JP2015138126A - Voice dictionary creation device, voice dictionary creation method and voice dictionary creation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve accuracy of correspondence between notation and reading when creating a dictionary using existing data.SOLUTION: A voice dictionary creation device includes: an input unit for acquiring a first notation and a first reading corresponding to the notation and the reading of character strings respectively; an extraction unit for extracting characters of a removal object estimated to have no portions corresponding to the first reading, out of the character strings in the first notation; a creation unit for creating a second notation in which the characters of the removal object are removed from the first notation; and a selection unit for, out of a plurality of temporary readings created based on the reading of characters in each of the first notation and the second notation, selecting the temporary reading which is the closest to the first reading as the second reading; and a registration unit for registering the notation and the reading of unknown words contained in the first notation in the dictionary based on the notation corresponding to the second reading out of the first notation and the second notation and the second reading.

Description

  The present invention relates to a speech dictionary creation device, a speech dictionary creation method, and a speech dictionary creation program.

  In recent years, services for acquiring information by voice have become widespread. A voice interaction system is used for this type of service. For example, a speech dialogue system has a speech dictionary in order to correctly recognize a user's speech and read (synthesize) a reply speech correctly. In the speech dictionary, character notations and readings are accumulated. For example, a dictionary creation device that creates a speech dictionary or the like estimates a reading for a notation obtained by morphological analysis, and registers the notation and the reading in association with each other (for example, Patent Documents 1-4).

  In recent years, there has been proposed a dictionary creation device that creates a dictionary using existing data stored in an address book of a mobile phone (for example, Patent Document 5). For example, this type of dictionary creation apparatus acquires text data from a storage unit that stores text data in which notation and reading are paired. Then, the dictionary creation device divides text data in which notation and reading are paired into units of words and phrases, and registers the notations and readings of the divided words and phrases in the dictionary.

JP 2002-41081 A JP 2009-204732 A JP 2000-207394 A JP 2004-151847 A JP 2009-258293 A

  In existing data, the correspondence between notation and reading may not be exact. For example, in data in which reading of some words in a character string is omitted, there is a discrepancy between the notation and reading in the data. If there is a discrepancy between the notation and the reading in the existing data used for creating the dictionary, the correspondence between the notation and the reading may fail. That is, in the method of creating a dictionary using existing data, when there is a discrepancy between the notation and the reading in the data, there is a possibility that the accuracy of the correspondence between the notation and the reading may be lowered.

  In one aspect, the speech dictionary creation device, the speech dictionary creation method, and the speech dictionary creation program disclosed herein improve the accuracy of correspondence between notation and reading when creating a dictionary using existing data. With the goal.

  According to one aspect, the speech dictionary creation device includes an input unit that acquires a first notation and a first reading corresponding to a notation and a reading of a character string, and a first reading among the character strings in the first notation. An extraction unit that extracts a character to be removed that is estimated to have no corresponding part, a creation unit that creates a second notation in which the character to be removed is removed from the first notation, and each of the first notation and the second notation A selection unit that selects, as a second reading, a temporary reading that is most similar to the first reading from a plurality of temporary readings that are generated based on the reading of characters in the notation; Based on the notation corresponding to the second reading and the second reading, it has a registration unit for registering the notation and reading of the unknown word included in the first notation in the dictionary.

  According to another aspect, the speech dictionary creation method acquires the first notation and the first reading corresponding to the notation and the reading of the character string, respectively, and corresponds to the first reading among the character strings in the first notation. Based on the reading of the characters in each of the first and second notations, extracting the characters to be removed that are presumed to have no part, creating the second notation by removing the characters to be removed from the first notation A temporary reading most similar to the first reading is selected as the second reading, and the notation corresponding to the second reading of the first notation and the second notation and the second reading are selected. Based on the above, the notation and reading of the unknown word included in the first notation are registered in the dictionary.

  According to another aspect, the speech dictionary creation program acquires the first notation and the first reading corresponding to the notation and the reading of the character string, respectively, and corresponds to the first reading among the character strings in the first notation. Based on the reading of the characters in each of the first and second notations, extracting the characters to be removed that are presumed to have no part, creating the second notation by removing the characters to be removed from the first notation A temporary reading most similar to the first reading is selected as the second reading, and the notation corresponding to the second reading of the first notation and the second notation and the second reading are selected. Based on the above, the computer is caused to execute processing for registering the notation and reading of the unknown word included in the first notation in the dictionary.

  The speech dictionary creation device, the speech dictionary creation method, and the speech dictionary creation program of the present disclosure can improve the accuracy of association between notation and reading when creating a dictionary using existing data.

It is a figure which shows one Embodiment of an audio dictionary creation apparatus and an audio dictionary creation method. It is a figure which shows an example of the extraction part shown in FIG. It is a figure which shows an example of the operation principle of the speech dictionary creation apparatus shown in FIG. It is a figure which shows an example of matching operation | movement. It is a figure which shows an example of operation | movement of the audio | voice dictionary creation apparatus shown in FIG. It is a figure which shows an example of operation | movement of the extraction part shown in FIG. It is a figure which shows an example of operation | movement of the selection part shown in FIG. It is a figure which shows an example of the operation | movement which estimates reading from the matching relationship before and behind a mismatching location. It is a figure which shows an example of operation | movement of the dictionary registration part shown in FIG. It is a figure which shows an example of the hardware constitutions of the audio | voice dictionary creation apparatus shown in FIG. It is a figure which shows an example of the speech dialogue system carrying the speech dictionary creation apparatus shown in FIG.

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

  FIG. 1 shows an embodiment of a speech dictionary creation device and a speech dictionary creation method. The speech dictionary creation apparatus DSYS of this embodiment creates a dictionary DICT for a speech dialogue system using data that can be acquired from an existing database DB, for example. The existing database DB is, for example, a Web page or a business database. Data transferred from the database DB to the speech dictionary creation device DSYS is, for example, data with reading (notation data ND10, reading data RD10).

  The notation data ND (ND10, ND20, ND30) is data indicating the notation of a character string, and the reading data RD (RD10, RD20, RD30, RDt1, RDt2) is data indicating the reading of a character string. Hereinafter, the notation indicated by the notation data ND is also referred to as notation ND, and the reading indicated by the reading data RD is also referred to as reading RD.

  In addition, for example, general words are registered in the dictionary DICT in advance. Hereinafter, a word registered in the dictionary DICT or the like in advance is also referred to as a known word. A word other than the known word is also referred to as an unknown word. Note that general words may be registered in advance in a dictionary different from the dictionary DICT.

  The speech dictionary creation device DSYS has, for example, an input unit INP, an extraction unit EXTR, a notation creation unit NGEN, a selection unit SEL, and a dictionary registration unit REGP. The input unit INP is an example of an input unit that acquires a first notation (notation ND10) and a first reading (reading RD10) corresponding to the notation and reading of a character string, respectively. For example, the input unit INP acquires data including notation data ND10 and reading data RD10 of the character string to be processed from the database DB. Then, the input unit INP transfers the notation data ND10 and the reading data RD10 to the extraction unit EXTR and the notation creation unit NGEN.

  The extraction unit EXTR is an example of an extraction unit that extracts characters to be removed that are estimated to have no portion corresponding to the first reading (reading RD10) in the character string in the first notation (notation ND10). For example, the extraction unit EXTR performs character analysis such as morphological analysis on the notation ND10, and extracts words (characters to be removed) that may not appear in reading. For example, in a database that collects restaurant information, words with the same reading notation arranged in different character types, words for advertisement, words of fixed phrases, etc. may not appear in reading.

