JP2005241829A - System and method for speech information processing, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition system that eliminates the need to repeat the same correction without word registration by a user. <P>SOLUTION: A speech input part 101 inputs a voice of a character string. A speech recognition part 201 converts the character string into a KANA(Japanese syllabary)-KANJI(Chinese character) mixed sentence. This KANA-KANJI mixed sentence is displayed on a display screen and correction by a user is accepted. A correction result generation part 102 generates correction result information, a corrected word generation part 103 generates a corrected word consisting of the input voice and the correction result information, and the corrected word is registered in a corrected word dictionary 106. The speech recognition part 201 performs recognition processing by using a recognition vocabulary dictionary 107 and a corrected word dictionary 106 in combination. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a voice information processing system, a voice information processing method, and a program for inputting a sentence by voice.

  In recent years, a Japanese dictation system capable of inputting Japanese sentences by voice has been put into practical use and has begun to be used as a practical system in various fields. The number of vocabulary that can be recognized by the system has also increased considerably from tens of thousands to hundreds of thousands of words. However, in reality, it is impossible to register all proper nouns, words unique to individual users, and coined words created every day in the system dictionary. On the other hand, if these words remain unregistered in the dictionary, the speech recognition system cannot correctly recognize the words no matter how carefully the user speaks. In the current system, these unregistered words are dealt with by additionally registering them in the recognized vocabulary dictionary referred to by the speech recognition system in the form of user words.

  The user word dictionary registration is usually realized by the user inputting and specifying “notation”, “reading (or pronunciation)”, and “part of speech” and pressing the dictionary registration button. Generally, the above three sets of information are input for each word, and user words are registered in the dictionary one word at a time.

The conventional speech recognition technology is disclosed in detail in Non-Patent Document 12, for example.
“Journal of Information Processing Society”, April 2000 issue (Vol.41 No.04), pp. 436-439, feature name: signpost, title: speech recognition technology, so far, author: Tatsuya Kawahara (Graduate School of Informatics, Kyoto University)

  The performance of speech recognition systems is improving year by year, but the recognition rate is not 100%. For this reason, when the system misrecognizes, it is necessary for the user to re-speak the misrecognized portion or to correct it by an input operation such as a keyboard. The main cause of misrecognition is that the word input by the user is not registered in the recognition vocabulary dictionary. Therefore, if a misrecognized word is registered in the dictionary as a user word, it can be prevented from being misrecognized thereafter. However, it is very troublesome to perform user word registration work by interrupting the input work and the text review every time erroneous recognition occurs during text input while considering the text. As a result, many users do not carry out user word registration work, but use the keyboard or the like to carry out only misrecognition correction work on the spot. However, since the word that the user wanted to input remains unregistered in this case, when the user utters the same word during subsequent sentence input, the speech recognition system again performs the same erroneous recognition, and the user repeats the same correction work. It must be made.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide a voice information processing system, a voice information processing method, and a program that do not require the user to repeat the same correction without registering words. To do.

  The speech information processing system according to the present invention includes a recognized vocabulary dictionary in which a plurality of first dictionary data including information related to reading of a vocabulary to be recognized by kana and information related to notation in kana / kanji and means for inputting speech And a means for generating a kana character string based on the input speech; and a kana / kanji character string generating means for generating a kana / kanji character string for the generated kana character string based on the recognition vocabulary dictionary; Display means for displaying the generated kana-kanji character string on a display screen; accepting means for accepting correction to the displayed kana-kanji character string; and kana based on the kana-kanji character string related to the correction Dictionary data generating means for generating second dictionary data including a character string and information on the contents of correction, and the generated second dictionary data as the recognized vocabulary dictionary Comprises a registration means for registering the different specific dictionary, the kana-kanji character string generation means, and performs the generation also based on the particular dictionary.

  In the present invention, the correction operation performed by the user on the kana / kanji character string generated by the recognition processing of the input speech is monitored, and the dictionary data including the kana character string and the information related to the correction is sent to a specific dictionary. Register the dictionary. As a result, even if the user does not register the word, the dictionary data can be used in addition to the normal recognition vocabulary dictionary in the next and subsequent language processing. Since the user can recognize, the user does not have to repeat the same correction operation.

The present invention relating to the apparatus is also established as an invention relating to a method, and the present invention relating to a method is also established as an invention relating to an apparatus.
Further, the present invention relating to an apparatus or a method has a function for causing a computer to execute means corresponding to the invention (or for causing a computer to function as means corresponding to the invention, or for a computer to have a function corresponding to the invention. It can also be realized as a program (for realizing the program), and can also be realized as a computer-readable recording medium on which the program is recorded.

  According to the present invention, the user does not have to repeat the same correction without registering words.

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

(First embodiment)
FIG. 1 shows a configuration example of a speech recognition system (natural language processing system) according to the first embodiment of the present invention.

  As shown in FIG. 1, the speech recognition system includes a speech input unit 101, a correction result generation unit 102, a correction word generation unit 103, a correction word dictionary registration unit 104, a correction word dictionary combined type speech recognition unit 105, a correction word. A dictionary 106 and a recognized vocabulary dictionary 107 are provided.

  The recognition vocabulary dictionary 107 stores a plurality of pieces of word information including sets of information such as “notation”, “reading”, and “part of speech”.

  The voice input unit 101 is a part that receives input voice data (200) from the user (100).

  The corrected word dictionary combined speech recognition unit 105 generates a speech recognition result (201) as described later. Here, the speech recognition result is a sentence mixed with kana and kanji.

  The generated kana-kanji mixed sentence is displayed on a display screen of a predetermined display device (not shown). The displayed kana-kanji mixed sentence accepts a correction operation from the user via a predetermined input device (not shown), and the correction result is displayed on the display screen of the predetermined display device. The

  The correction result generation unit 102 specifies a correction range from the speech recognition result (201) output from the corrected word dictionary combined speech recognition unit 105 and the correction operation 202 performed by the user on the speech recognition result, Correction result information (203) including a correction position in the recognition result and a correction result character string is generated. Here, for the correction operation that can be performed by the user, not only re-input by voice, but also all input devices for inputting character information such as a keyboard, a mouse, and a pen can be used. Further, a plurality of these devices may be combined during the correction operation.

