JP2015215390A - Speech recognition dictionary update device, speech recognition dictionary update method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech recognition dictionary update device which can suppress selection of unnecessary words by an operator.SOLUTION: A speech recognition dictionary update device comprises: a recognition dictionary storage unit which stores a recognition dictionary; an unregistered word management unit which uses related documents as an input to extract and store an unregistered word appeared in the related documents by a trigger of the word addition instruction of the operator, and not registered in the recognition dictionary; a registration management unit which selects the unregistered words by prescribed criteria to provide the operator with the selected and unregistered words, outputs the unregistered words that are instructed to exclude from registration by the operator as excluded words, and updates an extended recognition dictionary to which the unregistered words other than the excluded words are added as a new recognition dictionary; an excluded word storage unit which stores the excluded words; and an unregistered word correction unit which removes words that matches with the excluded words from among the words newly stored in the unregistered word management unit by a trigger of the word addition instruction by the operator.

Description

  The present invention relates to a speech recognition dictionary updating apparatus, a speech recognition dictionary updating method, and a program for updating a speech recognition dictionary used for speech recognition by adding words.

  Japanese Patent Application Laid-Open No. 2004-151867 discloses a speech recognition dictionary updating apparatus that automatically selects additional words to be registered in a recognition dictionary used in the speech recognition system and adds the selected words to the recognition dictionary. This speech recognition dictionary update device extracts unregistered words that are not registered in the recognition dictionary from related documents input by a speech recognition system operator (hereinafter also referred to simply as an operator) and accumulates them during operation of the speech recognition system. The task relevance and average recognition reliability of each unregistered word are calculated using the voice and the recognition result, and the registration priority of each unregistered word is calculated from the task relevance and average recognition reliability. Select a high unregistered word to add.

  Hereinafter, the speech recognition dictionary updating apparatus of Patent Document 1 will be described with reference to FIGS. 1 and 2. FIG. 1 is a block diagram showing the configuration of the speech recognition dictionary updating device 9 of Patent Document 1. FIG. 2 is a flowchart showing the operation of the speech recognition dictionary update device 9 of Patent Document 1.

  As shown in FIG. 1, the speech recognition dictionary update device 9 of Patent Document 1 includes a related document storage unit 710, an unregistered word extraction unit 110, an unregistered word storage unit 810, and an unregistered word feature amount extraction unit 210. An input voice storage unit 720, a voice recognition unit 120, a recognition result storage unit 820, a recognition result feature quantity extraction unit 220, a task relevance calculation unit 310, a recognition dictionary storage unit 730, and a provisional recognition dictionary registration Unit 130, provisional recognition dictionary storage unit 830, registration priority calculation unit 320, provisional speech recognition unit 140, provisional recognition result storage unit 840, recognition reliability calculation unit 230, recognition dictionary registration unit 330, The configuration includes an additional registered word check exclusion unit 030, a recognition dictionary update unit 040, and an extended recognition dictionary storage unit 900.

  A voice recognition system user (hereinafter also simply referred to as a user) inputs voice to the voice recognition dictionary update device 9. The input voice (hereinafter referred to as input voice) is stored in the input voice storage unit 720, and at the same time, the voice recognition unit 120 recognizes the voice using the recognition dictionary stored in the recognition dictionary storage unit 730, and the recognition result. Is stored in the recognition result storage unit 820.

  The voice recognition system operator inputs a related document to the voice recognition dictionary update device 9. The input related document is stored in the related document storage unit 710.

  In response to a word addition instruction from the voice recognition system operator, a word addition process and a recognition dictionary update process are performed as follows in the procedure shown in FIG.

  (A) The unregistered word extraction unit 110 appears in the related document from the related document stored in the related document storage unit 710 and the recognition dictionary stored in the recognition dictionary storage unit 730, but is registered in the recognition dictionary. Unregistered unregistered words are extracted and stored in the unregistered word storage unit 810 (S110).

  (A) The unregistered word feature extraction unit 210 is a word included in a set of a sentence in a related document including an unregistered word and n sentences (co-occurrence windows) before and after the unregistered word. A co-occurrence frequency vector having a co-occurrence frequency between a plurality of words and the unregistered word as a vector value is generated (S210).

  (C) From the recognition result stored in the recognition result storage unit 820, the recognition result feature amount extraction unit 220 sets the occurrence frequency of words included in the set of utterance windows obtained by dividing the recognition result for every m sentences as a vector value. A word frequency vector is generated (S220).

  (D) The task relevance calculation unit 310 calculates the task relevance of each unregistered word from the co-occurrence frequency vector and word frequency vector of each unregistered word. The task relevance level is calculated from the cosine distance between the co-occurrence frequency vector and the word frequency vector, and represents the height of the semantic relationship between the contents of the input voice of the system user and the unregistered word (S310).

  (E) The provisional recognition dictionary registration unit 130 generates a provisional recognition dictionary in which all the unregistered words stored in the unregistered word storage unit 810 are added to the recognition dictionary stored in the recognition dictionary storage unit 730, and provisional recognition is performed. The data is stored in the dictionary storage unit 830 (S130).

  (F) The provisional speech recognition unit 140 uses the provisional recognition dictionary to generate a provisional recognition result obtained by recognizing the speech stored in the input speech storage unit 720, and stores the provisional recognition result in the provisional recognition result storage unit 840 (S140). .

  (G) The recognition reliability calculation unit 230 uses the provisional recognition result based on the recognition reliability that is an index representing the correctness of the recognition result output of each unregistered word and the number of appearances of the unregistered word. Then, the average recognition reliability of each unregistered word is calculated (S230).

  (H) The registration priority calculation unit 320 calculates the registration priority of each unregistered word using the task relevance and the average recognition reliability of each unregistered word. The registration priority is calculated, for example, as a weighted sum of task relevance and average recognition reliability (S320).

  (K) The recognition dictionary registration unit 330 selects all unregistered words whose registration priority is equal to or higher than a preset threshold value θ (S3301), and selects the selected unregistered words to the additional registered word check exclusion unit 030. Output. Increasing the threshold θ increases the risk that unregistered words to be registered will not be selected, while decreasing the threshold θ increases the risk that unnecessary unregistered words will be selected. A value of about 0.1 is usually used as θ. The additional registered word confirmation exclusion unit 030 presents the list of input unregistered words to the voice recognition system operator (for example, by displaying it on a display) and determines that the voice recognition system operator should not register. An unregistered word is designated using an interface (for example, a check box). The additional registered word confirmation exclusion unit 030 outputs the unregistered unspecified word to the recognition dictionary registration unit 330 (S030). The recognition dictionary registration unit 330 generates an extended recognition dictionary in which the unregistered unregistered words input from the additional registered word check exclusion unit 030 are added to the recognition dictionary stored in the recognition dictionary storage unit 730, and the extended recognition dictionary The data is stored in the storage unit 900 (S3302). The entire procedure including the above steps S3301, S030, and S3302 is referred to as step S330.