  In this case, for example, with reference to the notation ND10, the extraction unit EXTR searches for a word in which the same reading notation is arranged in another character type, a word for advertisement, and a word of a fixed phrase. Then, when the extraction unit EXTR detects from the notation ND10 a word in which the same reading notation is arranged in another character type, a word for advertisement, or a word of a fixed sentence, it extracts the detected word as a character to be removed To do.

  In this manner, the extraction unit EXTR performs character analysis such as morphological analysis, and extracts characters to be removed that are estimated to have no portion corresponding to the reading RD10 from the character string in the notation ND10. Then, the extraction unit EXTR outputs the extraction data ED10 indicating the extraction result to the notation creation unit NGEN. For example, the extracted data ED10 includes data indicating a result of character analysis, data indicating a character to be removed, and the like.

  The notation creation unit NGEN is an example of a creation unit that creates a second notation (notation ND20) obtained by removing characters to be removed from the first notation (notation ND10). For example, the notation creation unit NGEN receives the notation data ND10, the reading data RD10, and the extraction data ED10. Then, the notation creation unit NGEN removes the character to be removed indicated by the extracted data ED10 from the notation ND10 and creates the notation ND20. The created notation data ND20 is transferred to the selection unit SEL together with the provisional reading data RDt2 corresponding to the notation ND20.

  For example, the notation creation unit NGEN transfers the notation data ND10, the provisional reading data RDt1, the notation data ND20, and the provisional reading data RDt2 to the selection unit SEL. The temporary reading RDt1 (reading indicated by the temporary reading data RDt1) is, for example, a reading candidate for the notation ND10.

  The provisional reading RDt1 of the notation ND10 is generated by the extraction unit EXTR, for example, when searching for words in which the same reading notation is arranged in different character types. In this case, provisional reading data RDt1 corresponding to the notation ND10 is included in the extracted data ED10. The notation creation unit NGEN creates a notation ND10 in which a plurality of provisional readings RDt1 are associated with each other when there are a plurality of provisional readings RDt1.

  The provisional reading RDt2 (the reading indicated by the provisional reading data RDt2) is a candidate for reading of the notation ND20. For example, the notation creation unit NGEN removes the temporary reading corresponding to the character to be removed from the temporary reading RDt1 to generate the temporary reading RDt2. Note that when there are a plurality of provisional readings RDt2 for one notation ND20, the notation creation unit NGEN creates a notation ND20 in which a plurality of provisional readings RDt2 are associated.

  Thus, the notation creation unit NGEN creates a plurality of notations ND (ND10, ND20). Therefore, a plurality of provisional readings RDt (RDt1, RDt2) generated based on the reading of characters in each notation ND (ND10, ND20) are also generated.

  The selection unit SEL is an example of a selection unit that selects a temporary reading (provisional readings RDt1 and RDt2) most similar to the first reading (reading RD10) as the second reading (reading RD20). For example, when there is a provisional reading RDt (RDt1, RDt2) that matches the reading RD10, the provisional reading RDt that is most similar to the reading RD10 is the provisional reading RDt that matches the reading RD10. The selection unit SEL includes, for example, a matching unit MAT and a determination unit JUDG.

  For example, the matching unit MAT receives the reading data RD10, the notation data ND10, the provisional reading data RDt1, the notation data ND20, and the provisional reading data RDt2 from the notation creation unit NGEN. Note that the matching unit MAT may receive the reading data RD10 from the input unit INP. Then, for example, the matching unit MAT performs DP matching between the reading RD10 and the provisional reading RDt1, and DP matching between the reading RD10 and the provisional reading RDt2. DP matching is matching in which one time axis to be compared is nonlinearly expanded and contracted so that both patterns are best matched.

  As described above, the matching unit MAT performs matching on the reading RD10 by using both the notation ND10 in which the removal target character is left and the notation ND20 in which the removal target character is removed. Then, matching unit MAT outputs the result of each matching to determination unit JUDG. For example, the matching unit MAT outputs a cost indicating the similarity to the determination unit JUDG. For example, the matching cost decreases as the degree of similarity increases.

  For example, the determination unit JUDG selects the temporary readings RDt1 and RDt2 that are most similar to the reading RD10 as the reading RD20. For example, the determination unit JUDG selects a provisional reading with the lowest matching cost as the reading RD20. Then, for example, the determination unit JUDG notifies the dictionary registration unit REGP of the notation ND corresponding to the reading RD20 of the notations ND10 and ND20 and the reading RD20. For example, the determination unit JUDG outputs information when the reading RD 20 is associated with the notation ND to the dictionary registration unit REGP.

  As described above, the selection unit SEL selects the provisional reading RDt most similar to the reading RD10 from the plurality of provisional readings RDt (RDt1, RDt2) generated based on the reading of the characters in the notations ND10 and ND20. Is selected as RD20.

  The dictionary registration unit REGP is an example of a registration unit that registers the notation and reading of an unknown word included in the first notation (notation ND10) in the dictionary DICT. For example, the dictionary registration unit REGP receives the notation data ND corresponding to the reading RD20 and the reading data RD20 from the determination unit JUDG. Then, the dictionary registration unit REGP registers, in the dictionary DICT, unknown words associated with notation and reading among unknown words included in the notation ND10.

  For example, the dictionary registration unit REGP receives information from the determination unit JUDG when the reading RD 20 is associated with the notation ND. The dictionary registration unit REGP may receive information from the extraction unit EXTR when the reading RD 20 is associated with the notation ND. For example, the dictionary registration unit REGP can identify which part of the reading RD 20 the character in the writing ND corresponds to based on information when the reading RD 20 is associated with the writing ND.

  For example, the dictionary registration unit REGP specifies the reading RD30 associated with the unknown word notation ND30 included in the notation ND. Then, the dictionary registration unit REGP outputs the unknown word notation data ND30 and the reading data RD30 to the dictionary DICT. Thus, the dictionary registration unit REGP registers the notation ND30 and the reading RD30 of the unknown word included in the notation ND10 in the dictionary DICT based on the notation ND20 and the reading RD20 corresponding to the reading RD20 of the notations ND10 and ND20. To do.

  Note that the dictionary registration unit REGP extracts two words when one of two words of different character types and arranged in the same reading is extracted as a character to be removed and the other of the two words is registered in the dictionary DICT. ND30 and reading RD30 may be registered in the dictionary DICT.

  Further, for example, when the reading RD20 does not match the reading RD10, the dictionary registration unit REGP may estimate the reading of the notation corresponding to the mismatched portion based on the reading RD10. As described above, the dictionary registration unit REGP, based on at least one of the readings RD10 and RD20 and the notation ND corresponding to the reading RD20 among the notations ND10 and ND20, the notation ND30 and the reading of the unknown word included in the notation ND10. Register RD30 in dictionary DICT.

  Note that the configuration of the speech dictionary creation apparatus DSYS is not limited to this example. For example, the speech dictionary creation device DSYS may include an output unit that outputs a notation or reading of an unknown word that could not be registered to a display or the like.

  FIG. 2 shows an example of the extraction unit EXTR shown in FIG. The extraction unit EXTR includes, for example, a character analysis unit ANALY, a reading estimation unit PRED, a continuous notation extraction unit CEXT, a pattern storage unit PMEM, and a pattern extraction unit PEXT.

