JP4756499B2 - Voice recognition result inspection apparatus and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an inspection device of the voice recognition results, where a viewpoint whether they are suitable to use for translation is introduced in the checks of the results. <P>SOLUTION: This inspection device 60 is used together with the collection 56 of the phrases extracted from the bilingual corpus 34 to check whether to use the words constituting the word series of the voice recognition result 50 of the voice recognition device 48 for the purpose of processing in the translation machine 68 of a phrase base. A reliability level is attached in advance to each word to check. This inspection device 60 includes a suitability evaluation section 70 to give a suitability degree for the translation 68 to each word as a function of the word string including each word and having a matching phrase in the phrase collection 56 among the word series making part of the word string in the voice recognition result, and an accept/no-accept deciding section 74 of the word after comparing the threshold level 82 determined according to the suitability degree of the word and the reliability level of the word. <P>COPYRIGHT: (C)2007,JPO&amp;INPIT

Description

  The present invention relates to an apparatus for inspecting whether or not a word constituting a word string of a speech recognition result should be accepted as a target of subsequent processing, and in particular, has a reliability of a result of speech recognition processing for speech translation processing. In particular, the present invention relates to an apparatus for performing inspection in consideration of suitability for subsequent translation processing.

  Speech translation is a process of receiving speech in a certain language, translating the utterance content into another language and outputting it. The technology to automate and improve the performance of speech translation is one of the goals of spoken language processing technology. Speech translation is generally realized by the following two technologies. That is, voice recognition for identifying a word or sentence uttered from an uttered voice and machine translation for converting the identified word or sentence into a word or sentence in another language.

  In order to improve speech translation performance, it is necessary to closely link speech recognition and machine translation. Various techniques for this have been proposed. For example, Non-Patent Document 1 discloses a technique that uses one integrated statistical model for speech recognition and machine translation. Non-Patent Document 2 discloses a technique for generating a plurality of candidates as a speech recognition result, performing machine translation on each candidate, and selecting an optimum candidate from the plurality of translation results.

  In order to realize high-performance speech translation, it is necessary to improve both speech recognition performance and machine translation performance. As a technique for improving machine translation performance, a technique for prolonging the unit of translation processing has been proposed in statistical machine translation (SMT). For example, Non-Patent Document 3 discloses a phrase-based statistical translation technique using a “phrase” composed of a word or a plurality of connected words as a unit of processing. In this technique, a phrase automatically extracted from a bilingual corpus in the course of learning a statistical model for statistical translation (hereinafter referred to as a translation model) from the bilingual corpus is used as a translation unit. By using a translation processing unit as a phrase, it is possible to obtain a natural translation result in which words naturally connect to each other as compared to the case where translation is performed in units of words.

  Recent progress in research and improvements in computer performance have enabled speech recognition with considerable accuracy. However, there are many factors that hinder speech recognition, such as noise, speaker changes, and non-grammatical speech, making it difficult to achieve sufficient speech recognition performance. Therefore, it is important to accurately detect and correct a speech recognition error in the speech recognition process before the speech recognition result is translated. As a technique for that purpose, a technique for evaluating whether or not a word string (or a word constituting the word string) obtained as a speech recognition result is reliable has been developed. For example, in Non-Patent Document 4, a generalized word posterior probability (GWPP) based on a word graph output as a result of speech recognition and a word posterior probability assigned to each word. GWPP is used to evaluate each word of the speech recognition result.