  (E) The recognition dictionary update unit 040 updates the contents of the recognition dictionary storage unit 730 with the contents of the extended recognition dictionary (S040).

  The recognition dictionary is updated by the above procedure, and the speech input by the speech recognition system user thereafter is recognized by the speech recognition unit 120 using the updated recognition dictionary. In the updated recognition dictionary, words with high task relevance and high recognition reliability (= highly likely to be uttered in the input speech) are newly registered. Speech recognition can be performed with high accuracy.

  In the above procedure (K) / (S330), when an unregistered word is confirmed / excluded by the voice recognition system operator, an unnecessary word that will not be clearly spoken is selected by mistake. The risk that a simple word is added to the recognition dictionary and appears as a recognition error in the recognition result can be reduced.

In the present invention, the following mathematical formula of Patent Document 1 is used.
Formula (1) “Procedure (b) Generation of co-occurrence frequency vector”
Formula (2) “Procedure (C) Generation of Word Frequency Vector”
Formula (3) "Procedure (ki) Calculation of average recognition reliability"
Formula (4) “Procedure (g) Calculation of recognition reliability for each word used for calculation of average recognition reliability”
Formula (5) "Procedure (ku) Registration priority calculation"

JP 2013-171222 A

  When a speech recognition system using the conventional technology such as Patent Document 1 is operated for a long period of time, the speech recognition system operator does not give a word addition instruction only once. When new related documents are available during system operation, or when the words uttered in the input speech change over time, the speech recognition system operator must give a word addition instruction each time. The recognition dictionary is updated to maintain the speech recognition accuracy.

  In the prior art, when the voice recognition system operator issues a word addition instruction for the second and subsequent times, unnecessary words excluded in the first procedure (K) / (S330) are unregistered words to be added again. In many cases, it will be elected. This is because the word addition process uses a related document and input speech different from the first time, but the algorithm for selecting unregistered words is the same as the first time. For this reason, every time a voice recognition system operator gives an instruction to add a word, an operation to exclude the same unnecessary word occurs every time, and the efficiency of the system operation operation decreases. Since there are unnecessary words that have been newly selected for the second time, the amount of unnecessary word exclusion tends to increase each time the third and fourth times and the word addition execution is repeated. Incurs a decline. Unnecessary words that are left out in this work are registered in the extended recognition dictionary and continue to remain after that, so care must be taken when removing unnecessary words. As described above, the voice recognition system operator has to carefully exclude unnecessary words that are increased every time he / she takes time, and there is a problem that the system operation cost increases.

  Therefore, an object of the present invention is to provide a speech recognition dictionary updating apparatus that can suppress an operator's unnecessary word selection work.

  The speech recognition dictionary update device of the present invention includes a recognition dictionary storage unit, an unregistered word management unit, a registration management unit, an excluded word storage unit, and an unregistered word correction unit.

  The recognition dictionary storage unit stores a recognition dictionary. The unregistered word management unit extracts and stores an unregistered word that is a word that appears in the related document and is not registered in the recognition dictionary with an input of the related document and an operator's word addition instruction. The registration management unit selects unregistered words based on a predetermined criterion, presents the selected unregistered words to the operator, and has received an operator instruction to exclude the selected unregistered words. The recognition dictionary stored in the recognition dictionary storage unit is updated with the extended recognition dictionary in which the unregistered word other than the excluded word is added to the recognition dictionary as a new recognition dictionary. The excluded word storage unit stores excluded words. An unregistered word correction | amendment part deletes the word which corresponds to an exclusion word among unregistered words newly memorize | stored in the unregistered word management part by an operator's word addition instruction | indication.

  According to the speech recognition dictionary updating apparatus of the present invention, it is possible to suppress an operator's unnecessary word selection work.

The block diagram which shows the structure of the speech recognition dictionary update apparatus of patent document 1. FIG. 10 is a flowchart showing the operation of the speech recognition dictionary update device of Patent Document 1. 1 is a block diagram illustrating a configuration of a speech recognition dictionary update device according to Embodiment 1. FIG. 5 is a flowchart showing the operation of the speech recognition dictionary updating apparatus according to the first embodiment. The block diagram which shows the structure of the speech recognition dictionary update apparatus of Example 2. FIG. 9 is a flowchart showing the operation of the speech recognition dictionary updating apparatus according to the second embodiment. FIG. 9 is a block diagram illustrating a configuration of a speech recognition dictionary updating apparatus according to a third embodiment. 10 is a flowchart showing the operation of the speech recognition dictionary updating apparatus according to the third embodiment. FIG. 9 is a block diagram illustrating a configuration of a speech recognition dictionary updating apparatus according to a fourth embodiment. 10 is a flowchart showing the operation of the speech recognition dictionary updating apparatus according to the fourth embodiment. The block diagram which shows the structure of the speech recognition dictionary update apparatus of the modification of Example 1. FIG. 7 is a flowchart showing the operation of the speech recognition dictionary updating apparatus according to a modification of the first embodiment. FIG. 9 is a block diagram illustrating a configuration of a speech recognition dictionary update device according to a modification of the second embodiment. 10 is a flowchart showing the operation of the speech recognition dictionary updating apparatus according to a modification of the second embodiment. FIG. 10 is a block diagram illustrating a configuration of a speech recognition dictionary update device according to a modification of the third embodiment. 10 is a flowchart showing the operation of the speech recognition dictionary updating apparatus according to a modification of the third embodiment. FIG. 10 is a block diagram illustrating a configuration of a speech recognition dictionary update device according to a modification of the fourth embodiment. 10 is a flowchart showing the operation of the speech recognition dictionary updating apparatus according to a modification of the fourth embodiment.

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

  Hereinafter, the speech recognition dictionary updating apparatus 1 according to the first embodiment of the present invention will be described with reference to FIGS. FIG. 3 is a block diagram illustrating the configuration of the speech recognition dictionary updating apparatus 1 according to the present embodiment. FIG. 4 is a flowchart showing the operation of the speech recognition dictionary updating apparatus 1 according to the present embodiment.