  For example, the character analysis unit ANALY receives the notation data ND10 and the reading data RD10. Then, the character analysis unit ANALY performs morphological analysis and the like, and performs character analysis of the notation ND10. Note that the character analysis unit ANALY may delimit the character string in the notation ND10 based on a space delimiter, a symbol (parentheses, “&”, etc.) delimiter, a character type boundary, and the like. The character type is, for example, hiragana, katakana, kanji, or English.

  For example, the reading estimation unit PRED associates a temporary reading with words separated by character analysis of the character analysis unit ANALY. For example, the reading estimation unit PRED refers to the dictionary DICT and assigns a temporary reading to the word. Note that the reading estimation unit PRED assigns a temporary reading to an unknown word using single kanji reading, romaji reading, alphabet reading, phonics, or the like.

  The continuous notation extraction unit CEXT extracts, as a character to be removed, one of the words in the notation ND10 in which the same reading notation is arranged in another character type. For example, the continuous notation extraction unit CEXT searches for words in which the same reading notation is arranged in different character types based on the words separated by character analysis and the provisional reading. Then, for example, when the continuous notation extraction unit CEXT detects a word in which the same reading notation is arranged in different character types, the subsequent word is removed from the two detected words (words in line). Extract as target character. Note that the continuous notation extraction unit CEXT may extract the earlier word of the detected two words (words lined up) as a character to be removed.

  For example, in a database DB that collects information on restaurants, there are cases where words with the same reading notation arranged in different character types, promotional words, fixed-word words, etc. do not appear in reading. For example, when the notation ND10 is “Izakaya Marumaru marumaru” and the reading RD10 is “Izakaya Marumaru”, one of “Marumaru” and “marumaru” has no corresponding part in the reading RD10. That is, in the notation ND10, one of “maru” and “maramaru” does not appear in the reading RD10. In this case, the continuous notation extracting unit CEXT extracts one of “marumaru” and “maramaru” (for example, “maramaru”) as a character to be removed.

  For example, when the notation ND10 is “Izakaya Marumaru marumaru” and the reading RD10 is “Izakaya Marumaru”, “Izakaya Marumaru Marumaru” is assigned as one of the provisional readings RDt1 of the notation ND10. Therefore, the continuous notation extraction unit CEXT detects “Marumaru” and “marumaru” as words in which the same reading notation is arranged in different character types (kanji and English). Then, the continuous notation extraction unit CEXT extracts “marumaru”, which is a later word of the two detected words, as a character to be removed.

  The pattern storage unit PMEM stores, for example, advertising words, fixed phrases, and the like in advance. For example, the pattern storage unit PMEM stores a removal list in which advertising words, fixed phrase words, and the like are registered in advance. The advertising word is registered in advance in the removal list, for example, as a character string of a characteristic pattern (for example, a character string surrounded by an asterisk (*) or an asterisk). Also, for example, words indicating a cooking genre such as “Chinese cuisine” and “Yakiniku”, a word indicating a place name / road name / building name + branch name such as “XX shop”, etc., as removal words Registered in advance.

  The pattern extraction unit PEXT extracts words (promotional words, fixed phrases, etc.) stored in the pattern storage unit PMEM from the character string in the notation ND10 as characters to be removed. For example, if the notation ND10 is “Fuji Shokudo ** Immediately across the bridge across the station” and the reading RD10 is “Fujishoku Dou”, then the advertising word “Immediately across the bridge across the station” is “ There is no corresponding part in the reading RD10.

  In other words, among the notation ND10, “* immediately after the bridge crossing over the station” does not appear in the reading RD10. Further, the pattern storage unit PMEM stores, for example, a removal list indicating that a character string surrounded by asterisks (*) is a word for advertising. In this case, the pattern extraction unit PEXT selects “* ○○ station direction immediately across the bridge *” that applies to the character string of the characteristic pattern stored in the pattern storage unit PMEM among the character strings in the notation ND10. Extract as characters to be removed.

  Further, for example, when the notation ND10 is “Fuji Shokudo Kashimada store” and the reading RD10 is “Fujishoku Dou”, the word “Kashimada store”, which is a fixed phrase, has no corresponding part in the reading RD10. That is, in the notation ND10, “Kashimada store” does not appear in the reading RD10. Further, for example, in the removal list stored in the pattern storage unit PMEM, the place name “Kashimada” is registered as a word of a fixed phrase. In this case, the pattern extraction unit PEXT extracts “Kashimada branch” of the place name + branch name, which is a fixed phrase word stored in the pattern storage unit PMEM, as a character to be removed from the character string in the notation ND10. To do.

  In this manner, the extraction unit EXTR performs character analysis such as morphological analysis, and extracts characters to be removed that are estimated to have no portion corresponding to the reading RD10 from the character string in the notation ND10. Then, the extraction unit EXTR outputs the extraction data ED10 indicating the extraction result to the notation creation unit NGEN. For example, the extracted data ED10 includes data indicating a result of character analysis, data indicating a character to be removed, and the like.

  Note that the configuration of the extraction unit EXTR is not limited to this example. For example, when the word for advertisement and the word of the fixed phrase are not targeted for removal, the pattern storage unit PMEM and the pattern extraction unit PEXT may be omitted. In addition, for example, in the case where words with the same reading notation arranged in different character types are not the characters to be removed, the reading estimation unit PRED and the continuous notation extraction unit CEXT may be omitted. In this case, for example, the notation creation unit NGEN may include a reading estimation unit PRED.

  Alternatively, the extraction unit EXTR may include a dictionary in which general words are registered, for example, when general words are registered in advance in a dictionary different from the dictionary DICT. Further, the extraction unit EXTR may receive only the notation data ND10 out of the notation data ND10 and the reading data RD10.

  FIG. 3 shows an example of the operating principle of the speech dictionary creation device DSYS shown in FIG. Note that FIG. 3 shows a case where the notation ND10 is “Rakubara Marumaru” and the reading RD10 is “Rakunomiya Marumaru”.

  The speech dictionary creation device DSYS associates the tentative reading RDt1 with the character in the notation ND10 by, for example, word extraction and notation creation. For example, “Raku” for “Raku”, “Gaku” ..., “Flea” for “Drink”, “Inmi” ..., “Ya”, “Oku” for “Ya”, etc. .. “Maru”, “Gun”... Are associated with “Maru”. In addition, “marmaru” is associated with “Marmaru”, “M.U.

  Similarly, the speech dictionary creation device DSYS associates the provisional reading RDt2 with the characters in the notation ND20 by, for example, word extraction and notation creation. For example, “marumaru” and “marumaru” are determined to be the same reading, and “marumaru” is extracted as a character to be removed. Then, “Rakutakuya Marumaru” is created as the notation ND20 by removing the character “mararumaru” to be removed from the character string in the notation ND10. For example, the provisional reading RDt2 associated with the characters “Easy”, “Drink”, “Ya”, and “Maru” in the notation ND20 is the same as the provisional reading RDt1 of the “Raku drinking shop Marumaru” in the notation ND10 It is.

  Then, the speech dictionary creation device DSYS executes matching between the temporary reading RDt1 and the reading RD10, and matching between the temporary reading RDt2 and the reading RD10. As described above, the speech dictionary creation device DSYS executes a process of matching the notation ND10 with the provisional reading RDt1 and a process of matching the notation ND20 with the provisional reading RDt2. Then, the speech dictionary creation device DSYS determines each matching result, and registers the unknown word notation ND30 and the reading RD30 in the dictionary DICT based on the combination of the reading RD20 and the notation ND that most closely matches the reading RD10.

