JP2009271117A - Voice retrieval device and voice retrieval method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice retrieval device and a voice retrieval method, wherein voice recognition accuracy is improved, and a recognition result that is easily understood by a user is presented. <P>SOLUTION: A voice recognition section 4 performs voice recognition on an input voice, by referring to a sound reference pattern data base 2 and a word dictionary 3 for recognition. A data base retrieval section 9 obtains a retrieval result corresponding to the recognition result by referring to a data base 8 for retrieval, and stores it with a retrieval score which indicates a similarity degree with the recognition result, in a retrieval result data storage section 10. A recognition result correction section 11 collates a word included in the retrieval result to a node network through the recognition result, and performs correction by replacing it with a similar word, and performs ordering of the retrieval result by correcting the retrieval score based on the corrected recognition result. A candidate presenting section 12 presents the retrieval result of a retrieval score order, and the corrected recognition result corresponding to each retrieval result. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a voice search device and a voice search method for presenting a search result including ambiguity searched based on a result of voice recognition.

  Voice input has an advantage that even a beginner can input faster than input using a keyboard or a touch panel, and input can be executed even when eyes or hands are blocked by other tasks. In recent years, full-text search and name search using speech have been studied by combining large vocabulary continuous speech recognition and database search. At this time, it is necessary to consider that there are restrictions on vocabulary that can be recognized in speech recognition, and that speech recognition results include recognition errors.

  As a specific measure for this, for example, in Patent Document 1, in order to perform speech recognition in consideration of an unknown word that is not registered in the speech recognition dictionary, subwords (syllables) that are shorter than words and represented by a smaller number of types are used. ) And a word model used together, a search device that performs speech recognition is disclosed. This search device first searches for a search target document using only a part recognized as a word in a user utterance, acquires a search result, and then is an unknown part that is not recognized as a word in a user utterance. In order to interpolate words, keywords that match or are similar to the subword chain (syllable string) are determined from the keywords in the search results. Then, the search accuracy is improved by searching the search target document again using the keyword in which the unknown word is interpolated.

JP 2003-271629 A

Since the conventional voice search apparatus and voice search method are configured as described above, it is possible to improve the accuracy of voice document search in consideration of vocabulary restrictions and ambiguity of voice recognition due to misrecognition. ing. In the case of the search device disclosed in Patent Document 1, speech recognition is performed using limited words and syllables, a document is searched using the recognized words, syllable strings included in the search results, and keywords in the search results Match with. If the user utterance is a document, there is a high possibility that a document search result with a certain degree of accuracy can be obtained by searching a search target document using a keyword.
However, when the user utterance is a facility name such as a large-scale facility name, the speech recognition may fail due to the fact that the keyword is not included in the search result or the keyword is erroneously recognized as a similar keyword. There was a problem that the facility name of could not be searched.

  The present invention has been made to solve the above-described problems, and performs a search in consideration of ambiguity due to voice recognition, and verifies the validity of the voice recognition result by using the information of the search result. The purpose of the correction is to improve the recognition accuracy and present a recognition result that is easy for the user to understand when presenting the recognition result to the user.

  The speech search device according to the present invention uses a speech recognition unit that outputs a word string corresponding to an input speech as a recognition result, and divides the recognition result into subwords in units smaller than words, and uses them as search keys. A search unit that searches a search target document including a delimited index and outputs a search result, a recognition result correction unit that corrects a recognition result based on the search result, and a recognition result corrected by the search result and the recognition result correction unit And a candidate presenting unit that presents at least one of them.

  According to the present invention, a search target document including an index divided into subwords by using a word string corresponding to the input speech as a recognition result, dividing the recognition result into subwords smaller than words and using them as search keys. Is used as a search result, the recognition result is corrected based on the search result, and at least one of the search result and the corrected recognition result is presented. Recognition results that are easy for a person to understand can be presented.