  The correction word generation unit 103 generates a correction word (204) in which the pronunciation string of the input speech (200) received by the voice input unit 101 and the correction result information (203) generated by the correction result generation unit 102 are paired. . Here, the generation unit is called a word, but the character string to be registered may be a phrase or a short sentence. There is no restriction on the registered character string.

  The correction word dictionary registration unit 104 registers the correction word (204) generated by the correction word generation unit 103 in the correction word dictionary 106.

  The corrected word dictionary combined type speech recognition unit 105 recognizes speech by using the corrected word dictionary 106 and the recognized vocabulary dictionary 107 together (to recognize the input speech 200 to generate a kana character string and to store the kana character string. And the kana-kanji mixed character string 201 is generated). For the case where the correction word reading (the pronunciation string of the input speech) overlaps with the words in the recognition vocabulary dictionary, for example, the correction word is prioritized, or both words are presented and the user selects them. May be adopted.

  FIG. 2 shows an example of a schematic procedure related to recognition processing and correction processing in the speech recognition system of this embodiment.

  An input from the user is accepted (step S1). If this is an input to be recognized (step S2), a recognition process is performed (step S3), and a recognition result is output (step S4). On the other hand, if the input is for correction (step S2), correction information (in this embodiment, correction result information) is generated (step S5), and dictionary registration (in this embodiment) is performed based on the correction information. And registration of the correction word to the correction word dictionary) (step S6).

  In the conventional speech recognition system, unregistered words continue to be erroneously recognized thereafter unless the user additionally registers them in the recognition vocabulary dictionary. However, in the speech recognition system of the present embodiment, when a user performs a correction operation when erroneous recognition is performed and the operation result is automatically registered in the correction word dictionary as a correction word, when only the conventional recognition vocabulary dictionary is used Even in a place where erroneous recognition is continued, the correction operation previously performed by the user is applied and the recognition is correctly performed.

  Hereinafter, the present embodiment will be described using a specific example with reference to FIG.

  FIG. 3 shows specific examples of the operation of the correction result generation unit 102, the correction word generation unit 103, and the correction word dictionary registration unit 104, correction result information (203) and correction word (204) created in the process. It is.

  In this specific example, a case where a sentence “I will make efforts to break down the bone and bones” is considered. At this time, the user utters “Funkotsusai Shindo Rokushi”. Here, it is assumed that the word “crushed” is not registered in the recognition vocabulary dictionary 107.

  In Fig. 3, (a) is the first case (before correction / registration) that says "I'll do my best", and (b) is "I'll do my best" Is the second case (after correction / registration).

At this time, the voice recognition system performs the following erroneous recognition (S11).
Input voice: "Funkotsusaishindo Rokushi"
Recognition result: “I will do my best in powdered bones.”
In order to correct this erroneous recognition, the user performs the following correction operation (A).
Correction operation (A):
(I) Move the cursor to the right of “Latest”
(Ii) Delete the last two characters “latest”
(Iii) Say “Kukuku Shintai”
(Iv) Delete “ku” and “body” from the “crushed body” of the recognition result
(V) Move the cursor to the end of the sentence and prepare for the next utterance. In this embodiment, voice input is used to input a character string during correction operation. It does not matter.

The correction result by the above operation is as follows (S12).
Correction result: “I will make efforts to break down the bone and bone.”
From this correction operation result of the user, the correction result generation unit 102 detects that the character string at the position of the recognition result “latest” has been corrected to “crushed”, and “correction position: recognition” as correction result information (203). “Latest” in the result, correction result character string: shatter ”is output.

  Then, the correction word generation unit 103 uses the input speech and the correction result information (203) generated by the correction result generation unit 102 to generate “spoken string corresponding to the correction position of the input speech: Saishin” and “correction result character string: The correction word 204 “notation: shatter, reading: saishin” is generated in association with “shatter”.

  The correction word dictionary registration unit 104 registers the correction word in the correction word dictionary 106 (S13).

  Thereafter, the correction word dictionary combined speech recognition unit 105 performs speech recognition using the correction word dictionary 106 in which the correction word is registered and the recognition vocabulary dictionary 107 in combination. As a result, when the user inputs “Funkotsusaishin” next time, the correction word dictionary is referred to in the input portion of “Saishin”, thereby displaying the notation of “crushed” (S14).

  In the conventional method, unless the user registers in the dictionary the word “crushed”, the erroneous recognition of “latest bone” is repeated many times, but according to this embodiment, the user corrects only once. After that, the same misrecognition will not occur after that. Since the correction work is a work that the user naturally performs when the voice recognition system misrecognizes, no new effort or load is given to the user.

(Second Embodiment)
FIG. 4 shows a configuration example of a speech recognition system (natural language processing system) according to the second embodiment of the present invention.

  As shown in FIG. 4, the speech recognition system includes a speech input unit 101, a recognition vocabulary dictionary 107, a correction procedure generation unit 108, a correction macro generation unit 109, a correction macro dictionary registration unit 110, and a correction macro dictionary combined type speech recognition. A section 111 and a correction macro dictionary 112 are provided. In addition, the same code | symbol is attached | subjected to the part similar to FIG.

  An example of a schematic procedure related to recognition processing and correction processing in the voice recognition system of the present embodiment is the same as that in FIG.

  Below, it demonstrates centering on the point which is different from 1st Embodiment.

  The corrected macro dictionary combined speech recognition unit 111 generates a speech recognition result (201) as described later. Here, the speech recognition result is a sentence mixed with kana and kanji.

