Morphological analysis uses a grammar and word dictionary as an information source to divide a sentence written in natural language into a sequence of morphemes (Morpheme, the smallest unit that has meaning in the language). For example, the sentence “I was doing yesterday” can be divided into many words. Depending on how it is separated, the word does not make sense.
・ Yes / No / Wa / Toshi / Ha / Sh / te / Yes ・ Yes / I / Ha / Sh / te / Yes ・ Yes / I / Hash / Te / was

Next, the search control unit 131 selects N candidates with high scores in order to divide them into optimal phrases. By using the score, N candidates are selected not only by the appearance frequency of kana / kanji characters but also by the connection strength of words or phrases.
Σ ^M _{m = 1} α _m Σ ^Nm _{n = 1} P _m (x _mn )
M: number of words w _{m: m-th} word α _m: language of the word w _m score N _m: number of words phoneme frame included in the w _m x _mn: n-th phoneme frame that is included in the word w _m P _m (x _mn ): acoustic score of the acoustic model of the word w _{m in} the phoneme frame

  In the speech recognition system according to the prior art, the kana characters with the highest acoustic score and N-gram score are narrowed down from the candidate kana sequences. On the other hand, according to the present invention, the correct sentence structure candidates are selected by morphological analysis, and then narrowed down to reading kana with higher scores. This also improves the prediction accuracy.

  The case frame processing unit 1126 selects a conversion candidate according to the case structure of the sentence. The case frame means that the verb has information such as “with which particle (in order to be used together)”. A word having a case frame that matches the input reading is set as a first candidate for conversion. For example, “I wear a dress” and “I cut a dress with a knife” can be used to distinguish the homonyms “wear” and “cut”.

  In the conventional speech recognition system, it is difficult to convert homonyms of non-adjacent words into correct candidates due to the nature of N-gram. On the other hand, according to the present invention, the case structure of a sentence is conceptually recognized by case frame processing, and “which word is involved” is analyzed based on the particle after the noun. Thereby, even if the clause (case structure) is not adjacent (particularly apart), an appropriate candidate can be selected from the homonyms.

  The candidate selection screen control unit 132 causes the user to select any one kana / kanji character from a plurality of conversion candidates. Thereby, the user can visually recognize the voice recognition function and the kana-kanji conversion function as the same candidate selection screen.

  FIG. 4 is a flowchart in the present invention.

(S401) An audio signal uttered by a user is acquired using a microphone.
(S402) A phoneme frame is extracted from an audio signal by acoustic analysis.
(S403) Thereafter, steps up to S408 are repeated.
(S404) An acoustic score given by the acoustic model is calculated for each of a plurality of candidate reading names for the phoneme frame.
(S405) An N-gram score is calculated according to the connection strength between words using a language model for the candidate kana sequence.
(S406) A case grammar score is calculated with respect to the candidate kana sequence using the language model according to the connection strength between the clauses.
(S407) A history score is calculated according to the past learning history using the history information for the candidate reading kana series.
(S408) An integrated score based on the acoustic score and the language score (N-gram score, case grammar score, and history score) is calculated for the candidate kana sequence.
(S409) While selecting a plurality of candidate kana / kanji characters based on the morphological analysis, steps S403 to S408 are repeated to search for kana / kanji characters having the highest integrated score.
(S410) A conversion candidate is selected according to the case structure of the sentence including the searched kana / kanji.
(S411) The user is allowed to select any one kana / kanji from a plurality of conversion candidates. From the viewpoint of the user, the voice recognition function and the kana-kanji conversion function are visually recognized as the same candidate selection screen.

  FIG. 5 is a kana-kanji predictive conversion sequence comparing the prior art and the present invention.

  According to FIG. 5, in the case of speech recognition in the prior art, since the search is repeated with the acoustic score and the N-gram score, it is likely to be influenced by adjacent words and misconverted.

  On the other hand, in the case of speech recognition according to the present invention, the case grammar score and the history score are further included in the integrated score, and the search is repeated to classify them into optimum morphemes. Further, the case structure of the sentence is determined by the case frame processing, and finally converted into an optimal kana / kanji character.

  As described above in detail, according to the terminal, method, and program of the present invention, by using the kana-kanji conversion system for speech recognition, it is possible to integrate the language model resources of the entire system and to perform erroneous speech recognition conversion. Can be reduced.

  According to the present invention, the recognition accuracy in speech recognition is improved by compensating for misrecognition in speech recognition by the language score calculation unit in the kana-kanji conversion system. Also, by integrating the language model, the resources that the terminal should have are reduced, and can be incorporated into a portable terminal. Since the prediction candidate selection screens in the speech recognition system and the kana-kanji conversion system are shared, it is easy for the user to understand. Furthermore, the kana-kanji selected by the user in the kana-kanji conversion system is calculated with a high score in the same manner in the speech recognition system. Conversely, the kana / kanji character selected by the user in the speech recognition system is similarly calculated with a high score in the kana / kanji conversion system. That is, the same kana-kanji conversion is executed in both systems from the viewpoint of the user.