Embodiment 1 FIG.
FIG. 1 is a block diagram showing a configuration of a speech search apparatus according to Embodiment 1 of the present invention. A speech search apparatus 1 shown in FIG. 1 includes an acoustic standard pattern database 2 that stores acoustic feature quantities for each minimum unit of speech used for speech recognition, a recognition word dictionary 3 that includes a word definition dictionary and a word connection dictionary, and input speech. On the other hand, a speech recognition unit 4 that performs speech recognition including speech analysis processing, collation processing, and search processing, a database search unit 9 that refers to the search database 8 using the recognition results, and obtains the search results. A search result data storage unit 10 to be stored, a recognition result correction unit 11 for correcting the recognition result using the search result of the search result data storage unit 10, and a candidate presentation unit 12 for presenting the search result and the corrected recognition result are provided.

  This speech search apparatus 1 relates to a search apparatus that performs a search using input objects as search targets that are generally allowed to be omitted, although they are short word strings as in a large-scale facility name search. The database search unit 9 searches the search database 8 using a set of subwords such as phonemes, which are units smaller than words, as search keys. Unlike the conventional speech search device, the speech recognition apparatus 1 includes a database search unit 9 that uses a subword unit delimiter, a recognition result correction unit 11, and a candidate presentation unit 12. In the database search unit 9, a set of subwords smaller than words. Is used as a search key to search the search database 8, and the recognition result correction unit 11 compares the obtained search result with the recognition result of speech recognition, and ranks the search results based on the corrected correction recognition result candidates. The candidate is corrected and presented to the user in the candidate presentation unit 12.

  The acoustic standard pattern database 2 shown in FIG. 1 stores an acoustic pattern composed of a spectrum and temporal characteristics for each minimum unit of speech recognition such as phonemes. FIG. 2 is an explanatory diagram showing the structure of a hidden Markov model (HMM) used in the speech search apparatus according to Embodiment 1 of the present invention, and shows an example of the topology of the hidden Markov model. As an acoustic standard pattern, for example, as shown in FIG. 2, a hidden Markov model in which three states in units of phonemes are arranged in time series is used. Each state 21 has an autoregressive arc 22 and no back arc. The acoustic feature amount corresponding to each state is represented by an 8-mixed Gaussian distribution that maintains covariance. The parameters of the acoustic standard pattern are estimated in advance from learning voice data of a large number of speakers.

The recognition word dictionary 3 is a dictionary used to describe the input speech to be recognized as a combination of acoustic standard patterns. Specifically, it consists of a word definition dictionary describing linguistic output targets and constraints to be recognized in an acoustic standard pattern, and a word connection dictionary describing connections between words.
When searching for a large-scale facility name, it is not realistic from the viewpoint of memory and processing speed that the voice search device 1 registers and processes all words in the recognition word dictionary 3. For this reason, the speech search apparatus 1 uses a recognition word dictionary 3 in which typical words and other morphemes are expressed as subwords. FIG. 3 is an explanatory diagram showing an example of a recognition word dictionary of the speech search apparatus according to Embodiment 1 of the present invention. In the example of the subword and word definition dictionary shown in FIG. 3, a word or subword and a chain of acoustic standard patterns are paired.
FIG. 4 is an explanatory diagram showing an example of a recognition word dictionary of the speech search apparatus according to Embodiment 1 of the present invention. In the example of the word connection dictionary shown in FIG. 4, a combination of a chain (trigram) of three words or subwords and a probability of chaining are paired.

  The speech recognition unit 4 shown in FIG. 1 includes a speech analysis unit 5 that performs speech analysis processing, a verification unit 6 that performs verification processing, and a search unit 7 that performs search processing. The recognition word dictionary 3 is collated, and a combination of words having a high score representing the degree of collation is output as a recognition result.

  The voice analysis unit 5 converts the input voice into an acoustic feature value suitable for voice recognition. As an acoustic feature amount calculation method, for example, the speech analysis unit 5 performs A / D conversion on input speech at a sampling period of 16 kHz and 16 bits, performs 256-point Fourier transform at intervals of 10 ms in a time frame, and obtains a power spectrum. Inverse Fourier transform is performed after logarithmizing the frequency axis. A total of 24 dimensions including the 12-dimensional mel cepstrum thus calculated and the 12-dimensional primary regression coefficient in the time direction is used as the acoustic feature amount.