  The generated kana-kanji mixed sentence is displayed on a display screen of a predetermined display device (not shown). The displayed kana-kanji mixed sentence accepts a correction operation from the user via a predetermined input device (not shown), and the correction result is displayed on the display screen of the predetermined display device. The

  The correction procedure generation unit 108 specifies a correction range from the speech recognition result (201) output from the correction macro dictionary combined speech recognition unit 111 and the correction operation (202) performed by the user on the speech recognition result. Then, a correction procedure (206) representing the correction operation procedure is generated.

  The correction macro generation unit 109 generates a correction macro (207) in which the pronunciation sequence of the input voice (200) received by the voice input unit 101 and the correction procedure (206) generated by the correction procedure generation unit 108 are combined.

  The correction macro dictionary registration unit 110 registers the correction macro (207) generated by the correction macro generation unit 109 in the correction macro dictionary 112.

  The correction macro dictionary combined speech recognition unit 111 recognizes speech by using the correction macro dictionary 112 and the recognition vocabulary dictionary 107 together (to recognize the input speech 200 to generate a kana character string, and to store the kana character string. And the kana-kanji mixed character string 201 is generated). For the case where the reading of the correction macro (the pronunciation string of the input speech) overlaps with the word in the recognition vocabulary dictionary, for example, both the word related to the correction macro and the word related to the recognition vocabulary dictionary are given priority. An existing method such as presenting and allowing the user to select may be adopted.

  Hereinafter, the present embodiment will be described using a specific example with reference to FIG.

  FIG. 5 shows specific examples of operations of the correction procedure generation unit 108, the correction macro generation unit 109, and the correction macro dictionary registration unit 110 and the correction procedure (206) and correction macro (207) created in the process. is there.

  In this specific example, a case where a sentence “I will make efforts to break down the bone and bones” is considered. At this time, the user utters “Funkotsusai Shindo Rokushi”. Here, it is assumed that the word “crushed” is not registered in the recognition vocabulary dictionary.

  In Fig. 5, (a) is the first case (before correction / registration) that says "I'll do my best", and (b) is "I'll do my best" Is the second case (after correction / registration).

At this time, the voice recognition system performs the following erroneous recognition (S21).
Input voice: "Funkotsusaishindo Rokushi"
Recognition result: “I will do my best in powdered bones.”
In order to correct this misrecognition, the user performs a correction operation (A) (see the first embodiment).

The correction result by this operation is as follows (S22).
Correction result: “I will make efforts to break down the bone and bone.”
From this correction operation result of the user, the correction procedure generation unit 108 detects that the correction operation (A) has been performed on the character string at the position of the recognition result “latest”, and the “correction position: “Latest” in the recognition result, corrective action: Delete “Latest” → Say “Kakuku Shintai” → Output “Delete” and “Delete body”.

  Then, the correction macro generation unit 109 deletes “the utterance string corresponding to the correction position of the input speech: Saishin” and “correction operation:“ latest ”from the input speech and the correction procedure 206 generated by the correction procedure generation unit 108. → “Kakushintai” and utterance → “ku” and “Delete body” are associated and “Operation: Delete“ Latest ”→ Say“ Kakushintai ”→ Delete“ ku ”and“ Body ” A correction macro 207 “Read: Saishin” is generated.

  The correction macro dictionary registration unit 110 registers this correction macro in the correction macro dictionary 112 (S23).

  Thereafter, the corrective macro dictionary combined speech recognition unit 111 recognizes speech by using the corrective macro dictionary 112 in which the corrective macro is registered and the recognition vocabulary dictionary 107 in combination. As a result, the next time the user inputs “Funkotsusaishin”, the correction macro dictionary is referenced in the input portion of “Saishin”, so that “Delete“ Latest ”→ Say“ Kakushinsai ”→“ The correct operation “Delete“ Body ”” is automatically executed, and finally the notation “Crush” is displayed.

  The “reading” assigned to the correction macro does not have to be the utterance string of the corrected portion of the input speech before correction. For example, the utterance string input during the correction operation may be assigned as “reading”. In the above example, consider the assignment of the utterance string for the correction operation “Kakukushintai”. At this time, if you enter “Funkotsusaishindokushikaku”, you will misrecognize that “I will make the latest effort with powdered bones.” However, if you say “Kakushinshintai” here, the correction macro will be executed. The character string “latest” in the previous recognition result is replaced with “crushed”. Conventionally, it has been necessary to move the cursor and delete extra character strings, but this embodiment can eliminate the user's trouble.

  In the above example, the correction macro and the normal speech recognition word are handled without distinction, but a rule that utters a reserved word (for example, the word “correction macro”) before or after the correction macro may be used. . For example, in the above example, the character string “latest” in the immediately preceding recognition result is replaced with “crush” only when “correction macro” is spoken. Thereby, it is possible to avoid a correction macro being erroneously executed during normal voice input.

(Third embodiment)
FIG. 6 shows a configuration example of a speech recognition system (natural language processing system) according to the third embodiment of the present invention.

  As shown in FIG. 6, the speech recognition system includes a speech input unit 101, a correction result generation unit 102, a correction word generation unit 103, a recognition vocabulary dictionary 107, a user word dictionary registration automatic activation unit 113, and a user word dictionary combined type. A voice recognition unit 114 and a user dictionary 115 are provided. In addition, the same code | symbol is attached | subjected to the part similar to FIG.

  An example of a schematic procedure related to recognition processing and correction processing in the voice recognition system of the present embodiment is the same as that in FIG.

  Below, it demonstrates centering on the point which is different from 1st Embodiment.

  Similar to the recognition vocabulary dictionary 107, the user dictionary 115 is a plurality of registered word information composed of information sets such as “notation”, “reading”, and “part of speech”. As in the past, the recognized vocabulary dictionary 107 is a general dictionary, whereas the user dictionary 115 is a dictionary in which the user registers user words (108) as appropriate.