  Various changes, modifications, and omissions of the above-described various embodiments of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

DESCRIPTION OF SYMBOLS 1 Terminal 100 Display 111 Key operation part 112 Kana-kanji conversion engine 1121 Dependency grammar score calculation part 1122 Case grammar score calculation part 1123 History score calculation part 1124 Integrated score calculation part 1125 Morphological analysis part 1126 Case frame process part 113 Language model storage part 114 history storage unit 115 kana / kanji selection screen control unit 121 microphone 122 speech recognition engine 1221 acoustic analysis unit 1222 acoustic score calculation unit 1223 N-gram score calculation unit 1224 integrated score calculation unit 1225 candidate range control unit 123 acoustic model storage unit 124 language Model accumulation unit 125 Speech recognition selection screen control unit 130 Integrated score calculation unit 131 Search control unit 132 Candidate selection screen control unit 133 Language model accumulation unit

Claims (9)

Acoustic model storage means for storing an acoustic model used in speech recognition;
Language model storage means for storing a word dictionary and a language model corresponding to N-gram and case grammar;
Microphone means for acquiring a voice signal uttered by a user;
Acoustic analysis means for extracting phoneme frames from the speech signal;
Acoustic score calculation means for calculating an acoustic score given by the acoustic model for each of a plurality of candidate reading pseudonyms for the phoneme frame;
N-gram score calculating means for calculating an N-gram score according to the concatenation strength between words given by the language model for candidate reading kana series;
Case grammar score calculating means for calculating a case grammar score according to the connection strength between the clauses for the candidate kana series,
An integrated score calculation means for calculating an integrated score based on the acoustic score, the N-gram score, and the case grammar score for the candidate kana series,
In order to select a plurality of candidate kana-kanji characters based on morphological analysis and to search for kana-kanji characters having the highest integrated score, the acoustic score calculation means, the N-gram score calculation means, the case grammar score calculation means, and Search control means for repeating the integrated score calculation means;
A terminal having a case frame processing means for selecting a conversion candidate according to a case structure of a sentence including a searched kana / kanji, and realizing a voice recognition function. History accumulating means for storing kana-kanji characters converted in the past;
A history score calculating unit that calculates a history score according to a past learning history with respect to candidate reading kana series using the history storage unit,
The terminal according to claim 1, wherein the integrated score calculation unit further includes the history score in the integrated score. A key operation means for obtaining a hiragana string input by the user using a key;
The kana-kanji conversion function is realized by the N-gram score calculation means, the case grammar score calculation means, the integrated score calculation means, the search control means, and the case frame processing means. The terminal described in. Candidate selection screen control means for causing the user to select any one kana / kanji from a plurality of conversion candidates;
The terminal according to claim 3, wherein the candidate selection screen control means causes the user to visually recognize the same candidate selection screen for the voice recognition function and the kana-kanji conversion function. A speech recognition method on a terminal,
The terminal
An acoustic model storage unit that stores an acoustic model used in speech recognition;
A language model storage unit that stores a word dictionary and a language model corresponding to N-gram and case grammar;
A first step of acquiring an audio signal uttered by a user with a microphone unit;
A second step of extracting a phoneme frame from the speech signal;
A third step of calculating an acoustic score given by the acoustic model for each of a plurality of candidate reading pseudonyms for the phoneme frame;
A fourth step of calculating an N-gram score according to a connection strength between words given by the language model for a candidate kana sequence;
A fifth step of calculating a case grammar score according to the connection strength between clauses for the candidate kana series;
A sixth step of calculating an integrated score based on the acoustic score, the N-gram score, and the case grammar score for candidate kana series;
Selecting a plurality of candidate kana-kanji characters based on morphological analysis and repeating the third to sixth steps to search for kana-kanji characters having the highest integrated score;
A speech recognition method comprising: an eighth step of selecting a conversion candidate according to a case structure of a sentence including a searched kana / kanji character, and realizing a speech recognition function. The terminal further includes a history storage unit for storing kana-kanji characters converted in the past,
Using the history accumulating unit, further comprising a step of calculating a history score according to a past learning history for a candidate reading kana series;
The voice recognition method according to claim 5, wherein the sixth step further includes calculating the history score in the integrated score. The terminal further includes a key operation unit that acquires a hiragana sequence input by the user using a key,
The speech recognition method according to claim 5 or 6, wherein the kana-kanji conversion function is realized by the fourth step to the eighth step.   The speech recognition function and the kana-kanji conversion function are the same candidates for the user by further comprising the step of causing the user to select any one kana-kanji character from a plurality of conversion candidates. The speech recognition method according to claim 7, wherein the selection screen is visually recognized. A speech recognition program that allows a computer installed in a terminal to function,
Acoustic model storage means for storing an acoustic model used in speech recognition;
Language model storage means for storing a word dictionary and a language model corresponding to N-gram and case grammar;
Microphone means for acquiring a voice signal uttered by a user;
Acoustic analysis means for extracting phoneme frames from the speech signal;
Acoustic score calculation means for calculating an acoustic score given by the acoustic model for each of a plurality of candidate reading pseudonyms for the phoneme frame;
N-gram score calculating means for calculating an N-gram score according to the concatenation strength between words given by the language model for candidate reading kana series;
Case grammar score calculating means for calculating a case grammar score according to the connection strength between the clauses for the candidate kana series,
An integrated score calculation means for calculating an integrated score based on the acoustic score, the N-gram score, and the case grammar score for the candidate kana series,
In order to select a plurality of candidate kana-kanji characters based on morphological analysis and to search for kana-kanji characters having the highest integrated score, the acoustic score calculation means, the N-gram score calculation means, the case grammar score calculation means, and Search control means for repeating the integrated score calculation means;
A speech recognition program for causing a computer to function as a case frame processing means for selecting a conversion candidate according to a case structure of a sentence including a searched kana / kanji character.

Images