Y. Gao, “Coupling vs Unification: Modeling Techniques for Speech-to-Speech Translation”, Eurospech 2003 Proceedings, pages 365-368 (2004) (Y. Gao, “Coupling vs. unifying: Modeling technique for speech-speech translation” , Proc. Of Eurospeech 2003, pp.365-368 (2004)) R. Tsang et al., “Improved Spoken Language Translation Using N-Best”, ICSLP2004 Proceedings, 1629-1632 (2004) (R. Zhang et al., “Improved Spoken Language Translation Using N-Best Speech Recognition Hypothesis ", Proc. Of ICSLP 2004, pp.1629-1632 (2004)) E. Sumita et al., “Example Translation (EBMT), Statistical Translation (SMT), Hybrid, etc .: ATR Spoken Language Translation System”, IWSLT 2004 Proceedings, pp. 13-20, (2004) (E. Sumita et al., “EBMT, SMT, Hybrid and More: ATR Spoken Language Translation System "Proc. Of IWSLT 2004, pp.13-20 (2004)) Frank K. Soon et al., “Optimal acoustic and language model weights to minimize word evaluation errors”, ICSLP 2004 Proceedings, 441-444, (2004) (Frank K. Soong et al., “Optimal Acoustic and Language Model Weights for Minimizing Word Verification Errors "Proc. Of ICSLP 2004, pp.441-444 (2004))

  In the technique described in Non-Patent Document 1, one integrated statistical model is used for speech recognition and machine translation. However, this integrated statistical model is not necessarily the optimal model for both speech recognition and machine translation. Therefore, there is a risk that the performance of either speech recognition or machine translation is sacrificed. In fact, in machine translation, a long word string such as a phrase can be processed as a processing unit as in the technique described in Non-Patent Document 3, whereas in speech recognition, such a long processing unit is used. If the process is performed, the recognition performance may be deteriorated. In the technique described in Non-Patent Document 2, it is necessary to perform machine translation processing on a plurality of candidates. The amount of processing increases accordingly.

  Therefore, in order to make the cooperation between speech recognition and machine translation closer, it is necessary to verify the speech recognition result from the viewpoint of whether or not the speech recognition result is suitable for machine translation. However, Non-Patent Documents 1 to 3 do not specifically describe a technique for performing such evaluation.

  In the technique described in Non-Patent Document 4, evaluation is performed as to whether each word of the speech recognition result is reliable as the recognition result. However, this technique does not evaluate from the viewpoint of “whether the speech recognition result is suitable for machine translation”. Therefore, even a speech recognition result highly appreciated by this technique may not be suitable for machine translation at all.

  Such a problem may occur not only in the combination of speech recognition and machine translation, but also in the combination of speech recognition and general natural language processing using the result.

  Therefore, it is an object of the present invention to introduce a viewpoint of whether or not the speech recognition result is suitable for subsequent natural language processing based on the reliability of the speech recognition result. It is to provide an inspection apparatus.

  Another object of the present invention is to provide an apparatus for inspecting a speech recognition result for maintaining a high level of speech recognition performance and machine translation performance, and making the processing of both processes closer. .

  The speech recognition result inspection apparatus according to the first aspect of the present invention receives a word string of a speech recognition result generated from a predetermined input speech by speech recognition processing, and is based on a predetermined phrase base following the speech recognition processing. This is an apparatus for examining whether or not a word constituting a word string of a speech recognition result should be accepted as a target of natural language processing. The speech recognition result inspection apparatus is used together with a set of phrases extracted by a predetermined extraction method from a corpus for natural language processing. Each word constituting the word string of the speech recognition result is given reliability in advance by the speech recognition process. The speech recognition result inspection device matches, for each word constituting the word sequence of the speech recognition result, a word sequence that includes the word and within a set of phrases in a word sequence that forms a partial word sequence of the speech recognition result As a function of a set of word strings having a phrase to be matched, a fitness level giving means for giving a fitness level for natural language processing, and for each word constituting the word string of the speech recognition result, the fitness level giving means assigns the word to the word. Determining means for determining whether or not to accept the word by comparing the threshold value determined according to the given degree of matching with the confidence level given to the word;

  First, the degree of conformity is given to the words constituting the word string of the speech recognition result by the degree-of-fit provision unit. The goodness-of-fit is given as a function of a set of words having a matching phrase in a set of phrases extracted from the corpus among word strings forming a partial word string of the speech recognition result including the word. If there is a partial word string having a matching phrase in the phrase set in the word string of the speech recognition result, it is considered that the partial word string is compatible with the corpus. That is, the partial word string is considered suitable for subsequent natural language processing performed using a corpus. The determining means determines whether or not each word should be accepted by comparing the threshold value determined according to the degree of matching with the reliability. Therefore, it is possible to introduce the viewpoint of “whether it is suitable for natural language processing using a corpus” in the inspection of each word based on reliability, without reducing the performance of speech recognition and machine translation. , Can strengthen the linkage between speech recognition and natural language processing.