  The user word dictionary combined speech recognition unit 114 generates a speech recognition result (201) as described later. Here, the speech recognition result is a sentence mixed with kana and kanji.

  The generated kana-kanji mixed sentence is displayed on a display screen of a predetermined display device (not shown). The displayed kana-kanji mixed sentence accepts a correction operation from the user via a predetermined input device (not shown), and the correction result is displayed on the display screen of the predetermined display device. The

  The user word dictionary registration automatic activation unit 113 performs an operation for registering the correction word (204) generated by the correction word generation unit 103 in the user dictionary 115 as the user word (108). For example, a user word registration screen is displayed, and information on “notation”, “reading”, and “part of speech” necessary for registration is generated from the corrected word (204) and substituted. Note that there are various methods for “part of speech” to be assigned to the user word, such as a method of setting the same part of speech as the part of speech of the original word to be corrected, or a method of uniformly setting “noun”. . Here, the user can confirm the registered contents. If there is no problem with the registered contents, all you have to do is press the registration button. If correction is necessary, the user word is registered in the dictionary after appropriate correction.

  The user word dictionary combined speech recognition unit 114 recognizes speech by using the user dictionary 115 and the recognition vocabulary dictionary 107 together (recognizes the input speech 200 to generate a kana character string, and based on the kana character string A character string 201 mixed with kana and kanji is generated. For the case where the reading of the user word (the pronunciation string of the input speech) overlaps with the word in the recognition vocabulary dictionary, for example, the user word is given priority, or both words are presented to the user to select. May be adopted.

  Hereinafter, this embodiment will be described using a specific example with reference to FIG.

  FIG. 7 shows specific examples of operations of the correction result generation unit 102, the correction word generation unit 103, and the user word dictionary registration automatic activation unit 113, correction result information (203) and correction word (204) created in the process. It is a thing.

  In this specific example, a case where a sentence “I will make efforts to break down the bone and bones” is considered. At this time, the user utters “Funkotsusai Shindo Rokushi”. Here, it is assumed that the word “crushed” is not registered in the recognition vocabulary dictionary.

  In addition, in Fig. 7, (a) is the first case (case before correction / registration) that says "I'll do my best" and (b) is "I'll be my own." Is the second case (after correction / registration).

At this time, the voice recognition system performs the following erroneous recognition (S31).
Input voice: "Funkotsusaishindo Rokushi"
Recognition result: “I will do my best in powdered bones.”
In order to correct this misrecognition, the user performs a correction operation (A) (see the first embodiment).

The correction result by this operation is as follows (S32).
Correction result: “I will make efforts to break down the bone and bone.”
From this correction operation result of the user, the correction result generation unit 102 detects that the character string at the position of the recognition result “latest” has been corrected to “crushed”, and the correction result information 203 indicates “correction position: recognition result”. "The latest", correction result character string: shattered "is output.

  Then, the correction word generation unit 103 determines from the input speech and the correction result information 203 generated by the correction result generation unit 102 that “the utterance string corresponding to the correction position of the input speech: Saishin” and “correction result character string: shatter”. Are associated with each other to generate a correction word 204 of “notation: shattered, reading: saishin”.

  The user word dictionary registration automatic activation unit 113 registers the corrected word as the user word 208 in the user word dictionary 115 (S34). Here, before registering in the user word dictionary, a confirmation screen may be displayed to the user so that the registered content can be corrected (S33).

  The user word dictionary combined speech recognition unit 114 performs speech recognition using both the user word dictionary 106 in which the user word is registered and the recognition vocabulary dictionary 107. As a result, when the user inputs “Funkotsusaishin” next time, the user word dictionary is referred to in the input portion of “Saishin”, so that the notation “Cream” is displayed.

  In the conventional method, in order for the user to register the user word “crushing: saishin” in the dictionary, all of “notation: crushing”, “reading: saishin”, and “part of speech: noun” must be specified. According to this embodiment, since it is possible to automatically extract information necessary for user word registration from the contents of the user's correction operation, it is possible to easily register user words. As a result, the user is relieved from having to input all of “notation”, “reading”, and “part of speech” from 1 and the troublesomeness of user word registration is greatly reduced as compared with the case of performing normal user word registration. Is done.

(Fourth embodiment)
FIG. 8 shows a configuration example of a speech recognition system (natural language processing system) according to the fourth embodiment of the present invention.

  As shown in FIG. 8, the speech recognition system includes a speech input unit 101, a correction result generation unit 102, a correction word generation unit 103, a correction word dictionary registration unit 104, a correction word dictionary 106, a recognition vocabulary dictionary 107, and context. An extraction unit 120, a corrected word sequence table registration unit 121, a correction word dictionary and context table combined speech recognition unit 122, and a corrected word sequence table 123 are provided. In addition, the same code | symbol is attached | subjected to the part similar to FIG.

  An example of a schematic procedure related to recognition processing and correction processing in the voice recognition system of the present embodiment is the same as that in FIG.

  Below, it demonstrates centering on the point which is different from 1st Embodiment.

  The corrected word dictionary and the context table combined speech recognition unit 122 generates a speech recognition result (201) as described later. Here, the speech recognition result is a sentence mixed with kana and kanji.

  The generated kana-kanji mixed sentence is displayed on a display screen of a predetermined display device (not shown). The displayed kana-kanji mixed sentence accepts a correction operation from the user via a predetermined input device (not shown), and the correction result is displayed on the display screen of the predetermined display device. The

  The context extraction unit 120 corrects the speech recognition result (201) output from the corrected word dictionary and context table combined speech recognition unit 122 and the correction operation (202) performed by the user on the speech recognition result. The range is specified, and the contextual information (220) of the correction part is extracted.

  The correction word context table registration unit 121 sets the correction word 204 registered by the correction word dictionary registration unit 104 and the context information (220) of the correction portion generated by the context extraction unit 120 as a pair before and after the correction word. Relation information (221) is generated and registered in the corrected word context table 123.