The collation unit 6 collates the acoustic feature amount calculated by the voice analysis unit 5 with the acoustic standard pattern stored in the acoustic standard pattern database 2, and calculates a score representing the degree of collation.
The search unit 7 refers to the correspondence between the words or subwords in the recognition word dictionary 3 and the acoustic standard patterns, and combinations of the words or subwords, and the cumulative score for the entire speech increases based on the word connection relationship of the input speech. Search for recognition candidates and output a word string as a recognition result. In addition, the search part 7 may output a high-order multiple candidate with a high score as a recognition result, or may output a recognition result with a score.

  The detailed method of speech recognition by the speech recognition unit 4 is explained in “Basics of Speech Recognition (above) (below), co-authored by Lawrence Labiner, Biing-Hwang Jung, Translated by Sadahiro Furui, NTT Advanced Technology Co., Ltd.” Has been.

The search database 8 stores a search target word string such as a facility name. FIG. 5 is an explanatory diagram showing an example of a search database of the voice search device according to Embodiment 1 of the present invention. As shown in the example of FIG. 5, the search database 8 includes at least facility names and reading information divided into words.
The search database 8 is usually provided with a search index created in advance to improve the search efficiency. Information retrieval methods and index creation methods from databases are described in “Information retrieval algorithm, Kenji Kita, Kazuhiko Tsuda, Masami Isogobori, Kyoritsu Publishing Co., Ltd.”. In the present embodiment, an index divided into subwords is created in advance in the search database 8, and the database search unit 9 is configured to be able to search for an arbitrary subword.

  The database search unit 9 refers to the search database 8 and acquires a search result corresponding to the recognition result output by the voice recognition unit 4. As a method for taking into account the ambiguity of speech recognition, the database search unit 9 searches the search database 8 using all the subwords of the recognition result as search keys, and acquires candidate facility names. Furthermore, the database search unit 9 compares all the subwords used for the search with the subwords included in the candidate facility names, and scores the search results based on the number of matching subwords.

FIG. 6 is an explanatory diagram showing an example of a search result of the voice search device according to Embodiment 1 of the present invention. The database search unit 9 outputs a search result including the facility name, reading information, and search score as shown in FIG. 6 to the search result data storage unit 10 and the recognition result correction unit 11. The ID included in the search result is given for the purpose of identifying the facility name, and is unchanged through a series of processes of the voice recognition device 1.
The search result data storage unit 10 stores the search results output from the database search unit 9 as a list.

  The recognition result correction unit 11 refers to the search result data storage unit 10 to collate the list of search results with the recognition result acquired by the speech recognition unit 4, correct a word included in the recognition result, and obtain the corrected recognition result. Output to the candidate presentation unit 12. The recognition result correction unit 11 further corrects the ranking of the search result list by correcting the search score of the search result stored in the search result data storage unit 10 based on the correction recognition result.

  The candidate presenting unit 12 presents data of a predetermined rank of the search results stored in the search result data storage unit 10 to the user, and also presents a corrected recognition result obtained by correcting the recognition result. As a presentation method, for example, the candidate presentation unit 12 includes a monitor screen and displays a search result and a correction recognition result on the screen.

Next, the operation of the voice search device 1 will be described. FIG. 7 is a flowchart showing the operation of the speech search apparatus according to Embodiment 1 of the present invention. Here, taking a facility name search as an example, a recognition accuracy improvement based on a speech recognition result and a database search result, and a method of generating recognition result presentation contents to a user will be described. To simplify the explanation, it is assumed that the user intends to search for the name of the facility by saying “Marchy Dome Zaiten”.
Further, it is assumed that an index divided into subwords is created in advance in the search database 8 shown in FIG. 1, and that an arbitrary subword search by the database search unit 9 is possible.

  In step ST <b> 1 shown in FIG. 7, the speech recognition unit 4 first recognizes the input speech and outputs a recognition result that is input to the database search unit 9. Here, a recognition result of “Marque Doubts Eyetel” is acquired and output by the speech recognition unit 4 with respect to the input sound “Markov Dome Eyes”.