  Preferably, the matching degree assigning unit collates the phrase set with a word string forming a partial word string of the speech recognition result, thereby detecting a word string having a matching phrase in the phrase set. A degree of conformity is given to each word constituting the word sequence of the speech recognition result and the means according to a predetermined criterion based on a set of word sequences including the word among the word sequences detected by the matching unit Means.

  The collation means collates the word string forming the partial word string of the speech recognition result with the phrase set, thereby detecting a word string having a matching phrase in the phrase set. The assigning means assigns a fitness level to each word constituting the word string of the speech recognition result based on a set of word strings including the word among the detected word strings. By doing so, it is possible to efficiently detect a word string having a matching phrase and give a matching degree to each word.

  Preferably, the fitness level assigning unit gives the fitness level for natural language processing to each word constituting the word sequence of the speech recognition result as a function of the length of the word sequence included in the set of word sequences for the word. Means for doing so.

  For each word constituting the word string of the speech recognition result, the degree of fitness is given as a function of the length of the partial word string that includes the word and has a matching phrase in the phrase set. Therefore, it is possible to evaluate whether or not the recognition result is suitable for natural language processing in which the processing performance varies depending on the length of the portion where the matched phrase exists based on the degree of matching.

  The means for giving is, for each word constituting the word string of the speech recognition result, out of the word strings included in the set of word strings for the word, the number of words having the maximum number of words constituting the word string May be given as the fitness.

  The suitability with natural language processing in which the processing performance improves as the length of the matched phrase is longer can be expressed by this suitability.

  The natural language process may include a phrase-based statistical translation process between the language of the input speech and a predetermined target language. The set of phrases includes a set of phrases extracted in the process of learning a translation model for statistical translation processing from a bilingual corpus of the input speech language and a predetermined target language. The fitness level assigning means includes means for giving a fitness level for natural language processing as a function of a set of word strings for each word constituting the word sequence of the speech recognition result.

  The degree of matching given based on the detected phrase represents the matching between the bilingual corpus and the translation model learned from the corpus. Therefore, the viewpoint of whether each word of the speech recognition result is suitable for phrase-based statistical translation can be introduced into the speech recognition result inspection, and the cooperation between speech recognition and statistical translation can be strengthened.

  The determining means is for holding the matching level and the threshold in association with each other and for each word constituting the word string of the speech recognition result based on the matching level given to the word Is set by means for setting a threshold for the word according to the fitness and the threshold held by the means, and for each word constituting the word string of the speech recognition result. And comparing means for determining whether or not to accept the word by comparing the threshold value and the reliability assigned to the word.

  For each word, each of the constituents of the speech recognition result is determined by comparing whether the word should be accepted by comparing the threshold set according to the degree of fitness of the word and the reliability of the word. Words can be inspected for compatibility with the corpus.

  The threshold value held by the means for holding may be selected such that the threshold value decreases as the fitness level increases.

  By selecting such a threshold value, even a word with low reliability can be accepted if a partial word string including the word is suitable for subsequent natural language processing. If the partial word string is suitable for the subsequent phrase-based natural language processing, the probability that the natural language processing for the portion will fail is low. Therefore, as a result, the performance of a series of processes including a speech recognition process and a phrase-based natural language process performed using the result is improved.

  When the computer program according to the second aspect of the present invention is executed by a computer, it causes the computer to operate as any of the speech recognition result inspection apparatuses described above. Therefore, it is possible to obtain the same effects as those of the voice recognition result inspection apparatus described above.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the drawings used in the following description, the same reference numerals are assigned to the same items. Their names and functions are also the same. Therefore, description thereof will not be repeated.