  As shown in FIG. 3, the difference from the speech recognition dictionary update device 9 (FIG. 1) of Patent Literature 1 is that the additional registration word confirmation exclusion unit 030 in the speech recognition dictionary update device 9 of Patent Literature 1 stores additional registration word confirmation exclusion. This is a point that is replaced by the part 050, and an excluded word storage part 060 and an unregistered word correction part 070 that are not included in the speech recognition dictionary update device 9 of Patent Document 1 are added.

  Hereinafter, the operations of the additional registered word confirmation exclusion storing unit 050 and the unregistered word correcting unit 070 which are new configurations will be described with reference to FIG. In addition, since the operation | movement of the structure to which the same reference number as FIG. 1 is provided among the structures shown in FIG. 3 is the same as the above-mentioned, description is abbreviate | omitted suitably.

<Additional registered word confirmation exclusion storage unit 050>
Input: Unregistered word selected by recognition dictionary registration unit 330 Output: Unregistered word not specified by speech recognition system operator (to recognition dictionary registration unit 330), Unregistered word specified by speech recognition system operator (To excluded word storage unit 060)
Processing: The additional registered word confirmation exclusion saving unit 050 performs the following processing after the recognition dictionary registration unit 330 selects an unregistered word (S3301).
1) The additional registered word confirmation exclusion storage unit 050 presents the list of input unregistered words to the voice recognition system operator (for example, by displaying it on a display), and the voice recognition system operator should register An unregistered word determined not to be specified is specified using an interface (for example, a check box). The additional registered word confirmation excluding unit 050 outputs the unregistered unspecified word to the recognition dictionary registering unit 330 (S030).
2) The additional registered word confirmation exclusion storing unit 050 further stores the unregistered word (excluded word) designated by the voice recognition system operator in the excluded word storage unit 060 (S050).

<Unregistered word correction unit 070>
Input: Unregistered word stored in unregistered word storage unit 810, Excluded word stored in excluded word storage unit 060 Output: Unregistered word after removal of excluded word (to unregistered word storage unit 810)
Process: The unregistered word correction unit 070 stores the unregistered word extracted by the unregistered word extraction unit 110 in the unregistered word storage unit 810 (S110), and then performs the following process (S070).
1) The unregistered word correction unit 070 acquires an unregistered word stored in the unregistered word storage unit 810 and an excluded word stored in the excluded word storage unit 060.
2) The unregistered word correction unit 070 removes unregistered words that match the excluded words.
3) The unregistered word correction unit 070 rewrites the contents of the unregistered word storage unit 810 with the unregistered words remaining without being removed.
* Since no excluded word is stored in the excluded word storage unit 060 at the time of the first execution, there is no word removed in the above procedure 2), and the output is the same as that of the prior art.

By storing the words excluded by the voice recognition system operator through the above process, and removing the excluded words immediately after the unregistered word extraction process,
* Unnecessary unregistered words excluded in the past will not be selected again * Co-occurrence frequency vector generation processing (S210), task relevance calculation processing (S310), provisional recognition result generation processing (S140), registration priority calculation processing ( The processing time of S320) is shortened. * Accuracy of the average recognition reliability calculated in S230 is improved. As a result, the quality of the extended recognition dictionary is improved, and the effect of improving the speech recognition accuracy is obtained. It is done.

  Hereinafter, the speech recognition dictionary updating apparatus 2 according to the second embodiment of the present invention will be described with reference to FIGS. 5 and 6. FIG. 5 is a block diagram illustrating the configuration of the speech recognition dictionary updating apparatus 2 according to the present embodiment. FIG. 6 is a flowchart showing the operation of the speech recognition dictionary updating apparatus 2 of the present embodiment.

  As shown in FIG. 5, the only difference from the speech recognition dictionary updating apparatus 1 of the first embodiment is that the unregistered word correcting unit 070 in the first embodiment is replaced with an unnecessary word estimation removing unit 080 in the present embodiment. It is. The unregistered word correction unit 070 of the first embodiment only removes unregistered words that match the excluded words that were previously excluded by the speech recognition system operator. In addition, the unnecessary word estimation exclusion unit 080 of the second embodiment is used. Then, unnecessary words estimated from the excluded words (which do not necessarily match the excluded words) are also removed. In the second embodiment, more unnecessary words are removed than in the first embodiment. Therefore, it is possible to reduce the amount of unnecessary words excluded by the voice recognition system operator, reduce the processing time for adding words, and improve the accuracy of the average recognition reliability. The effect of improving the speech recognition accuracy is greater.

  Hereinafter, the operation of the unnecessary word estimation removing unit 080 having a novel configuration will be described with reference to FIG. Note that, in the configuration shown in FIG. 5, the operation of the configuration given the same reference numerals as those in FIG. 3 and FIG.

<Unnecessary word estimation removal unit 080>
Input: Unregistered word stored in unregistered word storage unit 810, Excluded word stored in excluded word storage unit 060 Output: Unregistered word after removal of excluded word (to unregistered word storage unit 810)
Process: The unnecessary word estimation removal unit 080 stores the unregistered word extracted by the unregistered word extraction unit 110 in the unregistered word storage unit 810 (S110), and then performs the following process (S080).
1) The unnecessary word estimation removal unit 080 acquires an unregistered word stored in the unregistered word storage unit 810 and an excluded word stored in the excluded word storage unit 060.
2) The unnecessary word estimation removing unit 080 obtains an estimated excluded word according to the following rule.
(A) Unnecessary word estimation removing unit 080 determines an unregistered word that matches an excluded word as an estimated excluded word.
(B) The unnecessary word estimation removing unit 080 collects symbol excluded words composed only of symbols (characters other than hiragana, katakana, kanji, and alphabets) among the excluded words, and calculates all the symbols that appear in the symbol excluded words. A set (exclusion symbol set) is generated. Among the unregistered words, words that are composed only of symbols included in the excluded symbol set are assumed to be excluded words.
3) The unnecessary word estimation removing unit 080 removes unregistered words that match the estimated excluded words.
4) The unnecessary word estimation removal unit 080 rewrites the contents of the unregistered word storage unit 810 with the unregistered words remaining without being removed.
* Similar to Example 1, since no excluded word is stored in the excluded word storage unit 060 at the first execution, no estimated excluded word is obtained in the above step 2), which is the same as the prior art. Output.