  For example, when the reading RD10 is “Rakunomiya Marumaru”, the provisional reading that most closely matches the reading RD10 is “Rakunomiya Marumaru” which is one of the provisional readings RDt2 associated with the notation ND20. In this case, a pair of “Rakubara” and “Rakunomiya” (notation ND30a, reading RD30a) and a pair of “Marumaru” and “Marumaru” (notation ND30b, reading RD30b) are registered in the dictionary DICT. Further, since the speech dictionary creating apparatus DSYS determines that the reading of “marumaru” is the same as “marumaru”, the pair of the notation ND30c and the reading RD30c of the character to be removed (“marumaru”, “Marumaru”) is also a dictionary. You may register with DICT.

  Here, for example, in a method in which the process of matching the provisional reading RDt2 with the notation ND20 is not performed, the unknown word notation is based on the result of the matching between the provisional reading RDt1 and the reading RD10 corresponding to the notation ND10. And the reading are registered in the dictionary DICT. For this reason, the notation “Marumaru marumaru” is associated with the provisional reading “Marmaru”, and a pair of “Marumaru marumaru” and “Marmaru” may be registered in the dictionary DICT. Also, for example, when the notation ND10 is “Fuji Shokudo Kashimada store” and the reading RD10 is “Fujishoku Dou”, in the method not using the notation ND20, a pair of “Fuji Shokudo Kashimada store” and “Fujishoku Dou” is registered in the dictionary DICT. There is a risk.

  On the other hand, as described with reference to FIG. 2, the speech dictionary creation device DSYS extracts “Kashimada store” as a fixed phrase word. Then, the speech dictionary creation device DSYS associates “Fujishoku Dou” as a temporary reading RDt2 with “Fuji Shokudo” which is the notation ND20 obtained by removing “Kashimada store” from the character string in the notation ND10. As a result, the voice dictionary creation device DSYS can register a pair of “Fuji Shokudo” and “Fujishokudo” in the dictionary DICT. Note that “Kashimada store” is not registered in the dictionary DICT because there is no corresponding reading in the reading RD10.

  As described above, the speech dictionary creation device DSYS creates the notation ND20 from which the character to be removed is removed and associates the provisional reading RDt2 even when there is a mismatch in the correspondence between the notation ND10 and the reading RD10. The accuracy of correspondence between reading and reading can be improved. That is, the speech dictionary creation device DSYS can improve the accuracy of association between notation and reading when creating a dictionary using existing data (a pair of the notation ND10 and the reading RD10).

  FIG. 4 shows an example of the matching operation. FIG. 4 shows an example of the matching operation when DP matching is used. DP matching is matching in which one time axis to be compared is nonlinearly expanded and contracted so that both patterns are best matched. In the example of FIG. 4, the cost of matching is calculated with “3” as the cost when there is a mismatch and “1” when the character is shifted by one.

  The actual notation ND10 and the actual reading RD10 in FIG. 4 indicate the notation ND10 and the reading RD10 acquired from the database DB by the input unit INP shown in FIG. For example, the actual notation ND10 is “Rakutakuya Fuji Shokudo”, and the actual reading RD10 is “Rakunomiyafujishokudo”. The provisional reading RDt1 is, for example, “Raku, Gaku ...”, “No, In ...”, “Mi”, “Ya, Ok ...”, “F” corresponding to each character in the notation ND10. , Tomi ... "," Ji, Shi ... "," Shoku, Ta, Ku ... "," Dou ".

  As an example of matching, three provisional readings RDt1 of “Rakunomiyafujishokudo”, “Gakunomiyafujishokudo”, and “Rakunomyokufujishokudo” are shown. When the provisional reading RDt1 is “Rakunomiya Fuji Dokdo”, since both the provisional reading RDt1 and the actual reading RD10 are “Rakunomiya Fuji Dokudo”, there is no misalignment or mismatched portion. For this reason, when the provisional reading RDt1 is “Rakunomya Fujishokudo”, the matching cost between the provisional reading RDt1 and the actual reading RD10 is “0”.

  When the provisional reading RDt1 is “Gakunomiyafujishokudo”, the provisional reading RDt1 is not misaligned with the actual reading RD10 “Rakunomiyafujishokudo”, but there is one mismatched portion. Exists (location surrounded by a thin broken line in FIG. 4). For this reason, when the provisional reading RDt1 is “Gakunomya Fujishokudo”, the matching cost between the provisional reading RDt1 and the actual reading RD10 is “3”, which is a mismatch cost.

  When the provisional reading RDt1 is “Rakunomiyakokudokudo”, the provisional reading RDt1 has a misalignment and a disagreement with the actual reading RD10 “Rakunomiyafujishokudo” ( (A portion surrounded by a thick broken line in FIG. 4). For this reason, when the provisional reading RDt1 is “Rakuno okufujishokudo”, the matching cost between the provisional reading RDt1 and the actual reading RD10 is the difference between the one character deviation cost “1” and the two character mismatch. It becomes “7” which is the sum of the cost “6”.

  Therefore, the selection unit SEL (for example, the determination unit JUDG of the selection unit SEL) selects “Rakunomiya Fuji” which is the provisional reading RDt1 with the lowest matching cost as the reading RD20 most similar to the actual reading RD10. select. In addition, the selection unit SEL selects the notation ND10 to which “Rakunomya Fujishokudo” is assigned as the provisional reading RDt1 as the notation to be referred to at the time of dictionary registration.

  For example, when “Fuji Shokudo” is stored in the pattern storage unit PMEM illustrated in FIG. 2, “Rakutakuya” is created as the notation ND20. In this case, the matching cost between the provisional reading RDt2 (for example, “Rakunomiya”) of the notation ND20 and the actual reading RD10 is the minimum value of the matching cost between the provisional reading RDt1 of the notation ND10 and the actual reading RD10 ( Greater than "0").

  Therefore, even when “Fuji Shokudo” is stored in the pattern storage unit PMEM, the selection unit SEL is the provisional reading RDt1 with the lowest matching cost as the reading RD20 most similar to the actual reading RD10. Select "Rakunomiya Fuji Shokoku". In addition, the selection unit SEL selects the notation ND10 to which “Rakunomya Fujishokudo” is assigned as the provisional reading RDt1 as the notation to be referred to at the time of dictionary registration. Thus, the selection unit SEL selects the notation ND corresponding to the reading RD20 from the notations ND10 and ND20.

  FIG. 5 shows an example of the operation of the speech dictionary creation apparatus DSYS shown in FIG. The operation of FIG. 5 may be realized only by hardware, or may be realized by controlling the hardware by software. For example, software such as a speech dictionary creation program may cause the computer to execute the operation of FIG. That is, the computer may read the storage medium storing the voice dictionary creation program and execute the operation of FIG. Note that the program that causes the computer to execute the operation of FIG. 5 is an aspect of the speech dictionary creation program.

  In step S100, the input unit INP acquires the notation ND10 and the reading RD10 corresponding to the notation and reading of the character string, respectively. For example, the input unit INP reads all pairs of the notation ND10 and the reading RD10 registered in the database DB from the database DB. The input unit INP stores, for example, a pair of the notation ND10 and the reading RD10 read from the database DB in a memory of the speech dictionary creation device DSYS.