[Overview]
In the present embodiment, whether or not each word of a speech recognition result should be accepted as a target for subsequent translation processing in speech translation by a combination of speech recognition and phrase-based statistical translation which is a kind of natural language processing. inspect. In this inspection, for each word, pass / fail as a processing target of each word is determined by comparing a value representing the reliability of the word (hereinafter, this value is referred to as “reliability”) with a threshold value. . In the present embodiment, evaluation from the viewpoint of suitability of speech recognition results for translation processing is introduced to the problem of checking each word by comparing the reliability with a threshold value. That is, it is evaluated whether the speech recognition result is suitable for translation, and the threshold value is changed for each word according to the result of the evaluation. Hereinafter, the fitness index is referred to as “fitness”.

  In the present embodiment, a phrase generated in the learning process of a translation model used in phrase-based statistical translation is used to obtain a goodness of fit. When learning a translation model from a bilingual corpus composed of a large number of parallel translations of a source language and a target language, a large number of phrases are extracted as by-products. They form a set of phrases. The phrase that is an element of this set is a part of a sentence in the bilingual corpus and can be said to be a phrase that frequently appears in the domain defined by the bilingual corpus. Therefore, if there is a part that matches such a phrase in the speech recognition result, it can be considered that the part is compatible with the domain. Furthermore, the longer the part, the higher the suitability of that part. Therefore, in the present embodiment, the length of the portion that matches the phrase extracted from the bilingual corpus is used as a measure for evaluating the suitability of each word included in the speech recognition result. The number of words is used as a measure of length. That is, for each word of the speech recognition result, a partial word string of the speech recognition result including the word, and considering a set having a matching phrase in the set of phrases described above, a portion included in this set The degree of suitability of the word is determined by the maximum length of the word string. In the following description, the length of the phrase is referred to as “phrase length”. Since the partial word string is also considered as one phrase, the length of the partial word string is also called “phrase length”. A partial word string that matches one of the phrases in the phrase set in the speech recognition result is referred to as a “matching portion”. Of the set of all matching parts including a word, the phrase length with the longest phrase length is called “maximum matching length (of the word)” and is represented by “Lmax”.

[Constitution]
(Overall structure of speech translation system)
FIG. 1 is a block diagram showing the configuration of a speech translation system 30 according to this embodiment. Referring to FIG. 1, the speech translation system 30 includes a translation model learning device 54 for learning a translation model from a predetermined bilingual corpus 34, and a translation for holding a translation model learned by the translation model learning device 54. Model part 52 is included. The translation model learning device 54 has a function of extracting a phrase from the bilingual corpus 34 in the course of learning the translation model. The speech translation system 30 further includes a phrase database 56 for storing phrases extracted by the translation model learning device 54.

  The speech translation system 30 further includes a translation speech 32 and a learning speech 38A prepared in advance for learning in response to an operation input 42 for selecting one of speech translation and the threshold learning process described above. .., 38P and a selector 44 for selecting the speech 46 to be processed, and speech recognition processing is performed on the speech 46 selected by the selector 44, and the speech recognition word string and each word of the word string The speech recognition device 48 for outputting the speech recognition result 50 including the reliability and the inspection of each word of the speech recognition result 50 are determined according to the maximum matching length Lmax of the word with respect to the phrase in the phrase database 56. Based on the threshold value and the reliability of the word, a test result 76 including the speech recognition result 50, the maximum match length Lmax of each word, and pass / fail information is generated. An inspection device 60 for outputting, a switch having first and second outputs, interlocking with the selector 44, and outputting the inspection result 76 to the first or second output according to the operation input 42 78.