  In this technical field, since a document collected from the Web is input as a related document, the related document often includes an unexpected symbol.

  There are important symbols and unnecessary symbols depending on the application of the speech recognition system. For example, “#” (sharp) and “♭” (flat) are important symbols in a system that recognizes audio related to music, but these symbols are often unnecessary for content other than music. Since it is difficult to distinguish important symbols from unnecessary symbols completely automatically, in Example 2, unnecessary symbols are identified by using words excluded by a voice recognition system operator who is familiar with the usage. . For example, unnecessary symbols are specified in the first sentence of the rule (b) in the procedure 2).

  Even if a symbol is unnecessary, it may be important when the symbol appears as part of a special word (for example, “#” is not necessary in the recognition dictionary, but “C #” in the programming language is desired to be registered. Voice recognition system operator). In the second sentence of the rule (b) in the procedure 2), a word composed only of unnecessary symbols is removed. As a result, when the operator of the speech recognition system excludes “#”, “##” is automatically removed because it is composed only of symbols included in the excluded symbol set when the next word addition is executed. Since “C #” includes a character (alphabet C) that is not included in the exclusion symbol set in the word description, it is passed to the selection process of unregistered words without being removed, and important words are erroneously removed. The situation is prevented.

By adding the rule (b) in step 2), the unnecessary word estimation removal unit 080 removes more words than the unregistered word correction unit 070 of the first embodiment.
* Unnecessary unregistered words excluded in the past will not be selected again, and unnecessary unregistered words consisting only of unnecessary symbols will not be selected, so work efficiency will be greatly improved compared to Example 1. * Co-occurrence The processing time of the frequency vector generation process (S210), the task relevance calculation process (S310), the provisional recognition result generation process (S140), and the registration priority calculation process (S320) is greatly shortened compared to the first embodiment * calculated in S230 The accuracy of the average recognition reliability is greatly improved as compared with the first embodiment, and as a result, the effect of improving the speech recognition accuracy can be obtained.

  Hereinafter, the speech recognition dictionary updating apparatus 3 according to the third embodiment of the present invention will be described with reference to FIGS. FIG. 7 is a block diagram illustrating the configuration of the speech recognition dictionary updating device 3 according to the present embodiment. FIG. 8 is a flowchart showing the operation of the speech recognition dictionary updating apparatus 3 of the present embodiment.

  As shown in FIG. 7, the difference from the speech recognition dictionary update apparatuses 1 and 2 of the first or second embodiment is that the additional registered word confirmation exclusion storing unit 050 in the first and second embodiments is replaced with a registration priority storing unit 090. In other words, an exclusion registration priority storage unit 091 and a registration priority threshold value calculation unit 092 which are not included in the speech recognition dictionary updating apparatuses 1 and 2 of the first and second embodiments are added. In the present embodiment, the above-described changes (changes in which 050 is replaced with 090 and 091 and 092 are added) are added based on the configuration of the second embodiment, but the present invention is not limited to this, and the configuration of the first embodiment Further, a configuration in which similar changes (replace 050 with 090 and changes to add 091 and 092) may be made.

  The recognition dictionary registration unit 330 selects unregistered words having a registration priority equal to or higher than a preset threshold value. In this embodiment, the registration priority (exclusion registration priority) of each unregistered word excluded by the voice recognition system operator is saved, and the threshold value calculated from the exclusion registration priority is registered in the recognition dictionary when the next word addition is executed. By setting in the unit 330, selection of unnecessary unregistered words is suppressed more than in the first or second embodiment.

  The operations of the registration priority storage unit 090 and the registration priority threshold calculation unit 092, which are new configurations, will be described below with reference to FIG. In the configuration shown in FIG. 7, the operation of the configuration given the same reference numerals as those in FIGS. 5, 3, and 1 is the same as described above, and therefore the description thereof is omitted as appropriate.

<Registration priority storage unit 090>
Input: Unregistered word selected by recognition dictionary registration unit 330 Output: Unregistered word not specified by speech recognition system operator (to recognition dictionary registration unit 330), Unregistered word specified by speech recognition system operator (To exclusion word storage unit 060), registration priority of unregistered words designated by the voice recognition system operator (to exclusion registration priority storage unit 091)
Processing: The registration priority storage unit 090 performs the following processing after the recognition dictionary registration unit 330 selects an unregistered word (S3301).
1) The registration priority storage unit 090 presents a list of input unregistered words to the voice recognition system operator (for example, by displaying it on a display), and the voice recognition system operator should not register. The determined unregistered word is designated using an interface (for example, a check box). The registration priority storage unit 090 outputs an unregistered unspecified word to the recognition dictionary registration unit 330 (the same processing as S030 of the conventional technology).
2) The registration priority storage unit 090 further stores the unregistered word (excluded word) designated by the voice recognition system operator in the excluded word storage unit 060 and sets the registration priority of the excluded word as the excluded registration priority. The data is stored in the storage unit 091 (S090). At this time, the exclusion registration priority stored at the time of past word addition is deleted, and then the current exclusion registration priority is stored. There are a plurality of exclusion registration priorities (the same number as the number of unregistered words excluded by the voice recognition system operator), and N_all indicates the number.

<Registration priority threshold calculation unit 092>
Input: Exclusion registration priority stored in exclusion registration priority storage unit 091 Output: Registration priority threshold (to recognition dictionary registration unit 330)
Process: The registration priority threshold calculation unit 092 performs the following process after the registration priority calculation unit 320 calculates the registration priority of each unregistered word (S320).
1) The registration priority threshold value calculation unit 092 acquires the registration priority threshold value θ set in the recognition dictionary registration unit 330.
2) The registration priority threshold value calculation unit 092 acquires all the exclusion registration priorities stored in the exclusion registration priority storage unit 091.
3) When the number of excluded registration priorities (N_all) is smaller than a predetermined value N, the registration priority threshold value calculation unit 092 sets the registration priority threshold value θ acquired in step 1 in the recognition dictionary registration unit 330 as it is. To finish the process. N represents the number of exclusion registration priorities necessary for determining the registration priority threshold change, and a value of about 50 is used.
4) The registration priority threshold calculation unit 092 calculates θ + Δθ by adding the registration priority threshold maximum increase width Δθ to the registration priority threshold θ when the number of exclusion registration priorities is N or more, and excludes it. The number (N_small) of low registration priorities whose values are equal to or less than θ + Δθ among the registration priorities is counted. Increasing Δθ increases the amount of increase in the registration priority threshold and makes it difficult to select unregistered words, while decreasing it decreases the amount of increase in the registration priority threshold and makes it difficult to suppress selection of unnecessary words. A value of about 0.05 to 0.1 is usually used for Δθ.
5) The registration priority threshold calculation unit 092 determines that the ratio (N_small / N_all) of the number of low registration priorities (N_small) out of the total number of excluded registration priorities (N_all) is equal to or greater than a predetermined value R The maximum registration priority among the low registration priorities is set in the recognition dictionary registration unit 330 as the update value of the registration priority threshold value θ. If N_small / N_all is less than R, the registration priority threshold value θ acquired in step 1 is set in the recognition dictionary registration unit 330 as it is. R is a parameter that indicates how dense the exclusion registration priority is in the vicinity of the threshold value, and the threshold value increase is determined. Usually, a value of about 0.8 is used.
* At the first execution, since no exclusion registration priority is stored in the exclusion registration priority storage unit 091, N_all does not exceed N in the above procedure 3, and the registration priority threshold is not changed. .