In step ST <b> 2, the database search unit 9 outputs a search result with reference to the search database 8 and stores it in the search result data storage unit 10.
The database search unit 9 subtracts “Marchy Doubutsuiteru” acquired as a recognition result from the subwords “ma”, “le”, “cue”, “do”, “bu”, “tu”, “a”, “b”. , “Te”, “le”. Next, the database search unit 9 uses the subwords as search keys to search the search database 8 in which the search target data shown in FIG. 5 is stored. As a specific search method, for example, in a search method using a vector space model used as a document search method, a method of performing a search using a subword instead of a word used for the search can be considered.
As a result of searching the database for search 8 shown in FIG. 5 by using the “Malque Dove's Eyetel” as a key by the database search unit 9, the search result of the facility name shown in FIG. 6 is acquired and a search score is given. Of these search results, the database search unit 9 outputs the facility names of ID = 8, 1, 9, 10 of the top N (N = 4 here) items as the search results.

In step ST3, the recognition result correction unit 11 first creates a network expression from the top N search results. FIG. 8 is an explanatory diagram showing an example of a network structure diagram created by the recognition result correction unit of the speech search apparatus according to Embodiment 1 of the present invention. This network expression is a network that accepts a paraphrase expression of words (units delimited by “|” in the reading information) included for each facility name of search result ID = 8, 1, 9, 10. The feature of this network is that, for each word 31, a fuzzy network 32, 33 having a structure in which the possibility of subwords that are likely to be mistaken is expanded is generated, and a filler element 34 that passes an arbitrary subword is generated. There is.
Here, the network is such that all words can be connected in an arbitrary order, but may be appropriately modified in consideration of language restrictions. Moreover, you may give a score to the arc which connects each word.

In step ST <b> 4, the recognition result correction unit 11 collates the recognition result “Marquee dovezitel” of the speech recognition unit 4 through the network expression, and creates a correction recognition result candidate.
FIG. 9 is an explanatory diagram showing an example of a correction recognition result candidate created by the recognition result correction unit of the speech search apparatus according to Embodiment 1 of the present invention. The correction recognition result candidate 41 shown in FIG. 9 is created as data obtained by dividing each candidate into words by “|”. Further, a portion that does not match each word constituting the network expression and passes through the filler element is a pseudo word surrounded by “(” and “)”.
For example, among the correction recognition result candidates 41 shown in FIG. 9, “Marchudo | (Butua) | The recognition result “Butua” passes through the filler element three times, and the recognition result “Itel” passes through the ambiguous network of the word “Shoten”.

  A correction score is assigned to each correction recognition result candidate, and represents the similarity between the recognition result “Marque Doubts Eyetel” and each correction recognition result candidate. The correction score is scored such that the recognition result correction unit 11 has a higher evaluation as the degree of coincidence increases based on the similarity of the subwords between the recognition result and the correction recognition result candidate, and is further included in the correction recognition result candidate. When the proportion of subwords that have passed through the filler element is large, scoring that lowers the evaluation is performed.

Subsequently, the recognition result correction unit 11 compares the correction recognition result candidate with the search result stored in the search result data storage unit 10 and corrects the search score included in the search result list.
For example, the recognition result correction unit 11 corrects the search score according to the following equation based on the number of deficiencies and excess words in the correction recognition result candidate compared to the number of words in the search result and the correction score of the correction recognition result candidate. To do.
Corrected search score = search score × {1− (number of deficiencies / number of search result words) × α
− (Excess number / number of search result words) × β
− (1−correction score) × γ}

  For example, if (α, β, γ) = (0.05, 0, 20, 0.03), the shortage of words is allowed as much as possible, the excessive utterance is hardly allowed, and the original recognition is performed. It is possible to make corrections taking into account the gap from the result to some extent.

  When the search score of the search result is actually corrected and the correction recognition result is selected from the correction recognition result candidates according to the correction search score, the result is as follows. Here, a corrected search score is calculated for each of the top M correction scores (for example, M = 5), and a correction recognition result candidate having the maximum corrected search score is selected from the M correction scores. For “Marchodo | Shoten | Mutsai | Ten” with search result ID = 1, among the correction recognition result candidates 41, “Marchodo | Mutsai | Ten” has a corrected search score of 0.828 and the optimum correction recognition result. Selected as. Also, for “Marquedeaux” with ID = 8, “Marquedeau | (Butsitel)” is selected with a corrected search score of 0.668. Further, for “Marquedeaux | Shoten” with ID = 9, “Marquedo | (Butua) | Shoten” is selected with a corrected search score of 0.731. In addition, for “Marquedeau | Chillyoin” with ID = 10, “Marquedeaux | (Butsitel)” is selected with a corrected search score of 0.708.