  The speech translation system 30 is further connected to the first output of the switch 78. When a test result 76 is given from the switch 78, the speech translation system 30 responds to the pass / fail of each word with respect to the speech recognition result 50 included in the test result 76. Pre-processing device 64 for performing predetermined pre-processing and phrase-based statistical translation using translation model unit 52 translates speech recognition result 66 pre-processed by pre-processing device 64 into translation result 36. And a translation device 68. In the present embodiment, the preprocessing device 64 removes rejected words from the word sequence of the speech recognition result 50 based on the inspection result 76, and generates and outputs a preprocessed speech recognition result 66. .

  The speech translation system 30 is further connected to the second output of the switch 78. When the test result 76 for the learning speech 38A,..., 38P is given from the switch 78, the test result 76 and the learning speech 38A,. , 38P, the threshold value learning unit 84 for learning the threshold value for each maximum matching length Lmax from the correct answers (hereinafter referred to as “correct answer test results”) 40A,. And a threshold value table unit 82 for holding a threshold value table composed of threshold values for each maximum matching length Lmax obtained by learning by the threshold value learning unit 84. The switch 78 is interlocked with the selector 44. That is, when the selector 44 selects the translation voice 32 according to the operation input 42, the inspection result 76 is output to the preprocessing device 64, and the selector 44 selects the learning voices 38A,. Sometimes, the inspection result 76 is output to the threshold value learning unit 84.

  The inspection device 60 obtains the maximum matching length Lmax of each word of the speech recognition result 50 for the phrase in the phrase database 56, and generates and outputs a fitness evaluation result 72 composed of the speech recognition result 50 and the maximum matching length Lmax of each word. The pass / fail of each word of the speech recognition result 50 included in the suitability evaluation result 70 and the appraisal result 72 of the suitability, the reliability of the word, the maximum match length Lmax, and the threshold in the threshold table 82 And a pass / fail determination unit 74 for generating and outputting a test result 76 that is determined by using the value table and includes the evaluation result 72 of the matching degree and the pass / fail of each word.

(Speech recognition result 50 by the speech recognition device 48)
FIG. 2 schematically shows the outline of the speech recognition processing performed by the speech recognition device 48 and the configuration of the speech recognition result 50 output by the speech recognition device 48 as a result. With reference to FIG. 2, the speech recognition device 48 first performs speech recognition processing 100 of the given speech 46, and generates a word graph 102 that expresses the utterance content that can output the speech 46 by a combination of words. The word graph 102 is a graph composed of a route network composed of paths corresponding to words, and a score given by the speech recognition device 48 is stored in each path. A route when traveling from the start point to the end point through the route network corresponds to a word string for outputting the voice 46. Further, the posterior probability and GWPP of each word are calculated based on this route network. Therefore, the voice recognition device 48 performs processing 104 for selecting a route from the route network of the word graph 102 based on the score of each word. The route selected by this processing 104 is the speech recognition result 50 according to the present embodiment.

The speech recognition result 50 includes the word string 106 composed of the words W _{1 to} W _N of the route selected from the word graph 102 and the GWPP (GWPP 1 to GWPP 1 to GW) calculated from the word graph 102 for each of the words W _{1 to} W _N. And a sequence 108 consisting of GWPPN). In the speech translation system 30 according to the present embodiment, the GWPP calculated for each word is used as the reliability of the word in the speech recognition result 50.

(Phrase database 56)
FIG. 3 schematically shows the data structure of the phrase database 56 (see FIG. 1). Referring to FIG. 3, phrase database 56 includes phrases 120 </ b> A,. Phrases 122A,..., 122Q. Phrases 120A,..., 120Q and 122A,... Therefore, these phrases have no relation to grammatical units such as phrases. The phrase length varies depending on the phrase. Each of these phrases is obtained by segmenting a sentence in the bilingual corpus 34 by a predetermined statistical method. Therefore, all of these phrases can be said to be frequent phrases in the domain defined by the bilingual corpus 34.