  The recognition dictionary registration unit 330 selects unregistered words whose registration priority is equal to or higher than the threshold θ. At this time, an unregistered word whose registration priority slightly exceeds the threshold value θ has a lower registration priority among the selected unregistered words, and has a relatively high possibility of being an unnecessary word.

  When the speech recognition system operator excludes unnecessary words, if many unregistered words slightly exceeding the threshold θ are excluded, it is considered that the threshold should be increased. Example 3 is based on this idea, and uses the registration priority (exclusion registration priority) of unnecessary words excluded by the voice recognition system operator so that unnecessary words are not easily selected at the next and subsequent word addition executions. It is determined whether or not the threshold value θ is increased.

By the above steps 3 to 5, the registration priority threshold is set as follows.
* When the voice recognition system operator excludes a certain number of unregistered words,
* If many unregistered words (ratio R or more) that slightly exceed the registration priority threshold (by Δθ) are excluded,
* When the next word addition is executed, the registration priority threshold is reset so that all unregistered words that slightly exceed the registration priority threshold θ are excluded (to the maximum value of the low registration priority).

  In the third embodiment, as described above, when the speech recognition system operator excludes many unregistered words near the threshold at the previous execution, the threshold is changed so that they are not selected again. Thereby, in addition to the effect of Example 1 and 2, selection of an unnecessary word can be suppressed and the unnecessary word exclusion operation | work of a speech recognition system operator can further be reduced.

  Hereinafter, the speech recognition dictionary updating apparatus 4 according to the fourth embodiment of the present invention will be described with reference to FIGS. FIG. 9 is a block diagram showing the configuration of the speech recognition dictionary update device 4 of this embodiment. FIG. 10 is a flowchart showing the operation of the speech recognition dictionary update device 4 of this embodiment.

  As shown in FIG. 9, the difference from the speech recognition dictionary updating apparatus 1 of the first embodiment is that the unregistered word correction unit 070 in the first embodiment is replaced with a similar unregistered word exclusion unit 095.

  There is a high possibility that a word similar to an unnecessary word excluded by the voice recognition system operator in the past is unnecessary (for example, when “PC-9801” is unnecessary, “PC-8801” or “PC-”). 9801F "is also considered unnecessary). Based on this idea, the speech recognition dictionary update device 4 according to the fourth embodiment suppresses selection of unnecessary words more than the speech recognition dictionary update device 1 according to the first embodiment.

  The unregistered word correction unit 070 of the first embodiment removes only words that completely match words (excluded words) that were previously excluded by the speech recognition system operator from the unregistered words. In addition to that, the unregistered word exclusion unit 095 also removes words similar to the excluded words (words having an editing distance close to the excluded words) from the unregistered words. Thereby, selection of an unnecessary unregistered word is suppressed more than in the first embodiment.

  Hereinafter, the operation of the similar unregistered word exclusion unit 095 will be described with reference to FIG. Note that, in the configuration shown in FIG. 9, the operation of the configuration given the same reference numerals as in FIG. 3 is the same as that in the first embodiment described above, and therefore the description thereof is omitted as appropriate.

<Similar Unregistered Word Exclusion Unit 095>
Input: Unregistered word stored in the unregistered word storage unit 810, Excluded word stored in the excluded word storage unit 060: Unregistered word (unregistered word removed from the same word as the excluded word and similar words) Registered word storage unit 810)
Process: The similar unregistered word exclusion unit 095 stores the unregistered word extracted by the unregistered word extraction unit 110 in the unregistered word storage unit 810 (S110), and then performs the following process (S095).
1) The similar unregistered word exclusion unit 095 acquires an unregistered word stored in the unregistered word storage unit 810 and an excluded word stored in the excluded word storage unit 060.
2) The similar unregistered word exclusion unit 095 performs the following process for each unregistered word.
2-1) The similar unregistered word exclusion unit 095 calculates an edit distance between the unregistered word and each excluded word, and calculates a normalized edit distance obtained by dividing each edit distance by the number of characters of the unregistered word. calculate. The edit distance is calculated by an existing method disclosed in Reference Non-Patent Document 1, for example. The normalized edit distance is a real value indicating how much of the character string constituting the unregistered word is different from the excluded word, and when the unregistered word completely matches the excluded word 0, and the value increases as the number of different parts increases (for example, if there are 4 unregistered words and 4 excluded words and there is no matching character, edit distance = 4, normalized edit Distance = 4 ÷ 4 = 1.0).
2-2) The similar unregistered word exclusion unit 095 removes the unregistered word when at least one of the normalized edit distances is equal to or less than a predetermined normalized edit distance threshold τ. When τ = 0 is set, only unregistered words that completely match the excluded words are removed, so the output is the same as in the first embodiment. Usually, a value of about τ = 0.2 is set. When τ = 0.2 is set, for example, when 4 or more characters out of 5 unregistered words match the excluded word, they are removed.
3) The similar unregistered word exclusion unit 095 rewrites the contents of the unregistered word storage unit 810 with the remaining unregistered words that have not been removed.
* Since no excluded word is stored in the excluded word storage unit 060 at the time of the first execution, there is no word removed in the above procedure 3, and the output is the same as that of the conventional technique.