  In step S200, the extraction unit EXTR extracts, for each notation ND10, characters to be removed that are estimated to have no portion corresponding to the reading RD10 from the character string in the notation ND10. For example, the extraction unit EXTR executes a process of extracting characters to be removed for all the notations ND10 acquired in step S100. As a result, words that may not appear in the reading, such as words in which the same reading notation is arranged in different character types, words for advertisement, words in fixed phrases, etc., are extracted for each notation ND10.

  In step S300, the notation creation unit NGEN creates the notation ND10 and the like associated with the provisional reading RDt. For example, the notation creation unit NGEN assigns a provisional reading RDt1 to the characters in the notation ND10 acquired in step S100. Note that the notation creation unit NGEN may use the result of the assignment performed in step S200, for example, when the provisional reading RDt1 is assigned to the characters in the notation ND10 in step S200.

  For example, when the removal target character is extracted in step S200, the notation creation unit NGEN removes the removal target character from the character string in the notation ND10 acquired in step S100 and creates the notation ND20. Then, the notation creation unit NGEN assigns a temporary reading RDt2 to the characters in the notation ND20. As a result, the notations ND10 and ND20 associated with the provisional reading RDt are created.

  In step S400, for example, the selection unit SEL selects a provisional reading most similar to the reading RD10 as the reading RD20 from a plurality of provisional readings RDt associated with the notation ND created in step S300. For example, the selection unit SEL selects the provisional reading most similar to the reading RD10 as the reading RD20 from the plurality of provisional readings RDt1 and RDt2 generated based on the reading of the characters in the notations ND10 and ND20. .

  In step S500, the dictionary registration unit REGP registers the unknown word notation ND30 and the reading RD30 included in the notation ND10 in the dictionary DICT based on the reading RD20 selected in step S400. For example, the dictionary registration unit REGP registers the notation ND30 and the reading RD30 of the unknown word included in the notation ND10 in the dictionary DICT based on the notation ND20 and the reading RD20 corresponding to the reading RD20 of the notations ND10 and ND20.

  As described above, when there is a mismatch in correspondence between the notation ND10 including the unknown word and the reading RD10, the speech dictionary creation device DSYS has no portion corresponding to the reading RD10 in the character string in the notation ND10. Extract the estimated characters to be removed. Then, the speech dictionary creation device DSYS performs matching between the provisional reading RDt2 and the reading RD10 corresponding to the notation ND20 from which the character to be removed is removed. Thereby, the speech dictionary creation device DSYS can improve the accuracy of the association between the unknown word notation ND30 and the reading RD30.

  The operation of the speech dictionary creation device DSYS is not limited to this example. For example, the speech dictionary creation device DSYS may output a notation or reading of an unknown word that could not be registered on a display or the like. Further, for example, the speech dictionary creation device DSYS may execute the processes of steps S200 to S500 again for the notation ND10 including the unknown word that could not be registered. In this case, for example, since the dictionary DICT updated by the first dictionary registration can be referred to, there is a possibility that the accuracy of provisional reading RDt allocation may be improved.

  FIG. 6 shows an example of the operation of the extraction unit EXTR shown in FIG. The operation of FIG. 6 corresponds to step S200 shown in FIG. For example, the operation of FIG. 6 is executed by the character analysis unit ANALY, the reading estimation unit PRED, the continuous notation extraction unit CEXT, and the pattern extraction unit PEXT shown in FIG.

  Note that the operation of FIG. 6 may be realized only by hardware, or may be realized by controlling the hardware by software. For example, software such as a speech dictionary creation program may cause the computer to execute the operation of FIG. That is, the computer may read the storage medium storing the voice dictionary creation program and execute the operation of FIG.

  In step S210, for example, the character analysis unit ANALY performs morphological analysis or the like, and performs character analysis of the notation ND10.

  In step S220, for example, the reading estimation unit PRED estimates the reading of the notation ND10. For example, the reading estimation unit PRED refers to the dictionary DICT or the like and associates the temporary reading with the words delimited by the character analysis in step S210. Thereby, provisional reading RDt1 with respect to notation ND10 is estimated. Note that the reading estimation unit PRED estimates a temporary reading for an unknown word using single kanji reading, romaji reading, alphabet reading, phonics, or the like.

  In step S230, for example, the pattern extraction unit PEXT extracts a character string of a word or a characteristic pattern pre-registered in a removal list or the like from a character string in the notation ND10. For example, the pattern extraction unit PEXT reads a removal list stored in the pattern storage unit PMEM. Then, the pattern extraction unit PEXT searches for a word pre-registered in the removal list or a word matching the character string of the characteristic pattern in the character string in the notation ND10.

  Then, for example, if the extraction unit EXTR detects from the notation ND10 a word that matches a word or the like registered in advance in the removal list, the extraction unit EXTR extracts the detected word as a character to be removed. For example, the words of the fixed phrases are registered in advance in a removal list or the like stored in the pattern storage unit PMEM. In addition, for example, advertising words are registered in advance in a removal list or the like as a character string of a characteristic pattern. In this case, the pattern extraction unit PEXT extracts a fixed phrase word or a promotion word as a character to be removed from the character string in the notation ND10.

  In step S240, for example, the continuous notation extraction unit CEXT extracts a word that is read in the same character type as the previous word from the character string in the notation ND10. For example, the continuous notation extraction unit CEXT refers to the provisional reading RDt1 of the notation ND10 estimated in step S220 and the notation ND10, and reads a word of the same reading with a different character type from the previous word as a character string in the notation ND10. Search from. Then, when the consecutive reading extraction unit CEXT detects a word having the same reading as another character type as the previous word, it extracts the detected word as a character to be removed.

  In this way, the extraction unit EXTR extracts characters to be removed from the character string in the notation ND10 in steps S230 and S240. For example, in step S240, the speech dictionary creation device DSYS performs the notation and reading registered in step S500 of FIG. 5 even if the notation ND10 includes words with the same reading notation in different character types. The accuracy of association can be improved.

  Further, for example, in step S230, the speech dictionary creation device DSYS improves the accuracy of the correspondence between the notation and the reading registered in step S500 of FIG. 5 even when the word for advertisement is included in the notation ND10. it can. Alternatively, the speech dictionary creation device DSYS can improve the accuracy of the correspondence between the notation and the reading registered in step S500 of FIG. 5 even when the fixed phrase word is included in the notation ND10 by step S230.

  The operation of the extraction unit EXTR is not limited to this example. For example, when the word for advertisement or the word of the fixed phrase is not used as the character to be removed, step S230 may be omitted. Further, for example, when words having the same reading notation in different character types are not to be removed, steps S220 and S240 may be omitted.

  FIG. 7 illustrates an example of the operation of the selection unit SEL illustrated in FIG. The operation in FIG. 7 corresponds to step S400 shown in FIG. The operation of FIG. 7 may be realized only by hardware, or may be realized by controlling the hardware by software. For example, software such as a speech dictionary creation program may cause the computer to execute the operation of FIG. That is, the computer may read the storage medium storing the voice dictionary creation program and execute the operation of FIG.

  In step S410, for example, the matching unit MAT associates a provisional reading RDt (RDt1, RDt2) with each of the notation ND20 from which the character to be removed is removed and the notation ND10 from which the character to be removed is left. The notation data ND is received. That is, the matching unit MAT receives the notation data ND created in step S300 of FIG. 5 from the notation creation unit NGEN.