(Fitness evaluation result 72 and inspection result 76)
The evaluation of the fitness by the inspection device 60 and the inspection using the reliability are performed in units of words. 4A and 4B show data structures of the fitness evaluation result 72 and the inspection result 76, respectively. Referring to FIG. 4A, the fitness evaluation result 72 includes a speech recognition result 50 shown in FIG. 2 and a sequence 140 composed of the maximum match length Lmax of words W _{1 to} W _N in the speech recognition result 50. Including. In the present embodiment, when the number of elements of a set including matching parts including the word is 0, the maximum matching length Lmax of the word is set to 0. Referring to FIG. 4B, the inspection result 76 includes a conformity evaluation result 72 and an inspection result column 144 representing pass / fail of the words W _{1 to} W _N. In FIG. 4B, the pass / fail of the words W _{1 to} W _N is represented by “OK (1)” and “NG (0)” in the inspection result column 144.

(Configuration of threshold value table unit 82)
FIG. 5 shows the configuration of the threshold value table held in the threshold value table unit 82. Referring to FIG. 5, threshold table 160 includes five threshold values T0, T1, T2, corresponding to Lmax = 0, Lmax = 1, Lmax = 2, Lmax = 3, and Lmax> 3, respectively. Includes T3 and T4. The threshold values T0, T1, T2, T3, and T4 are adjusted by learning by the threshold value learning unit 84 shown in FIG.

(Configuration of conformity evaluation unit 70)
FIG. 6 is a block diagram showing a functional configuration of the fitness evaluation unit 70. With reference to FIG. 6, the fitness evaluation unit 70 includes a recognition result storage unit 180 for storing the input speech recognition result 50, and a partial in the speech recognition result 50 stored in the recognition result storage unit 180. A partial sentence having a matching phrase in the phrase database 56 by comparing a simple word string (hereinafter referred to as “partial sentence”) with the phrases 120A,..., 120Q (see FIG. 3) in the phrase database 56. For each word of the speech recognition result 50, the collation unit 182 for detection is calculated by the Lmax calculation unit 184 for calculating the maximum match length Lmax based on the detection result by the collation unit 182 and the Lmax calculation unit 184. The maximum matching length Lmax of each word is given as the matching level for the word of the speech recognition result 50, and the matching for forming the matching score evaluation result 72 in the recognition result storage unit 180 Comprising a deposition unit 185, an output unit 186 for outputting it reads from the recognition result storage unit 180 and evaluation results 72 fitness is formed.

  The collation unit 182 is generated with a partial sentence generation unit 190 for generating all partial sentences from the word string 106 of the speech recognition result 50 and a position indicator representing the position of the partial sentence in the word string 106. The partial sentence storage unit 192 for storing the partial sentence and its position indicator, and a phrase that matches the partial sentence in the partial sentence storage part 192 are searched in the phrase database 56, and if there is a matching phrase, A search unit 194 for outputting the position indicator as a position indicator of the matching portion and a matching part storage unit 196 for storing the position indicator output from the search unit 194 are included. In the matching part storage unit 196, all the position indicators output from the search unit 194 are stored. That is, the matching part storage unit 196 stores the position markers of all the matching parts.

  Based on the position indicator in the matching part storage unit 196, the Lmax calculation unit 184 determines a set of matching parts including the word for each word in the word string 106 in the recognition result storage unit 180. The Lmax calculating unit 184 further obtains the phrase length of each matching portion included in the set, obtains the maximum matching length Lmax for the word from the phrase length, and outputs the maximum matching length Lmax to the matching degree assigning unit 185. Here, there is a functional relationship in which if a set of matching parts is obtained for a certain word, the maximum matching length for the word is obtained.