The words (excluded words) excluded by the voice recognition system operator by the above process are stored, and words similar to the excluded words are removed immediately after the unregistered word extraction process.
* Unnecessary unregistered words excluded in the past are not selected again, and unnecessary unregistered words similar to unregistered words excluded in the past are also not selected, so work efficiency is greatly improved compared to the first embodiment. * Co-occurrence frequency vector generation processing (S210), task relevance calculation processing (S310), provisional recognition result generation processing (S140), registration priority calculation processing (S320) processing time is greatly shortened compared to the first embodiment. * The accuracy of the average recognition reliability calculated in S230 is greatly improved as compared with the first embodiment, and as a result, the effect of improving the speech recognition accuracy can be obtained.
(Reference Non-Patent Document 1: Daniel Jurafsky and James H. Martin: “Speech and Language Processing: An introduction to natural language processing, computational linguistics, and speech recognition”, Prentice Hall, 2009)

<Effect of the invention>
According to the present invention, selection of unnecessary words at the second and subsequent word addition executions is suppressed, so that the number of words that should be excluded by the voice recognition system operator is reduced and the efficiency of operation work is improved. As a result, the recognition dictionary maintenance cost during operation of the speech recognition system is reduced.

  Furthermore, in the present invention, unnecessary words are excluded before the selection process of unregistered words (immediately after the prior art procedure (A) and immediately before the procedure (A)). The time required for each of the processes (procedure (b) co-occurrence frequency vector generation process, procedure (d) task relevance calculation process, and procedure (c) registration priority calculation process) is reduced. In addition, since unnecessary words are also excluded from the temporary recognition dictionary, the size of the temporary recognition dictionary is reduced, and the time taken for the temporary recognition result generation processing (voice recognition processing by the temporary recognition dictionary) of the procedure (f) is shortened. The Due to these processing time reduction effects, the time for the voice recognition system operator to wait until the word addition processing is completed after issuing the word addition execution instruction is reduced, and the work efficiency is further improved at the second and subsequent word addition.

  Furthermore, when unnecessary words are reduced from the provisional recognition dictionary, the accuracy of the average recognition reliability calculated in the procedure (g) is improved. As described in paragraph [0037] of Patent Document 1, the recognition reliability is calculated. The paragraph [0037] of Patent Document 1 is cited below.

<Patent Document 1 Paragraph [0037]>
Any known method can be used to calculate the recognition reliability. For example, it can be calculated as follows (for more details, see “Li, Kawahara, Shikano,“ Fast word posterior in the two-pass search algorithm. Probability calculation method based on probability ", Information Processing Society of Japan, Vol.2003, No.124, pp.281-286)). The speech recognition system outputs a word string having the highest likelihood of the acoustic model and the language model for the input speech as a recognition result. However, if the pronunciation of the input speech is ambiguous, or if there are multiple words that are not unnaturally connected, a large number of conflicting candidates with the same likelihood appear, and the speech recognition system It can be difficult to identify. If there are many influential conflict candidates when the speech recognition system determines the output, the recognition reliability is low. Conversely, if there are few conflict candidates, the recognition reliability is high.

<End of paragraph [0037] of Patent Document 1>
If there are many unnecessary words in the provisional recognition dictionary, there is a high possibility that these words will become the above-mentioned “potential conflict candidates” during the provisional speech recognition processing, and words that should originally be given a high average recognition reliability. The number of cases where the average recognition reliability of the is reduced. Since the accuracy of recognition reliability is improved by reducing unnecessary words from the provisional recognition dictionary in advance, the accuracy of registration priority calculated using the average recognition reliability is also improved. Unregistered words to be added are selected more accurately. The number of cases where words that should originally be added are not selected and cases where words that should not be added are reduced is reduced, so the quality of the extended recognition dictionary that is finally generated is improved and the speech recognition accuracy is increased after the recognition dictionary is updated. The improvement effect is obtained.

<Points of the present invention>
The present invention prevents an unnecessary word from being selected during the second and subsequent word addition execution based on the idea that “the word excluded by the voice recognition system operator is definitely an unnecessary word”. Specifically, unnecessary words excluded by the speech recognition system operator in the prior art procedure (K) are stored, and prior art unregistered word extraction processing (procedure) (A) Unnecessary words are excluded at the timing immediately after S110).

  By performing unnecessary word exclusion processing at this timing, it is possible to reduce the time required for various subsequent processing (the second paragraph of <Effects of the invention>) and the accuracy of selecting unregistered words to be added. The effect to improve (the third paragraph of <Effect of invention>) is obtained.

[Modification 1]
Hereinafter, with reference to FIG. 11 and FIG. 12, a modification of the first embodiment (modification 1) realized by consolidating some of the constituent elements of the prior art to focus on the main part of the present invention. The speech recognition dictionary update device 5 will be described. FIG. 11 is a block diagram showing the configuration of the speech recognition dictionary update device 5 of this modification. FIG. 12 is a flowchart showing the operation of the speech recognition dictionary update device 5 of this modification.

  As shown in FIG. 11, the speech recognition dictionary update device 5 according to the present modification includes all the constituent elements included in the first embodiment without excess or deficiency. However, the related document storage unit 710, the unregistered word extraction unit 110, and the unregistered word storage unit 810 are included in the unregistered word management unit 5100. The unregistered word feature quantity extraction unit 210, the input voice storage unit 720, the voice recognition unit 120, the recognition result storage unit 820, the recognition result feature quantity extraction unit 220, and the task relevance calculation unit 310 This is a configuration included in the association degree management unit 5210. In addition, the provisional recognition dictionary registration unit 130, the provisional recognition dictionary storage unit 830, the provisional speech recognition unit 140, the provisional recognition result storage unit 840, and the recognition reliability calculation unit 230 are included in the average recognition reliability management unit 5220. It is the composition which is. The speech recognition dictionary updating apparatus 5 includes a registration priority management unit 5200 that further aggregates three component requirements of a task relevance management unit 5210, an average recognition reliability management unit 5220, and a registration priority calculation unit 320. Further, the recognition dictionary registration unit 330, the additional registered word confirmation exclusion storage unit 050, the recognition dictionary update unit 040, and the extended recognition dictionary storage unit 900 are included in the registration management unit 5300. The speech recognition dictionary update device 5 of the present modification includes a recognition dictionary storage unit 730, an excluded word storage unit 060, and an unregistered word correction unit 070 in addition to the above-described new configuration requirements. The speech recognition dictionary update device 5 of the present modification performs the same operation as that of the first embodiment, but the respective component requirements existing in the first embodiment are aggregated in new component requirements including these components. An example of the operation of a new configuration requirement obtained by integrating the requirements is disclosed in different expressions.