  Further, the matching unit MAT receives the actual reading data RD10 from the notation creation unit NGEN. Thus, the matching unit MAT receives, for example, actual reading data RD10, notation data ND10, provisional reading data RDt1, notation data ND20, and provisional reading data RDt2. Note that the matching unit MAT may receive the actual reading data RD10 from the input unit INP.

  In step S420, for example, the matching unit MAT selects a provisional reading RDt to be matched from the plurality of provisional readings RDt received in step S410.

  In step S430, for example, the matching unit MAT performs matching between the provisional reading RDt selected in step S420 and the actual reading RD10 received in step S410. Thereby, for example, the cost of matching between the provisional reading RDt selected in step S420 and the actual reading RD10 received in step S410 is calculated.

  In step S440, for example, the matching unit MAT determines whether matching (matching process in step S430) has been completed for all the provisional readings RDt received in step S410. When there is a provisional reading RDt for which matching has not ended (No in Step S440), the operation of the selection unit SEL returns to Step S420. On the other hand, when matching has been completed for all provisional readings RDt (Yes in step S440), the operation of the selection unit SEL moves to step S450.

  In step S450, for example, the determination unit JUDG selects a temporary reading RDt that is most similar to the actual reading RD10. For example, the determination unit JUDG selects the provisional reading RDt with the lowest matching cost calculated in step S430 as the provisional reading RD20 most similar to the actual reading RD10.

  In step S460, for example, when the tentative reading RD20 selected in step S450 and the actual reading RD10 include a mismatched portion, the determination unit JUDG performs the correct reading of the mismatched portion of the temporary reading RD20 as the mismatched portion. Estimate from the matching relationship before and after. For example, as illustrated in FIG. 8, the determination unit JUDG refers to the actual reading RD10 and estimates a correct reading of the mismatched portion of the temporary reading RD20.

  Note that the operation of the selection unit SEL is not limited to this example. For example, step S460 may be omitted. Alternatively, step S460 may be executed by the dictionary registration unit REGP.

  FIG. 8 shows an example of an operation for estimating a reading from the matching relationship before and after the mismatched portion. In FIG. 8, the actual notation ND10 is “Adult Hideaway Drunk Restaurant Fuji Shokudo No. 1 Store”, and the actual reading RD10 is “Otonano Kakurega Yoinomiya Fuji Shokoku”. In addition, “No. 1 store” is extracted as a character to be removed. For this reason, “adult hideaway drunk bar Fuji restaurant” is created as the notation ND20 from which the character to be removed is removed. Also, as the reading RD20 most similar to the actual reading RD10, “Otonano Kakurega Yonoya Fujishokudo”, which is one of the provisional readings RDt2 of the notation ND20, is selected.

  In this case, “Otonano Kakurega” in the reading RD20 matches “Otonano Kakurega” in the reading RD10. Further, “Fujishoku Dou” of the reading RD20 matches “Fujishokudo” of the reading RD10. Note that “Yonoya” located between “Otonano Kakurega” and “Fujishoku Dou” in the reading RD20 does not coincide with “Yoinomiya” located between “Otonano Kakurega” and “Fujishoku Dou” in the reading RD10.

  For example, the determination unit JUDG presumes the correct reading of “Yonoya” in the reading RD20 from the reading RD10 because “Yonoya” before and after “Yonoya” in the reading RD20 (“Otonano Kakurega”, “Fujishokudo”) matches the reading RD10. . In this case, the determination unit JUDG estimates the reading of “drunk bar” corresponding to “Yonoya” in reading RD20 as “Yoinomiya” located between “Otonano Kakurega” and “Fujishoku Dou” in reading RD10. As a result, the reading of the mismatched portion (for example, “drunk bar”) is estimated to be the correct reading (for example, “Yoinomiya”).

  As described above, the speech dictionary creation device DSYS can guess the reading of the mismatched portion even when the temporary reading RDt that matches the actual reading RD10 cannot be created. As a result, the speech dictionary creation device DSYS can improve the accuracy of association between the unknown word notation ND30 and the reading RD30.

  FIG. 9 shows an example of the operation of the dictionary registration unit REGP shown in FIG. The operation of FIG. 9 corresponds to step S500 shown in FIG. The operation of FIG. 9 may be realized only by hardware, or may be realized by controlling the hardware by software. For example, software such as a speech dictionary creation program may cause a computer to execute the operation of FIG. That is, the computer may read the storage medium storing the voice dictionary creation program and execute the operation of FIG.

  In the operation of FIG. 9, when a plurality of readings RD30 are detected for one unknown word notation ND30, the plurality of readings RD30 are registered with priority. For example, in the operation of FIG. 9, the number registered in the dictionary DICT is registered as the priority order.

  In step S510, the dictionary registration unit REGP receives, from the selection unit SEL, a pair of the notation ND and the reading RD20 with the lowest matching cost.

  In step S520, the dictionary registration unit REGP selects a processing target word (notation and reading pair) from the pair of notation ND and reading RD20 received in step S510. For example, the dictionary registration unit REGP selects a word associated with a reading from words in the notation ND.

  In step S530, the dictionary registration unit REGP determines whether or not the word selected in step S520 is a known word. The known word is, for example, a general word registered in advance in the dictionary DICT. When the word selected in step S520 is a known word (Yes in step S530), the operation of the dictionary registration unit REGP proceeds to step S570. That is, when the word selected in step S520 is a known word, the dictionary registration unit REGP does not register the word selected in step S520 in the dictionary DICT.

  On the other hand, when the word selected in step S520 is not a known word (No in step S530), the operation of the dictionary registration unit REGP proceeds to step S540. In the description of steps S540, S550, and S560, the word notation and reading pair selected in step S520 is also referred to as an unknown word notation ND30 and reading RD30 pair.

  In step S540, the dictionary registration unit REGP determines whether the same pair (notation ND30, reading RD30) of the unknown word notation ND30 and reading RD30 selected in step S520 is already registered in the dictionary DICT. When the same pair as the unknown word notation ND30 and reading RD30 is already registered in the dictionary DICT (Yes in step S540), the operation of the dictionary registration unit REGP moves to step S560. On the other hand, when the same pair of the unknown word notation ND30 and reading RD30 is not registered in the dictionary DICT (No in step S540), the operation of the dictionary registration unit REGP proceeds to step S550.

  In step S550, the dictionary registration unit REGP registers the pair of unknown word notation ND30 and reading RD30 in the dictionary DICT. Thereby, the registration operation for the pair of the unknown word notation ND30 and the reading RD30 selected in step S520 is completed.

  In step S560, the dictionary registration unit REGP adds “1” to the value of the registered number n of the unknown word notation ND30 and the reading RD30. Thereby, the registration operation for the pair of the unknown word notation ND30 and the reading RD30 selected in step S520 is completed. Note that the registered number n corresponds to, for example, the priority of the reference order when a plurality of readings RD30 are registered for one notation ND30.

  In step S570, the dictionary registration unit REGP determines whether there is a word to be processed in the notation ND received in step S510. For example, the dictionary registration unit REGP determines whether or not the processing in step S530 and the like has been executed for all words associated with the reading in the notation ND received in step S510. When there is a word to be processed in the notation ND (Yes in step S570), the operation of the dictionary registration unit REGP returns to step S520. On the other hand, when there is no word to be processed in the notation ND (No in step S570), the operation of the dictionary registration unit REGP regarding the pair of the notation ND and the reading RD20 received in step S510 ends.