  In this way, the recognition result correction unit 11 corrects the search score stored in the search result data storage unit 10, and re-retrieves the top N search results in the order of ID = 1, 9, 10, 8 according to the corrected search score. To score.

  In step ST5, the candidate presentation unit 12 displays the search result on a screen (not shown). At this time, the candidate presentation unit 12 presents the search results in the order of the corrected search scores, and also presents the correction recognition results selected for each search result.

FIG. 10 is an explanatory diagram showing a presentation example of the candidate presentation unit of the speech search device according to Embodiment 1 of the present invention. In the display screen 51 shown in FIG. 10, a list 53 of search results stored in the search result data storage unit 10 is presented in the order based on the corrected search score. A correction recognition result corresponding to the search result being selected is also presented.
At the top of the search result list 53, “○ Kyodo Shoten Rokukai” with the highest corrected search score is displayed. In the recognition result display frame 52, the correction recognition result “Marchyo Domes Eye Ten” corresponding to the currently selected “○ Kyodo Bookstore 6kai” is displayed.
Note that “Shoten” not spoken by the user is not included in the recognition result, and a recognition result 52 that is easy for the user to understand is presented.

  When the user moves the selection of the search result list 53 on the display screen 51 presented by the candidate presentation unit 12 and makes “○ Kudo Shoten” selected, the candidate presentation unit 12 displays the display screen 51 in step ST6. Transitions to the display screen 54. On the display screen 54, “○ Kudo Shoten” in the search result list 56 is selected, the display of the recognition result display frame 55 is also changed to match the selected item, and becomes “Marquedo (Butsua) Shoten”. A portion surrounded by “(” and “)” indicates that there is no portion corresponding to the search result and an unknown recognition result is included in the system.

  As described above, according to the first embodiment, the speech recognition apparatus 1 uses the speech recognition unit 4 that outputs the facility name corresponding to the input speech as the recognition result, and the key obtained by dividing the recognition result into subword units. A database search unit 9 that searches the search database 8 that stores facility names including an index divided into subwords in units smaller than words, and outputs a facility name having a high similarity as a search result. A recognition result correction unit 11 that collates recognition results using a network expression that includes included words and outputs a corrected recognition result obtained by correcting the recognition result, and presents a search result and a corresponding correction recognition result. The candidate presentation unit 12 is provided. Therefore, a search considering the ambiguity of the voice recognition of the voice recognition unit 4 can be performed, and the validity of the voice recognition result can be verified and corrected using the search result, thereby improving the input voice recognition accuracy. At the same time, it is possible to present a recognition result that is easy for the user to understand.

  In addition, according to the first embodiment, the speech recognition apparatus 1 has corrected recognition results for the top N search results among the search results that the recognition result correction unit 11 has ranked based on the similarity to the recognition results. The ranking is corrected and the search result is presented according to the ranking corrected by the candidate presentation unit 12. As a result, the user is more likely to obtain the desired recognition result, and the corrected recognition result is also in accordance with the search result, so that the user can know how the voice search device has recognized the user's utterance. It becomes possible to present it in an easy-to-understand manner.

  Furthermore, since the correction recognition result presented by the candidate presentation unit 12 is corrected based on the search result, even when the first utterance is not utterance content sufficient to specify the facility name to be searched. By using the recognition results and further uttering additional keywords, an interface for implementing a refined search can be easily constructed.

  In the first embodiment, the candidate presentation unit 12 is configured to present the search result and the corrected recognition result at the same time, but may be configured to present only one of them.

Further, in the first embodiment, the word break is registered in advance in the reading information of the search database 8, but the database search unit 9 automatically divides the word as necessary when the recognition result is acquired. It may be configured.
Further, when there are a plurality of candidates at the word break of the reading information in the search database 8, a plurality of candidates may be written together and the database search unit 9 may collate with the recognition result.

  In the first embodiment, the recognition result correction unit 11 is configured to create a network expression including all the top N search results. However, a network expression is generated for each search result, and the correction recognition is performed. You may comprise so that a result candidate and a correction score may be calculated | required.

  Further, in the first embodiment, the recognition result correction unit 11 creates the network representation component in units of words. However, the configuration may be such that the component is created in units of morpheme.