  As described above, the recognition dictionary storage unit 730 stores a recognition dictionary. The unregistered word management unit 5100 extracts and stores an unregistered word that is a word that appears in the related document and is not registered in the recognition dictionary when the related document is input and the operator adds a word. (S5100).

  The registration priority management unit 5200 calculates the registration priority of each unregistered word (S5200). More specifically, the task relevance management unit 5210 takes the input voice of the user as an input, and indicates the degree of semantic relevance between the contents of the user input voice and each unregistered word. Is calculated for each unregistered word (S5210). The average recognition reliability management unit 5220 generates a temporary recognition result that is a result of voice recognition of the user's input voice using a temporary recognition dictionary in which an unregistered word is added to the recognition dictionary, and based on the temporary recognition result An average recognition reliability that is an average of the number of appearances of an index representing the correctness of the recognition result output of each unregistered word is calculated (S5220). Similar to the first embodiment, the registration priority calculation unit 320 calculates the registration priority of each unregistered word based on the task association degree and the average recognition reliability of each unregistered word.

  The registration management unit 5300 selects an unregistered word based on a predetermined criterion (for example, the above-described registration priority and its threshold value), presents the selected unregistered word to the operator, and among the selected unregistered words An unregistered word for which an operator instruction to exclude is output as an excluded word, and an extended recognition dictionary in which unregistered words other than the excluded word are added to the recognition dictionary is stored in the recognition dictionary storage unit 730 as a new recognition dictionary. The recognition dictionary is updated (S5300). The excluded word storage unit 060 stores excluded words (S060). The unregistered word correction unit 070 deletes a word that matches the excluded word from unregistered words newly stored in the unregistered word management unit 5100, triggered by the operator's word addition instruction (S070).

  As described above, according to the speech recognition dictionary update device 5 of the present modification, it is possible to suppress unnecessary word selection work by the operator, as with the speech recognition dictionary update device 1 of the first embodiment.

[Modification 2]
Hereinafter, with reference to FIG. 13 and FIG. 14, a modification of the second embodiment (modification 2) realized by consolidating some of the constituent elements of the prior art in order to focus on the main part of the present invention. The speech recognition dictionary update device 6 will be described. FIG. 13 is a block diagram showing the configuration of the speech recognition dictionary update device 6 of this modification. FIG. 14 is a flowchart showing the operation of the speech recognition dictionary update device 6 of this modification.

  As shown in FIG. 13, the speech recognition dictionary updating apparatus 6 of this modification includes all the constituent elements included in the second embodiment without excess or deficiency. However, the speech recognition dictionary update device 6 according to the present modification, like the first modification, is a new configuration requirement in which the configuration requirements existing in the second embodiment are aggregated, the unregistered word management unit 5100, the registration priority management. Part 5200 and registration management part 5300. Since these configurations are the same as those of the first modification, description thereof is omitted. The speech recognition dictionary updating apparatus 6 according to the present modification includes a recognition dictionary storage unit 730, an excluded word storage unit 060, and an unnecessary word estimation removal unit 080 in addition to the above-described new components.

  Steps S5100, S5200, S5300, and S060 are executed in the same manner as in the first modification. Unnecessary word estimation removal unit 080 takes a word that matches an excluded word and a word similar to the excluded word among unregistered words newly stored in the unregistered word management unit, triggered by the operator's word addition instruction. Delete (S080). As described above, for example, the unnecessary word estimation removal unit 080 generates a set of symbol excluded words, which are words composed only of symbols among the excluded words, as an excluded symbol set and newly adds them to the unregistered word management unit 5100. Of the stored unregistered words, unregistered words composed only of symbols included in the excluded symbol set are deleted (S080).

  Thus, according to the speech recognition dictionary update device 6 of this modification, the same effect as the speech recognition dictionary update device 2 of Example 2 can be obtained.

[Modification 3]
Hereinafter, with reference to FIG. 15 and FIG. 16, it is the modification (modification 3) of Example 3 implement | achieved by integrating some components of a prior art in order to pay attention to the principal part of this invention. The speech recognition dictionary update device 7 will be described. FIG. 15 is a block diagram showing the configuration of the speech recognition dictionary update device 7 of this modification. FIG. 16 is a flowchart showing the operation of the speech recognition dictionary update device 7 of this modification.

  As shown in FIG. 15, the speech recognition dictionary updating apparatus 7 of this modification includes all the constituent elements included in the third embodiment without excess or deficiency. However, the speech recognition dictionary update device 7 of the present modified example is a new constituent requirement in which constituent features existing in the third embodiment are aggregated, as in the first and second modified examples, the unregistered word management unit 5100, and the registration priority. A degree management unit 5200 is included. Since these configurations are the same as those of the first and second modifications, description thereof is omitted. In addition to the above-described configuration requirements, the speech recognition dictionary update device 7 of the present modification includes a recognition dictionary registration unit 330, a recognition dictionary update unit 040, an extended recognition dictionary storage unit 900, a registration priority storage unit 090, and exclusion registration priority. A registration management unit 7300 that includes a degree storage unit 091 and a registration priority threshold value calculation unit 092. Furthermore, the speech recognition dictionary update device 5 of the present modification includes a recognition dictionary storage unit 730, an excluded word storage unit 060, and an unnecessary word estimation removal unit 080.

  Steps S5100, S5200, S060, and S080 are executed in the same manner as in the second modification. The registration management unit 7300 sets and updates a registration priority threshold based on the registration priority of the output excluded word, and selects the unregistered word using the registration priority threshold as the predetermined reference (S7300).

  Thus, according to the speech recognition dictionary update device 7 of this modification, the same effect as the speech recognition dictionary update device 3 of Example 3 can be obtained.

[Modification 4]
Hereinafter, with reference to FIG. 17 and FIG. 18, a modified example (modified example 4) of the fourth embodiment realized by consolidating some of the constituent elements of the prior art in order to pay attention to the main part of the present invention. The speech recognition dictionary update device 8 will be described. FIG. 17 is a block diagram showing a configuration of the speech recognition dictionary update device 8 of the present modification. FIG. 18 is a flowchart showing the operation of the speech recognition dictionary update device 8 of the present modification.

  As shown in FIG. 17, the speech recognition dictionary update device 8 of the present modification includes all the constituent requirements included in the fourth embodiment without excess or deficiency. The voice recognition dictionary update device 8 of the present modification is almost the same as the configuration of the voice recognition dictionary update device 5 of the first modification, but the unregistered word correction unit 070 in the first modification is similar unregistered in this modification. The only difference is that the word exclusion unit 095 is changed.