  The operation of the dictionary registration unit REGP is not limited to this example. For example, the dictionary registration unit REGP may select the reading RD30 most frequently detected for one notation ND30 and register it in the dictionary DICT. In this case, one selected pair (an unknown word notation ND30 and a reading RD30 pair) is registered in the dictionary DICT. Alternatively, step S560 may be omitted. In this case, the reading RD30 first detected for one notation ND30 is registered in the dictionary DICT as the reading RD30 of the notation ND30. Further, steps S540 and S560 may be omitted. In this case, the last reading RD30 detected for one notation ND30 is registered in the dictionary DICT as the reading RD30 of the notation ND30.

  As described above, the speech dictionary creating apparatus DSYS, the speech dictionary creating method, and the speech dictionary creating program according to the embodiment shown in FIGS. 1 to 9 match the provisional reading RDt1 and the reading RD10, and the provisional reading RDt2 and the reading RD10. Perform matching with. Then, for example, the speech dictionary creation device DSYS uses the unknown word notation ND30 and the reading RD30 based on the matching result between the temporary reading RDt1 and the reading RD10 and the matching result between the temporary reading RDt2 and the reading RD10. Is registered in the dictionary DICT.

  For example, the speech dictionary creation device DSYS includes an input unit INP that acquires the notation ND10 and the reading RD10, an extraction unit EXTR that extracts characters to be removed from the notation ND10, a notation creation unit NGEN, a selection unit SEL, and a dictionary registration Part REGP. For example, the notation creation unit NGEN creates a notation ND20 in which the character to be removed is removed from the notation ND10.

  The selection unit SEL selects the provisional reading RDt most similar to the reading RD10 as the reading RD20 from a plurality of provisional readings RDt generated based on the reading of the characters in the notations ND10 and ND20. For example, the selection unit SEL selects the temporary reading RDt most similar to the reading RD10 based on the matching between the temporary reading RDt1 and the reading RD10 and the matching between the temporary reading RDt2 and the reading RD10.

  Further, the dictionary registration unit REGP registers the notation ND30 and the reading RD30 of the unknown word included in the notation ND10 in the dictionary DICT based on the notation ND20 and the reading RD20 corresponding to the reading RD20 of the notations ND10 and ND20. In this way, the speech dictionary creation device DSYS does notation of unknown words based on the matching of the notation ND20 and the provisional reading RDt2 obtained by removing the character to be removed from the notation ND10, and the matching of the notation ND10 and the provisional reading RDt1. ND30 and reading RD30 are registered in the dictionary DICT.

  Thereby, in this embodiment, even when the correspondence between the notation ND10 and the reading RD10 is inconsistent, the accuracy of the association between the unknown word notation and the reading can be improved. That is, in this embodiment, when the dictionary DICT is created using existing data (a pair of the notation ND10 and the reading RD10), the accuracy of association between the notation and the reading can be improved.

  FIG. 10 shows an example of the hardware configuration of the speech dictionary creation apparatus DSYS shown in FIG. The same elements as those described in FIGS. 1 to 9 are denoted by the same reference numerals, and detailed description thereof will be omitted.

  The computer device CP includes a processor PU, a memory MEM, a hard disk device HDD, an input / output interface IF, and an optical drive device ODR. The processor PU, the memory MEM, the hard disk device HDD, the input / output interface IF, and the optical drive device ODR are connected to each other via a bus BUS. For example, the function of the speech dictionary creation device DSYS is realized by a processor PU, a memory MEM, a hard disk device HDD, and an input / output interface IF.

  The optical drive device ODR can be mounted with a removable disk DIS such as an optical disk, and reads and records information recorded on the mounted removable disk DIS. The computer apparatus CP communicates with the outside of the computer apparatus CP via, for example, the input / output interface IF. For example, the computer device CP acquires the notation ND10 and the reading RD10 from the database DB via the input / output interface IF. The computer device CP may store the notation ND10 and the reading RD10 acquired from the database DB in the memory MEM or the hard disk device HDD.

  The memory MEM stores, for example, the operating system of the computer device CP. Further, the memory MEM stores an application program such as a speech dictionary creation program for the processor PU to execute the operation of the speech dictionary creation device DSYS.

  An application program such as an audio dictionary creation program can be recorded and distributed on a removable disk DIS such as an optical disk. For example, the computer device CP may read an application program such as a speech dictionary creation program from the removable disk DIS via the optical drive device ODR and store it in the memory MEM or the hard disk device HDD.

  In addition, the computer device CP may download an application program such as a speech dictionary creation program via a communication device connected to a network such as the Internet and store it in the memory MEM or the hard disk device HDD. Further, the computer device CP may acquire the notation ND10 and the reading RD10 from the Web page via a communication device connected to a network such as the Internet. Further, the computer device CP may store the notation ND10 and the reading RD10 acquired from the Web page in the memory MEM or the hard disk device HDD.

  The hardware configuration of the speech dictionary creation device DSYS is not limited to this example. For example, the optical drive device ODR may be omitted from the computer device CP.

  FIG. 11 shows an example of a spoken dialogue system VSYS equipped with the speech dictionary creation device DSYS shown in FIG. The same elements as those described in FIGS. 1 to 9 are denoted by the same reference numerals, and detailed description thereof will be omitted.

  The voice interaction system VSYS can communicate with a terminal TERM such as a mobile phone or a computer. For example, the terminal TERM receives the user's voice and outputs voice data to the voice dialogue system VSYS. Alternatively, voice data is received from the voice dialogue system VSYS and voice is output.

  The voice dialogue system VSYS has, for example, a dictionary DICT, a voice dictionary creation device DSYS, a voice recognition unit VREC, a dialogue control unit DCTL, and a voice synthesis unit VSYN. The dictionary DICT created by the speech dictionary creation device DSYS is used by, for example, the speech recognition unit VREC and the speech synthesis unit VSYN.

  For example, the voice recognition unit VREC receives voice data from the terminal TERM and generates text data (speech recognition result) indicating the contents of the voice data with reference to the dictionary DICT. Then, the voice recognition unit VREC outputs the voice recognition result (text data) to the dialogue control unit DCTL. The dialogue control unit DCTL analyzes the speech recognition result and acquires necessary text information from an external service or Web information. Then, the dialogue control unit DCTL outputs the acquired text information to the voice synthesis unit VSYN.

  The speech synthesizer VSYN performs speech synthesis and generates speech data corresponding to the text data. For example, the speech synthesis unit VSYN generates speech data corresponding to text information received from the dialogue control unit DCTL with reference to the dictionary DICT. Then, the speech synthesis unit VSYN outputs speech data (speech synthesis result) to the terminal TERM.

  Thereby, the user can acquire information by voice. For example, the user issues “I want to hear today's news” to the terminal TERM. The voice interactive system VSYS receives voice data corresponding to “I want to hear today's news” via the terminal TERM. Then, the voice dialogue system VSYS analyzes voice data with reference to the dictionary DICT or the like, and obtains text information relating to today's news (for example, “mixed medical prohibition is a legal supreme court”) from an external service or Web information. The voice interactive system VSYS generates voice data corresponding to text information about today's news and outputs it to the terminal TERM. Accordingly, for example, the terminal TERM outputs, as today's news, a voice saying “Prohibition of mixed medical treatment is a legal supreme court”.