Embodiment 2. FIG.
FIG. 11 is a block diagram showing the configuration of the speech search apparatus according to Embodiment 2 of the present invention. The speech search apparatus 1a according to the second embodiment corrects the recognition result correction unit 11 according to the first embodiment, because the process is simple and lightweight, and the recognition result is focused on the continuity of the subword sequence. In this configuration, the recognition result correction unit 61 is replaced. Since the configuration other than the recognition result correction unit 61 is the same as that of the voice search device 1 of the first embodiment, detailed description thereof is omitted.
The recognition result correction unit 61 collates the search result divided into subword units with the recognition result in subword units, and selects a word including a subword string that matches the search result among the recognition results as a subword string of the search result. Perform correction to replace words that contain.

  Next, the operation of the voice search device 1a will be described. FIG. 12 is a flowchart showing the operation of the speech search apparatus according to Embodiment 2 of the present invention. As in the first embodiment, it is assumed that the user intends to search for a facility name by saying “Marchyo Dome Aiten”.

Steps ST11 and ST12 shown in FIG. 12 are the same processes as steps ST1 and ST2 shown in FIG. 7, and the search result data storage unit 10 of the voice search device 1a stores the search results shown in FIG.
In subsequent step ST13, the recognition result correction unit 61 takes out the search results in the search result data storage unit 10 one by one, and assigns serial numbers to the subwords. Here, a case is considered where the recognition result correction unit 61 acquires “Market Shot Tenmite” with ID = 1 as a search result and performs matching with the recognition result.
FIG. 13 is an explanatory diagram for explaining the correction processing by the recognition result correction unit of the speech search apparatus according to Embodiment 2 of the present invention. The recognition result correction unit 61 converts the search result extracted from the search result data storage unit 10 into subword units such as phonemes and syllables, “ma”, “le”, “kyo”, “do”, “sho”, “te”. , “N”, “M”, “TSU”, “A”, “I”, “TE”, “N”, and a serial number is assigned in order from the top to obtain a numbered search result 71.

  In step ST <b> 14, the recognition result correction unit 61 assigns the same number to the same subword as the numbered search result 71 for each subword of the recognition result “Marque Doubts Iter”, and sets it as the number assignment recognition result 72. Since the subword “ma” of the number assignment recognition result 72 is the same subword as “ma = 1” of the numbered search result 71, “1” is assigned by the recognition result correction unit 61. In addition, the subword “queue” of the number assignment recognition result 72 is not assigned a number because the corresponding subword does not exist in the numbered search result 71. Further, since the subword “te” of the number assignment recognition result 72 is the same as the two subwords “te = 6” and “te = 12” of the numbered search result 71, the recognition result correction unit 61 sets “6”. "And" 12 "are assigned.

In step ST15, the recognition result correction unit 61 extracts consecutive number strings in the number assignment recognition result 72 in order of increasing number of consecutive numbers. When the recognition result correction unit 61 selects a number sequence in consideration of continuity from the number assignment recognition result 72, “1, 2,” “4”, “9, 10, 11, 12,” “6”, “2”. Are selected. The recognition result correction unit 61 first acquires “9, 10, 11, 12” having the longest continuous number from these number sequences. At this time, since “6” is a subword at the same position as “12”, it is excluded from the subsequent acquisition candidates. Next, the recognition result correction unit 61 obtains the second longest “1, 2” as a number string not including “9” to “12”. At this time, the number strings “4” and “2” remain, but since “2” has already been acquired, the recognition result candidate 61 acquires “4”.
Thus, in step ST15, the number strings “9, 10, 11, 12”, “1, 2,” and “4” indicated by the double square frames are acquired from the number assignment recognition result 72.