  Steps S5100, S5200, S5300, and S060 are executed in the same manner as in the first modification. Next, the similar unregistered word exclusion unit 095 is a part that is different from the excluded word in the character string that forms the unregistered word newly stored in the unregistered word management unit, triggered by the operator's word addition instruction. The normalized edit distance, which is a value representing the degree of, is calculated, and the newly stored unregistered word is deleted based on the normalized edit distance (S095).

  Thus, according to the speech recognition dictionary update device 8 of this modification, the same effect as the speech recognition dictionary update device 4 of Example 4 can be obtained.

  The various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Needless to say, other modifications are possible without departing from the spirit of the present invention.

  Further, when the above-described configuration is realized by a computer, processing contents of functions that each device should have are described by a program. The processing functions are realized on the computer by executing the program on the computer.

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

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

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

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

Claims (8)

A recognition dictionary storage unit for storing the recognition dictionary;
An unregistered word management unit that extracts and stores an unregistered word that is a word that appears in the related document and is not registered in the recognition dictionary, with an input of the related document and triggered by an operator's word addition instruction; ,
The unregistered word is selected based on a predetermined criterion, the selected unregistered word is presented to the operator, and an operator instruction is given to exclude the selected unregistered word. A registration management unit that updates the recognition dictionary stored in the recognition dictionary storage unit as a new recognition dictionary with an extended recognition dictionary added to the recognition dictionary as an unregistered word other than the excluded word,
An excluded word storage unit for storing the excluded word;
Triggered by the operator's word addition instruction, among unregistered words newly stored in the unregistered word management unit, an unregistered word correction unit that deletes a word that matches the excluded word;
A speech recognition dictionary update device including: A recognition dictionary storage unit for storing the recognition dictionary;
An unregistered word management unit that extracts and stores an unregistered word that is a word that appears in the related document and is not registered in the recognition dictionary, with an input of the related document and triggered by an operator's word addition instruction; ,
The unregistered word is selected based on a predetermined criterion, the selected unregistered word is presented to the operator, and an operator instruction is given to exclude the selected unregistered word. A registration management unit that updates the recognition dictionary stored in the recognition dictionary storage unit as a new recognition dictionary with an extended recognition dictionary added to the recognition dictionary as an unregistered word other than the excluded word,
An excluded word storage unit for storing the excluded word;
An unnecessary word that deletes a word that matches the excluded word and a word that is similar to the excluded word among unregistered words newly stored in the unregistered word management unit, triggered by the operator's word addition instruction An estimation removal unit;
A speech recognition dictionary update device including: The speech recognition dictionary update device according to claim 2,
The unnecessary word estimation removing unit
A set of symbol excluded words that are words composed only of symbols among the excluded words is generated as an excluded symbol set, and the excluded symbol set among the unregistered words newly stored in the unregistered word management unit The speech recognition dictionary update apparatus which deletes the unregistered word comprised only by the symbol contained in. A recognition dictionary storage unit for storing the recognition dictionary;
An unregistered word management unit that extracts and stores an unregistered word that is a word that appears in the related document and is not registered in the recognition dictionary, with an input of the related document and triggered by an operator's word addition instruction; ,
The unregistered word is selected based on a predetermined criterion, the selected unregistered word is presented to the operator, and an operator instruction is given to exclude the selected unregistered word. A registration management unit that updates the recognition dictionary stored in the recognition dictionary storage unit as a new recognition dictionary with an extended recognition dictionary added to the recognition dictionary as an unregistered word other than the excluded word,
An excluded word storage unit for storing the excluded word;
Normalization that is a value representing the degree of a portion that is different from the excluded word in the character string constituting the unregistered word newly stored in the unregistered word management unit, triggered by the operator's word addition instruction A similar unregistered word excluding unit that calculates an edit distance and deletes the newly stored unregistered word based on the normalized edit distance;
A speech recognition dictionary update device including: The speech recognition dictionary update device according to any one of claims 1 to 4,
The average recognition reliability is assumed to be the average number of appearances of an index representing the correctness of the recognition result output of each unregistered word,
The task relevance level represents the height of the semantic relevance between the content of the user's input speech and each of the unregistered words,
The registration priority is an index based on the task relevance and the average recognition reliability,
The registration management unit
A speech recognition dictionary updating apparatus that sets and updates a registration priority threshold based on the registration priority of the output excluded word, and selects the unregistered word using the registration priority threshold as the predetermined reference. A speech recognition dictionary update method executed by a speech recognition dictionary update device including a recognition dictionary storage unit that stores a recognition dictionary and an excluded word storage unit that stores an excluded word,
An unregistered word management step for extracting and storing an unregistered word that appears in the related document and is not registered in the recognition dictionary, with an input of the related document and triggered by an operator's word addition instruction; ,
The unregistered word is selected based on a predetermined criterion, the selected unregistered word is presented to the operator, and an operator instruction is given to exclude the selected unregistered word. A registration management step of updating the recognition dictionary stored in the recognition dictionary storage unit as a new recognition dictionary using an extended recognition dictionary in which unregistered words other than the exclusion word are added to the recognition dictionary as a new recognition dictionary; ,
Triggered by the operator's word addition instruction, an unregistered word correction step of deleting a word that matches the excluded word among the unregistered words newly stored in the unregistered word management step;
A speech recognition dictionary update method including: A speech recognition dictionary update method executed by a speech recognition dictionary update device including a recognition dictionary storage unit that stores a recognition dictionary and an excluded word storage unit that stores an excluded word,
An unregistered word management step for extracting and storing an unregistered word that appears in the related document and is not registered in the recognition dictionary, with an input of the related document and triggered by an operator's word addition instruction; ,
The unregistered word is selected based on a predetermined criterion, the selected unregistered word is presented to the operator, and an operator instruction is given to exclude the selected unregistered word. A registration management step of updating the recognition dictionary stored in the recognition dictionary storage unit as a new recognition dictionary using an extended recognition dictionary in which unregistered words other than the exclusion word are added to the recognition dictionary as a new recognition dictionary; ,
An unnecessary word that deletes a word that matches the excluded word and a word that is similar to the excluded word among unregistered words newly stored in the unregistered word management step, triggered by the operator's word addition instruction An estimated removal step;
A speech recognition dictionary update method including:   A program for causing a computer to function as the speech recognition dictionary updating apparatus according to any one of claims 1 to 5.

Images