  The corrected word dictionary and context table combined speech recognition unit 122 recognizes speech by using the corrected word dictionary 106 and the corrected word context table 123 and the recognition vocabulary dictionary 107 together (recognizing the input speech 200 and performing kana characters A string is generated, and a kana-kanji mixed character string 201 is generated based on the kana character string).

  In the first embodiment, correction word readings may overlap with words in the recognition vocabulary dictionary. For example, in this case, if the method of always giving priority to the correction word is adopted, as an example, in order to enter the sentence “I will make efforts to break down the bones and bones in the latest department.” The correct word is always used at “Saishin”, and it may be misrecognized as “I ’ll do my best to break down the bones in the department of shattering”.

  Therefore, in the present embodiment, the correction word context is extracted and registered as a correction word context table when the correction word dictionary is registered. When the input speech matches the readings of both the corrected word and the recognized vocabulary, the words before and after the corrected word are compared with the corrected word context table to determine which one of the corrected word and the recognized vocabulary is appropriate. select.

  Hereinafter, the present embodiment will be described using a specific example with reference to FIG.

  FIG. 9 shows the operation of the context extraction unit 120, the correction word context table registration unit 121, the correction word and context table combined speech recognition unit 122, and the context information (220) of the correction part created in the process. The specific example of the contextual information (221) of a correction word is shown.

  In this specific example, a case where a sentence “I will make efforts to break down the bone and bones” is considered. At this time, the user utters “Funkotsusai Shindo Rokushi”. Here, it is assumed that the word “crushed” is not registered in the recognition vocabulary dictionary.

  In addition, in Fig. 9, (a) is the first case (before correction / registration) that says "I'll do my best", and (b) is "I'll do my best" Is the second case (after correction / registration).

At this time, the voice recognition system performs the following erroneous recognition (S41).
Input voice: "Funkotsusaishindo Rokushi"
Recognition result: “I will do my best in powdered bones.”
In order to correct this misrecognition, the user performs a correction operation (A) (see the first embodiment).

The correction result by this operation is as follows (S42).
Correction result: “I will make efforts to break down the bone and bone.”
From the correction operation result of the user, the correction result generation unit 102 detects that the character string at the position of the recognition result “latest” has been corrected to “crushed”, and the correction result information 203 indicates “correction position: in recognition result. "Latest", correction result character string: shattered.

  At the same time, from the user correction operation 202 and the speech recognition result 201, the context extraction unit 120 detects that the correction operation (A) has been performed on the character string at the position of the recognition result “latest”, “Fracture” and “Effort” are detected as the words before and after the operation location, and “correction position: front =“ powder ”, after =“ effort ”” is output as the context 220 of the correction location.

  Then, the correction word generation unit 103 determines from the input speech and the correction result information 203 generated by the correction result generation unit 102 that “the utterance string corresponding to the correction position of the input speech: Saishin” and “correction result character string: shatter”. Are associated with each other to generate a correction word 204 of “notation: shattered, reading: saishin”. The corrected word dictionary registration unit 104 registers this corrected word in the corrected word dictionary 106 (S43).

  On the other hand, the correction word context table registration unit 121 sets “notation: shatter, reading: saishin” of the correction word 204 registered by the correction word dictionary registration unit 104 and “correction position: previous =“ From the word “powder” and after = “effort” ”, a corrected word context 221 of“ notation: shattered, reading: saishin ”: front =“ powder ”, back =“ effort ”is generated, and the correction word context table. 123 is registered (S44).

  The corrected word dictionary and context table combined speech recognition unit 122 performs speech recognition using the corrected word dictionary 106, the corrected word context table 123, and the recognition vocabulary dictionary 107 in combination. As a result, the next time the user inputs “Saishin no butsu”, the first “Saishin” part has a different context from that when the correction word was registered. Instead, the recognized vocabulary dictionary is referred to as before, and the notation "latest" is displayed (S45, S46). On the other hand, in the second “Saishin” part, because the preceding and following words match the words in the correction word context table, the correction word dictionary will be referred to, and the notation of “crushed” It is displayed (S45, S46). As a result, it is correctly recognized that “the latest department will make efforts to break down the bone and bone” (S47).

  As described above, when the correction word dictionary is used, the correction word can be applied only to an appropriate place by considering the context of the correction word at the time of registration.

  In the present embodiment, the front and rear of the correction word are used as the context, but only one of them may be used. In the present embodiment, the range used as the context is one word for both the front and rear, but a longer word string of two or more words may also be used. In the present embodiment, the word to be stored as the context is a word, but information attached to a word such as a part of speech may be used instead of the word. In addition, as a method for determining the degree of coincidence with the context table, not only whether or not the words before and after the correction word exist in the context table, but also a determination method using a probability value may be used.

(Fifth embodiment)
In the fifth embodiment of the present invention, the configuration considering the context information of the fourth embodiment is applied to the second embodiment. The effects and variations related to the configuration taking into account the contextual information shown in the fourth embodiment are also applicable to this embodiment.

  FIG. 10 shows a configuration example of a speech recognition system (natural language processing system) according to the fifth embodiment of the present invention.

  As shown in FIG. 10, the speech recognition system includes a speech input unit 101, a recognition vocabulary dictionary 107, a correction procedure generation unit 108, a correction macro generation unit 109, a correction macro dictionary registration unit 110, a correction macro dictionary 112, and a context. An extraction unit 120, a correction macro context table registration unit 124, a correction macro dictionary and context table combined speech recognition unit 125, and a correction macro context table 126 are provided. In addition, the same code | symbol is attached | subjected to the part similar to FIG.

  An example of a schematic procedure related to recognition processing and correction processing in the voice recognition system of the present embodiment is the same as that in FIG.

  Below, it demonstrates centering on the point which is different from 2nd Embodiment.