In step ST <b> 16, the recognition result correction unit 61 evaluates the correspondence between the serial number subword string and the numbered search result 71 word in the order of the long number string, and creates a correction recognition result 73.
First, the recognition result correction unit 61 applies the word “Mutsuai | Ten” in the numbering search result 71 with the same number assigned to the subword sequence to which the longest number sequence “9, 10, 11, 12” is assigned. ”And compare them. Then, the recognition result correction unit 61 corrects “bu” of “number” in the number assignment recognition result 72 to “mu” and “le” of “tel” to “n”.
Secondly, the recognition result correction unit 61 compares the word “Marko Doo” in the numbering search result 71 with the same number with the sub-word string to which the number string “1, 2” is assigned and compares them. . Furthermore, the recognition result correction unit 61 determines that “do = 4” of the number assignment recognition result 72 that matches the partial subword of the word “Marquis” is also associated with the word “Marquis”. As a result, the recognition result correction unit 61 corrects “cue” of “marked” in the number assignment recognition result 72 to “kyo”. In this way, the recognition result correction unit 61 performs correction to replace a word including a serial number subword string in the number assignment recognition result 72 with a word in the numbering search result 71 having the same number.
Note that the word “Shoten” included in the numbering search result 71 does not have a subword having the same number in the number assignment recognition result 72.

  In step ST17, the recognition result correction unit 61 makes it difficult to specify a subword string that does not exist in the numbering search result 71 among the number assignment recognition results 72, so that the correction recognition result 73 remains as a subword of the recognition result. To leave. In the example shown in FIG. 13, since all subwords in the recognition result are specified as search result words, the recognition result correction unit 61 does not perform the process in step ST205 and proceeds to the next process.

  When the recognition result correction unit 61 corrects the recognition results for all the search results stored in the search result data storage unit 10, the candidate presentation unit 12 performs the search results and the correction recognition in subsequent steps ST18 and ST19. Present the results. Steps ST18 and ST19 are the same processes as steps ST5 and ST6 shown in FIG.

  As described above, according to the second embodiment, in the speech recognition device 1a, the recognition result correction unit 61 divides the search result and the recognition result into subwords and collates them, and the subword that matches the search result among the recognition results. A correction recognition result in which a word including a column is replaced with a word including a sub-word string as a search result is output, and the candidate presentation unit 12 is configured to present the search result and a correction recognition result corresponding thereto. Therefore, as in the first embodiment, it is possible to present a correction recognition result that is easy for the user to understand together with the search result.

In the voice search device 1a according to the second embodiment, the recognition result correction unit 61 may have a subword pair that is likely to be mistaken during voice recognition and information on the ease of mistake. The subword pairs that are likely to be mistaken are, for example, “Kyo” and “Cue”, and information on the likelihood of mistake of this pair, that is, the possibility of mistakes is 0.2.
In step ST14 (FIG. 7) in which numbers are assigned to the recognition results, the recognition result correction unit 61 assigns numbers with weights based on the information on the ease of error when there is an easily mistaken subword, and the continuity of the subword string is increased. Judge with weight. For example, the recognition result correction unit 61 assigns the same number “3” as “Kyo = 3” to the “queue” of the number assignment recognition result 72 shown in FIG.

Further, the correction method of the recognition result correction unit 61 of the second embodiment is an algorithm that is simple in processing and places importance on the continuity of the subword sequence. It is also possible to apply as post-processing for each correction recognition result of the created correction recognition result candidate. That is, as described in the first embodiment, the recognition result correction unit collates the recognition result through the network expression, replaces a word included in the recognition result with a word similar to the search result, and outputs the corrected recognition result. After creating and correcting the ranking of search results, as described in the second embodiment, the search result and the corrected recognition result are collated in units of subwords, and the subword that matches the search result among the recognition results In this configuration, a word including a column is corrected to be replaced with a form word including a sub-word column as a search result.
In the case of this configuration, the speech recognition apparatus can create a corrected recognition result that further considers word chain.

  Further, in the first embodiment and the second embodiment, the description has been given by using the voice search device for Japanese as an example, but the target language is not limited, and in other languages than the word It is possible to configure a speech search device using phonemes or phoneme strings which are small units.