The invention described in the above embodiments is organized and disclosed as an appendix.
(Appendix 1)
An input unit for obtaining the first notation and the first reading corresponding to the notation and the reading of the character string, respectively;
An extraction unit for extracting a character to be removed that is estimated to have no portion corresponding to the first reading in the character string in the first notation;
A creation unit for creating a second notation in which the character to be removed is removed from the first notation;
Selection for selecting the temporary reading most similar to the first reading as the second reading from a plurality of temporary readings generated based on the reading of the characters in the first notation and the second notation And
A registration unit that registers the notation and reading of an unknown word included in the first notation in a dictionary based on the notation corresponding to the second reading and the second reading of the first notation and the second notation. A voice dictionary creation device comprising:
(Appendix 2)
In the speech dictionary creation device according to attachment 1,
The extraction unit generates the tentative reading corresponding to the first notation among a plurality of the tentative readings, and arranges the words in the first notation with different character types and in the same reading. One of two words is extracted as the character to be removed.
(Appendix 3)
In the speech dictionary creation device according to attachment 2,
When the registration unit extracts one of the two words as the character to be removed and registers the other notation and reading of the two words in the dictionary, the notation and reading of the two words Is registered in the dictionary.
(Appendix 4)
In the speech dictionary creation device according to any one of supplementary notes 1 to 3,
The extraction unit extracts at least one of a pre-registered removal target word and a pre-registered characteristic pattern character string as the removal target character.
(Appendix 5)
Obtain the first notation and the first reading corresponding to the notation and reading of the character string,
Extracting a character to be removed that is estimated to have no portion corresponding to the first reading from the character string in the first notation;
Creating a second representation in which the characters to be removed are removed from the first representation;
Selecting the temporary reading most similar to the first reading as a second reading from a plurality of temporary readings generated based on the reading of the characters in each of the first notation and the second notation;
Registering the notation and the reading of the unknown word included in the first notation in the dictionary based on the notation corresponding to the second reading of the first notation and the second notation and the second reading. A featured voice dictionary creation method.
(Appendix 6)
In the speech dictionary creation method according to attachment 5,
The temporary reading corresponding to the first notation among the plurality of temporary readings is generated, and one of two words arranged in the same reading with different character types among the words in the first notation Is extracted as the character to be removed.
(Appendix 7)
In the speech dictionary creation method according to attachment 6,
When one of the two words is extracted as the character to be removed and the other notation and reading of the two words are registered in the dictionary, the notation and reading of the two words are stored in the dictionary. A voice dictionary creation method characterized by registration.
(Appendix 8)
In the speech dictionary creation method according to any one of appendix 5 to appendix 7,
A speech dictionary creation method, wherein at least one of a word to be removed registered in advance and a character string having a characteristic pattern registered in advance is extracted as the character to be removed.
(Appendix 9)
Obtain the first notation and the first reading corresponding to the notation and reading of the character string,
Extracting a character to be removed that is estimated to have no portion corresponding to the first reading from the character string in the first notation;
Creating a second representation in which the characters to be removed are removed from the first representation;
Selecting the temporary reading most similar to the first reading as a second reading from a plurality of temporary readings generated based on the reading of the characters in each of the first notation and the second notation;
A process of registering an unknown word notation and a reading included in the first notation in a dictionary based on a notation corresponding to the second reading and the second reading of the first notation and the second notation. An audio dictionary creation program that is executed by a computer.
(Appendix 10)
A dictionary,
A voice dictionary creation device for creating the dictionary;
A voice recognition unit that generates text data indicating the content of voice data received from outside with reference to the dictionary;
A dialog control unit that receives the text data from the voice recognition unit and obtains text information according to the text data;
A voice synthesis unit that receives the text information from the dialogue control unit and generates voice data corresponding to the text information with reference to the dictionary;
The voice dictionary creation device
An input unit for obtaining the first notation and the first reading corresponding to the notation and the reading of the character string, respectively;
An extraction unit for extracting a character to be removed that is estimated to have no portion corresponding to the first reading in the character string in the first notation;
A creation unit for creating a second notation in which the character to be removed is removed from the first notation;
Selection for selecting the temporary reading most similar to the first reading as the second reading from a plurality of temporary readings generated based on the reading of the characters in the first notation and the second notation And
Registration for registering the notation and the reading of the unknown word included in the first notation in the dictionary based on the notation corresponding to the second reading and the second reading of the first notation and the second notation And a voice dialogue system characterized by comprising:

  From the above detailed description, features and advantages of the embodiments will become apparent. This is intended to cover the features and advantages of the embodiments described above without departing from the spirit and scope of the claims. Also, any improvement and modification should be readily conceivable by those having ordinary knowledge in the art. Therefore, there is no intention to limit the scope of the inventive embodiments to those described above, and appropriate modifications and equivalents included in the scope disclosed in the embodiments can be used.

  ANALY: Character analysis unit; CEXT: Continuous notation extraction unit: CP: Computer device; DCTL: Dialogue control unit; Unit: JUDG ... determination unit; MAT ... matching unit; MEM ... memory; NGEN ... notation creation unit; ODR ... optical drive device; PEXT ... pattern extraction unit; PMEM ... pattern storage unit; REGP ... dictionary registration part; SEL ... selection part; VREC ... voice recognition part; VSYN ... voice synthesis part; VSYS ... voice dialogue system

Claims (6)

An input unit for obtaining the first notation and the first reading corresponding to the notation and the reading of the character string, respectively;
An extraction unit for extracting a character to be removed that is estimated to have no portion corresponding to the first reading in the character string in the first notation;
A creation unit for creating a second notation in which the character to be removed is removed from the first notation;
Selection for selecting the temporary reading most similar to the first reading as the second reading from a plurality of temporary readings generated based on the reading of the characters in the first notation and the second notation And
A registration unit that registers the notation and reading of an unknown word included in the first notation in a dictionary based on the notation corresponding to the second reading and the second reading of the first notation and the second notation. A voice dictionary creation device comprising: The speech dictionary creation device according to claim 1,
The extraction unit generates the tentative reading corresponding to the first notation among a plurality of the tentative readings, and arranges the words in the first notation with different character types and in the same reading. One of two words is extracted as the character to be removed. In the voice dictionary creation device according to claim 2,
When the registration unit extracts one of the two words as the character to be removed and registers the other notation and reading of the two words in the dictionary, the notation and reading of the two words Is registered in the dictionary. The speech dictionary creation device according to any one of claims 1 to 3,
The extraction unit extracts at least one of a pre-registered removal target word and a pre-registered characteristic pattern character string as the removal target character. Obtain the first notation and the first reading corresponding to the notation and reading of the character string,
Extracting a character to be removed that is estimated to have no portion corresponding to the first reading from the character string in the first notation;
Creating a second representation in which the characters to be removed are removed from the first representation;
Selecting the temporary reading most similar to the first reading as a second reading from a plurality of temporary readings generated based on the reading of the characters in each of the first notation and the second notation;
Registering the notation and the reading of the unknown word included in the first notation in the dictionary based on the notation corresponding to the second reading of the first notation and the second notation and the second reading. A featured voice dictionary creation method. Obtain the first notation and the first reading corresponding to the notation and reading of the character string,
Extracting a character to be removed that is estimated to have no portion corresponding to the first reading from the character string in the first notation;
Creating a second representation in which the characters to be removed are removed from the first representation;
Selecting the temporary reading most similar to the first reading as a second reading from a plurality of temporary readings generated based on the reading of the characters in each of the first notation and the second notation;
A process of registering an unknown word notation and a reading included in the first notation in a dictionary based on a notation corresponding to the second reading and the second reading of the first notation and the second notation. An audio dictionary creation program that is executed by a computer.

Images