  The corrected macro dictionary and the context table combined speech recognition unit 125 generates a speech recognition result (201) as described later. Here, the speech recognition result is a sentence mixed with kana and kanji.

  The generated kana-kanji mixed sentence is displayed on a display screen of a predetermined display device (not shown). The displayed kana-kanji mixed sentence accepts a correction operation from the user via a predetermined input device (not shown), and the correction result is displayed on the display screen of the predetermined display device. The

  The context extraction unit 120 corrects from the speech recognition result (201) output by the correction macro dictionary and the context table combined speech recognition unit 125 and the correction operation (202) performed by the user on the speech recognition result. The range is specified, and the contextual information (220) of the correction part is extracted.

  The correction macro context table registration unit 124 sets the correction macro (207) registered by the correction macro dictionary registration unit 110 and the context information (220) of the correction part generated by the context extraction unit 120 as a pair. The context information (222) is generated and registered in the correction macro context table 126.

  The corrected macro dictionary and context table combined speech recognition unit 125 recognizes speech by using the correction macro dictionary 112 and the corrected macro context table 126 and the recognition vocabulary dictionary 107 together (speech recognition of the input speech 200 and kana characters). A string is generated, and a kana-kanji mixed character string 201 is generated based on the kana character string).

  Hereinafter, the present embodiment will be described using a specific example with reference to FIG.

  FIG. 11 shows the operation of the context extraction unit 120, the correction macro context table registration unit 124, the correction macro and context table combined speech recognition unit 125, and the context information (220) of the correction part created in the process. The specific example of the information (222) of the context of the correction macro is shown.

  In this specific example, a case where a sentence “I will make efforts to break down the bone and bones” is considered. At this time, the user utters “Funkotsusai Shindo Rokushi”. Here, it is assumed that the word “crushed” is not registered in the recognition vocabulary dictionary.

  In addition, in Fig. 11, (a) is the first case (before correction / registration) that says "I'll do my best" and (b) is "I'll be my own." Is the second case (after correction / registration).

At this time, the voice recognition system performs the following erroneous recognition (S51).
Input voice: "Funkotsusaishindo Rokushi"
Recognition result: “I will do my best in powdered bones.”
In order to correct this misrecognition, the user performs a correction operation (A) (see the first embodiment).

The correction result by this operation is as follows (S52).
Correction result: “I will make efforts to break down the bone and bone.”
From the correction operation result of the user, the correction procedure generation unit 108 detects that the correction operation (A) has been performed on the character string at the position of the recognition result “latest”. “Latest” in the result, corrective action: Delete “Latest” → Say “Kukuku Shintai” → Output “Delete” and “Delete body”.

  At the same time, the context extraction unit 120 detects that the correction operation (A) has been performed on the character string at the position of the recognition result “latest” from the user correction operation 202 and the speech recognition result 201, “Fracture” and “Effort” are detected as the words before and after the operation location, and “correction position: front =“ powder ”, after =“ effort ”” is output as the context 220 of the correction location.

  Then, the correction macro generation unit 109 deletes “the utterance string corresponding to the correction position of the input speech: Saishin” and “correction operation:“ latest ”from the input speech and the correction procedure 206 generated by the correction procedure generation unit 108 → Say “Kakuku Shintai” → “Delete Ku” “Delete body” and “Operation: Delete“ Latest ”→ Say“ Kaku Shintai ”→ Delete“ ku ”“ Body ”and read : A correction macro 207 "Saishin" is generated. The correction macro dictionary registration unit 110 registers this correction macro in the correction macro dictionary 112 (S53).

  On the other hand, the correction macro context table registration unit 124 deletes “operation: delete“ latest ”→ speak“ success ”from the correction macro 207 registered by the correction macro dictionary registration unit 110 →“ ku ”“ body ”. , "Reading: Saishin" and "Correction position: Before =" powder ", After =" Effort "" in the context 220 of the correction part, "Operation: Delete" Latest "→ Say" Kakushinshin "→" “Deleted” and “Body”, read: “Saishin”: Created a corrected macro context 222 of before = “flour” and after = “effort”, and registered it in the corrected macro context table 126 (S54).

  The correction macro dictionary and context table combined speech recognition unit 125 recognizes speech using the correction macro dictionary 112, the correction macro context table 126, and the recognition vocabulary dictionary 107 in combination. As a result, the next time the user inputs “Fill in the future,” the first “Saishin” part has a different context from that in the correction macro registration. Instead, the recognized vocabulary dictionary is referred to as before, and the notation “latest” is displayed (S55, S56). On the other hand, in the second “Saishin” part, since the preceding and following words match the words in the correction macro context table, the correction macro dictionary is referred to, and “Delete“ latest ”→ The corrective action of “Kakukushintai” utterance → “ku” “delete body” is automatically executed, and finally the notation “crushed” is displayed (S55, S56). As a result, it is correctly recognized that “the latest department will make efforts to break down the bone and bone” (S57).

  As described above, when the correction macro dictionary is used, the correction macro can be applied only to an appropriate portion by considering the context of the correction macro at the time of registration.

(Sixth embodiment)
In the sixth embodiment of the present invention, the configuration considering the contextual information of the fourth embodiment is applied to the third embodiment. The effects and variations related to the configuration taking into account the contextual information shown in the fourth embodiment are also applicable to this embodiment.

  FIG. 12 shows a configuration example of a speech recognition system (natural language processing system) according to the sixth embodiment of the present invention.

  As shown in FIG. 12, the speech recognition system includes a speech input unit 101, a correction result generation unit 102, a correction word generation unit 103, a recognition vocabulary dictionary 107, a user word dictionary registration automatic activation unit 113, a user dictionary 115, A relationship extraction unit 120, a user word context table registration unit 127, a user word dictionary and context table combined speech recognition unit 128, and a user word context table 129 are provided. In addition, the same code | symbol is attached | subjected to the part similar to FIG.