It is a block diagram which shows the structure of the speech search device concerning Embodiment 1 of this invention. It is explanatory drawing which shows the structure of the hidden Markov model used with the speech search device concerning Embodiment 1 of this invention. It is explanatory drawing which shows one example of the word dictionary for recognition of the speech search device concerning Embodiment 1 of this invention. It is explanatory drawing which shows one example of the word dictionary for recognition of the speech search device concerning Embodiment 1 of this invention. It is explanatory drawing which shows one example of the database for a search of the speech search device concerning Embodiment 1 of this invention. It is explanatory drawing which shows an example of the search result of the speech search device concerning Embodiment 1 of this invention. It is a flowchart which shows operation | movement of the speech search device based on Embodiment 1 of this invention. It is explanatory drawing which shows one example of the network structure figure which the recognition result correction | amendment part of the speech search device concerning Embodiment 1 of this invention produces. It is explanatory drawing which shows an example of the correction | amendment recognition result candidate which the recognition result correction | amendment part of the speech search device concerning Embodiment 1 of this invention produces. It is explanatory drawing which shows the example of a presentation of the candidate presentation part of the speech search device concerning Embodiment 1 of this invention. It is a block diagram which shows the structure of the speech search device which concerns on Embodiment 2 of this invention. It is a flowchart which shows operation | movement of the speech search device based on Embodiment 2 of this invention. It is explanatory drawing explaining the correction process by the recognition result correction | amendment part of the speech search device which concerns on Embodiment 2 of this invention.

Explanation of symbols

  1, 1a voice search device, 2 acoustic standard pattern database, 3 recognition word dictionary, 4 speech recognition unit, 5 speech analysis unit, 6 collation unit, 7 search unit, 8 search database, 9 database search unit, 10 search result Data storage unit, 11 recognition result correction unit, 12 candidate presentation unit, 21 state, 22 autoregressive arc, 31 words, 32, 33 fuzzy network, 34 filler element, 41 correction recognition result candidate, 51 display screen, 52 recognition result display Frame, 53 search result list, 61 recognition result correction unit, 71 numbered search result, 72 number assignment recognition result, 73 correction recognition result.

Claims (6)

A speech recognition unit that outputs a word string corresponding to the input speech as a recognition result;
A search unit that divides the recognition result into subwords in units smaller than words and uses them as a search key, searches a search target document including an index divided in subword units, and outputs a search result;
A recognition result correction unit that corrects the recognition result based on the search result;
A speech search apparatus comprising: a candidate presentation unit that presents at least one of the search result and the recognition result corrected by the recognition result correction unit.   The recognition result correction unit compares the search result divided into subwords and the recognition result in word units, and performs correction to replace a word included in the recognition result with a word similar to the search result. The voice search device according to claim 1. The search unit outputs multiple search results,
The recognition result correction unit calculates a search score according to the number of subword matches between the recognition result and each search result, assigns a ranking to the plurality of search results, and searches for a predetermined number of higher ranks among the plurality of search results For the result, the upper predetermined number of search results divided into subwords and the recognition result are collated in word units, and the words included in the recognition result are replaced with words similar to the upper predetermined number of search results. Performing correction, recalculating the search score of each search result using the corrected recognition result, correcting the order of the plurality of search results,
The voice search device according to claim 1, wherein the candidate presentation unit presents the plurality of search results according to the ranks given by the recognition result correction unit.   The recognition result correction unit collates the search result divided into subword units and the recognition result in subword units, and selects a word including a subword string that matches the search result among the recognition results. The speech search apparatus according to claim 1, wherein correction is performed by replacing with a word including a subword string. The search unit outputs multiple search results,
The recognition result correction unit calculates a search score according to the number of subword matches between the recognition result and each search result, assigns a ranking to the plurality of search results, and searches for a predetermined number of higher ranks among the plurality of search results For the result, the upper predetermined number of search results divided into subwords and the recognition result are collated in word units, and the words included in the recognition result are replaced with words similar to the upper predetermined number of search results. The correction is performed, the search score of each search result is recalculated using the corrected recognition result, the order of the plurality of search results is corrected, and each search result and the corrected recognition result are Correction is performed by substituting a word including a subword string that matches the search result among the recognition results with a word including the subword string in each search result by collating in units of subwords. Voice search apparatus according to claim 1, wherein a. A speech recognition step of outputting a word string corresponding to the input speech as a recognition result;
A search step of dividing the recognition result into subwords in units smaller than words and using them as search keys, searching for a search target document including an index divided in subword units, and outputting the search results;
A recognition result correction step for correcting the recognition result based on the search result;
A speech search method comprising: a candidate presentation step for presenting at least one of the search result and the recognition result corrected in the recognition result correction step.

Images