  An example of a schematic procedure related to recognition processing and correction processing in the voice recognition system of the present embodiment is the same as that in FIG.

  Below, it demonstrates focusing on the point which is different from 3rd Embodiment.

  The user word dictionary and the context table combined speech recognition unit 128 generates a speech recognition result (201) as described later. Here, the speech recognition result is a sentence mixed with kana and kanji.

  The generated kana-kanji mixed sentence is displayed on a display screen of a predetermined display device (not shown). The displayed kana-kanji mixed sentence accepts a correction operation from the user via a predetermined input device (not shown), and the correction result is displayed on the display screen of the predetermined display device. The

  The context extraction unit 120 corrects the speech recognition result (201) output from the user word dictionary and context table combined speech recognition unit 128 and the correction operation (202) performed by the user on the speech recognition result. The range is specified, and the contextual information (220) of the correction part is extracted.

  The user word context table registration unit 127 sets the user word (208) registered by the user word dictionary registration unit 113 and the context information (220) of the corrected portion generated by the context extraction unit 120 as a pair. Is created and registered in the user word context table 129.

  The user word dictionary and context table combined speech recognition unit 128 recognizes speech using the user word dictionary 115 and user word context table 129 and the recognition vocabulary dictionary 107 (recognizes the input speech 200 and recognizes the kana character string). And a character string 201 mixed with kana-kanji based on the kana character string).

  Hereinafter, this embodiment will be described using a specific example with reference to FIG.

  FIG. 13 shows the operation of the context extraction unit 120, the user word context table registration unit 127, the user word and context table combined speech recognition unit 128, and the context (220) of the correction points created in the process, The example of the information (223) of the context of a word is shown.

  In this specific example, a case where a sentence “I will make efforts to break down the bone and bones” is considered. At this time, the user utters “Funkotsusai Shindo Rokushi”. Here, it is assumed that the word “crushed” is not registered in the recognition vocabulary dictionary.

  In Fig. 13, (a) is the first case (before correction / registration) that says "I'll do my best", and (b) is "I'll do my best" Is the second case (after correction / registration).

  At this time, the voice recognition system performs the following erroneous recognition (S61).

Input voice: "Funkotsusaishindo Rokushi"
Recognition result: “I will do my best in powdered bones.”
In order to correct this misrecognition, the user performs a correction operation (A) (see the first embodiment).

The correction result by this operation is as follows (S62).
Correction result: “I will make efforts to break down the bone and bone.”
From this correction operation result of the user, the correction result generation unit 102 detects that the character string at the position of the recognition result “latest” has been corrected to “crushed”, and the correction result information 203 indicates “correction position: recognition result”. "The latest", correction result character string: shattered "is output.

  At the same time, from the user's correction operation 202 and the speech recognition result 201, the context extraction unit 120 detects that the correction operation (A) has been performed on the character string at the position of the recognition result “latest”, and “Fracture” and “Effort” are detected as the words before and after the operation location, and “correction position: front =“ powder ”, after =“ effort ”” is output as the context 220 of the correction location.

  Then, the correction word generation unit 103 determines from the input speech and the correction result information 203 generated by the correction result generation unit 102 that “the utterance string corresponding to the correction position of the input speech: Saishin” and “correction result character string: shatter”. Are associated with each other to generate a correction word 204 of “notation: shattered, reading: saishin”. The user word dictionary registration automatic activation unit 113 registers the corrected word as the user word 208 in the user word dictionary 115 (S64). Here, before registering in the user word dictionary, a confirmation screen may be displayed to the user so that the registered contents can be corrected (S63).

  On the other hand, the user word context table registration unit 127 displays “notation: shatter, reading: saishin” of the user word 208 registered by the user word dictionary registration automatic activation unit 113 and “correction position: previous” of the correction part context 220. = “Powdered bone”, After = “Effort” ”, the user word context 223 of“ notation: shattered, reading: Saishin ”: front =“ powdered bone ”, back =“ effort ”is generated, and the user word It registers in the context table 129 (S65).

  The user word dictionary and context table combined speech recognition unit 128 performs speech recognition using the user word dictionary 115, the user word context table 129, and the recognition vocabulary dictionary 107 in combination. As a result, the next time the user enters “Fill in the future,” the first “Saishin” part has a different context from that at the time of user word registration. Instead, the recognized vocabulary dictionary is referred to as usual, and the notation “latest” is displayed (S66, S67). On the other hand, in the second “Saishin” part, since the preceding and following words match the words in the user word context table, the user word dictionary will be referred to, and the notation of “crushed” It is displayed (S66, S67). As a result, it is correctly recognized that “the latest department will make efforts to break down the bone and bone” (S68).

  As described above, when the user word dictionary is used, the user word can be applied only to an appropriate portion by considering the context of the user word at the time of registration.

  By the way, in the first to sixth embodiments, the voice input unit recognizes input voice data from the user and outputs a kana character string, and each of the recognition units based on the kana character string, Instead of generating a kana-kanji mixed character string, instead, the input unit inputs a kana character string from the user by a kana character input device such as a keyboard device or a so-called soft key, and each recognition unit The present invention can also be applied to a case where a kana / kanji mixed character string is generated based on the kana character string. It is also possible to use both voice input and kana characters.

Each of the above functions can be realized even if it is described as software and processed by a computer having an appropriate mechanism.
The present embodiment can also be implemented as a program for causing a computer to execute predetermined means, causing a computer to function as predetermined means, or causing a computer to realize predetermined functions. In addition, the present invention can be implemented as a computer-readable recording medium that records the program.

  Note that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. In addition, various inventions can be formed by appropriately combining a plurality of components disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined.

The figure which shows the structural example of the speech recognition system which concerns on the 1st Embodiment of this invention. The flowchart which shows an example of the process sequence in the 1st-6th embodiment of this invention. The figure for demonstrating operation | movement of the correction result production | generation part of the speech recognition system which concerns on the 1st Embodiment of this invention, a correction word production | generation part, and a correction word dictionary registration part. The figure which shows the structural example of the speech recognition system which concerns on the 2nd Embodiment of this invention. The figure for demonstrating operation | movement of the correction procedure production | generation part of the same speech recognition system, a correction macro production | generation part, and a correction macro dictionary registration part The figure which shows the structural example of the speech recognition system which concerns on the 3rd Embodiment of this invention. The figure for demonstrating operation | movement of the correction result production | generation part of the same speech recognition system, a correction word production | generation part, and a user word dictionary registration automatic starting part. The figure which shows the structural example of the speech recognition system which concerns on the 4th Embodiment of this invention. The figure for demonstrating operation | movement of the context extraction part of the same speech recognition system, the correction word context table registration part, the correction word and context table combined speech recognition part The figure which shows the structural example of the speech recognition system which concerns on the 5th Embodiment of this invention. The figure for demonstrating operation | movement of the context extraction part of the same speech recognition system, a correction macro context table registration part, a correction macro, and a context table combined speech recognition part The figure which shows the structural example of the speech recognition system which concerns on the 6th Embodiment of this invention. The figure for demonstrating operation | movement of the context extraction part of the same speech recognition system, a user word context table registration part, a user word and a context table combined speech recognition part

Explanation of symbols

  DESCRIPTION OF SYMBOLS 101 ... Voice input part, 102 ... Correction result generation part, 103 ... Correction word generation part, 104 ... Correction word dictionary registration part, 105 ... Correction word dictionary combined type speech recognition part, 106 ... Correction word dictionary, 107 ... Recognition vocabulary dictionary 108 ... correction procedure generation unit 109 ... correction macro generation unit 110 ... correction macro dictionary registration unit 111 ... correction macro dictionary combined speech recognition unit 112 ... correction macro dictionary 113 ... user word dictionary registration automatic activation unit, 114... User word dictionary combined speech recognition unit 115. Correction word context table 124: Correction macro context table registration unit 125 ... Correction macro dictionary and context table combined speech recognition unit 126 ... Correction macro Rear relationship table, 127 ... user word context table registration unit, 128 ... user word context table 129 ... user word dictionary and context table combination speech recognition unit

Claims (8)

A recognized vocabulary dictionary in which a plurality of first dictionary data including information related to reading of kana words for information to be processed and information related to notation in kana kanji are registered;
Means for inputting voice;
Means for generating a kana character string based on the input voice;
A kana / kanji character string generating means for generating a kana / kanji character string for the generated kana character string based on the recognition vocabulary dictionary;
Display means for displaying the generated kana-kanji character string on a display screen;
Accepting means for accepting corrections to the displayed kana-kanji character string;
Dictionary data generating means for generating second dictionary data including a kana character string that is the basis of the kana-kanji character string related to the correction and information about the content of the correction;
Registration means for registering the generated second dictionary data in a specific dictionary different from the recognized vocabulary dictionary;
The kana / kanji character string generation means performs the generation based on the specific dictionary.   The dictionary data generation means includes at least a corrected kana-kanji character string related to the correction in the second dictionary data as information relating to the content of the correction. Voice information processing system.   3. The voice information processing system according to claim 2, wherein the specific dictionary is a user dictionary capable of user registration of dictionary data related to a desired vocabulary.   2. The voice according to claim 1, wherein the dictionary data generation unit includes information indicating an operation procedure related to the correction in the second dictionary data as information related to the content of the correction. Information processing system. Whether or not to apply the second dictionary data related to the correction in the generation of the kana / kanji character string by the kana / kanji character string generation means based on the generated kana / kanji character string and the correction to the kana / kanji character string Reference data generation means for generating reference data serving as a reference for determining whether or not
Storage means for storing the generated reference data in association with the second dictionary data;
5. The speech information processing system according to claim 1, wherein the kana-kanji character string generation unit performs the generation based on the reference data.   The reference data indicates that the second dictionary data is applied only when a specific kana / kanji character string exists at a location having a specific positional relationship with the kana character string related to the correction. The voice information processing system according to claim 5, wherein: A speech information processing method in a language processing apparatus having a recognized vocabulary dictionary in which a plurality of first dictionary data including information related to reading of a vocabulary to be processed by kana and information related to notation in kana / kanji are registered,
A voice input step for inputting voice;
A kana character string generation step for generating a kana character string based on the input voice;
A kana / kanji character string generating step for generating a kana / kanji character string for the generated kana character string based on the recognition vocabulary dictionary;
A display step of displaying the generated kana-kanji character string on a display screen;
An accepting step of accepting correction for the displayed kana-kanji character string;
A dictionary data generation step of generating second dictionary data including a kana character string that is a source of the kana-kanji character string related to the correction and information on the content of the correction;
A registration step of registering the generated second dictionary data in a specific dictionary different from the recognized vocabulary dictionary;
In the kana / kanji character string generation step, the generation is performed based on the specific dictionary. In a program for causing a computer to function as a speech information processing system including a recognized vocabulary dictionary in which a plurality of first dictionary data including information related to reading of a vocabulary to be processed by kana and information related to notation in kana / kanji are registered,
The program is
A voice input step for inputting voice;
A kana character string generation step for generating a kana character string based on the input voice;
A kana / kanji character string generating step for generating a kana / kanji character string for the generated kana character string based on the recognition vocabulary dictionary;
A display step of displaying the generated kana-kanji character string on a display screen;
An accepting step of accepting correction for the displayed kana-kanji character string;
A dictionary data generation step of generating second dictionary data including a kana character string that is a source of the kana-kanji character string related to the correction and information on the content of the correction;
Causing the computer to execute a registration step of registering the generated second dictionary data in a specific dictionary different from the recognized vocabulary dictionary;
In the kana-kanji character string generation step, the generation is performed based on the specific dictionary.

Images