(Configuration of the pass / fail determination unit 74)
FIG. 7 is a block diagram showing a functional configuration of the pass / fail determination unit 74 (see FIG. 1). With reference to FIG. 7, the pass / fail determination unit 74 includes a storage unit 260 for storing the evaluation result 72 of the fitness level output from the fitness level evaluation unit 70, and the evaluation result 72 of the fitness level in the storage unit 260. A word selection unit 270 for sequentially selecting words to be processed and sequentially outputting GWPP and maximum matching length Lmax of the word, and a threshold value according to the maximum matching length Lmax output from the word selection unit 270 In order to compare the threshold value setting unit 272 for selecting and outputting from the threshold value table 160 (see FIG. 5) with the threshold value output from the GWPP and the threshold value setting unit 272, and sequentially outputting the results. The comparison unit 274 is included. Further, the pass / fail determination unit 74 gives the comparison result regarding the words W _{1 to} W _N output by the comparison unit 274 to the evaluation result 72 of the fitness level in the storage unit 260 as the test result for the words W _{1 to} W _N. Thus, when the inspection result giving unit 264 generates the inspection result 76 and the inspection result adding unit 264 generates the inspection result 76, the output unit 268 reads out the inspection result 76 from the storage unit 260 and outputs it. Including.

(Configuration of threshold learning unit 84)
FIG. 8 is a block diagram showing a functional configuration of the threshold learning unit 84 shown in FIG. Referring to FIG. 8, threshold learning unit 84 assigns corresponding correct test results 40A,..., 40P to test result 76 given from switch 78 (see FIG. 1) to threshold learning unit 84. A correct answer assigning unit 280 for generating a learning test result 282 with correct answer, a learning test result storage unit 284 for storing the generated learning test result 282 with correct answer, and a learning test result storage A correct / incorrect determination unit 286 for determining the correctness / incorrectness of the test result of each word based on the correct answer test result from the learning test result 282 with correct answer stored in the unit 284, and for the correct / incorrect determination And a correct / incorrect storage 290 for storing the result 288.

  Further, when the correctness determination by the correctness determination unit 286 is completed, the threshold value learning unit 84 performs an inspection by the pass / fail determination unit 74 (see FIG. 1) based on the correctness determination result 288 stored in the correctness / incorrectness storage unit 290. Threshold values for adjusting the values of threshold values T0 to T4 in threshold value table 160 (see FIG. 5) using information in correctness / error storage unit 290 according to the result When the threshold value is adjusted, the adjustment unit 294 and the re-inspection unit 296 for re-inspecting the learning test result 282 in the learning test result storage unit 284 based on the adjusted threshold value. Including.

  The learning test results 282 stored in the learning test result storage unit 284 are prepared in advance corresponding to the test results 76 for the learning voices 38A,..., 38P and the learning voices 38A,. Of the correct answer results 40A,..., 40P, the correct answer results 40A,. The correct test result includes a correct answer corresponding to each test result of each word. The correctness determination unit 286 has a function of starting correctness determination in response to a change in information in the learning test result storage unit 284. The correct / wrong determination result 288 generated by this determination includes the GWPP sequence 108 (see FIG. 2) and the maximum match length Lmax sequence 140 (see FIG. 4) corresponding to the word sequence of the correct test result, and the test result sequence 144. And a correct / incorrect indicator string composed of correct / incorrect indicators indicating the correctness / incorrectness of the inspection result of each word. Each correct / incorrect mark indicates whether or not the test result for the corresponding word matches the correct answer, and if not, what the correct answer is. That is, a value indicating that the test result matches the correct answer, a value indicating that the word to be passed (OK) is rejected (NG), and a word to be rejected (NG) are passed. It is a value indicating that it is set to (OK). The threshold adjustment unit 294 has a function of evaluating the performance of the inspection by the pass / fail determination unit 74 (see FIG. 1) for each maximum match length Lmax using the reliability determination error rate (CER) shown below. Have.

しきい値調整部２９４はさらに、ＣＥＲがそれぞれ最小になるように、しきい値Ｔ0〜Ｔ4の各々を調整する機能を持つ。

The threshold adjustment unit 294 further has a function of adjusting each of the thresholds T0 to T4 so that the CER is minimized.

[Operation]
The speech translation system 30 according to the present embodiment operates as follows.

(Learning translation models and extracting phrase data)
Referring to FIG. 1, speech translation system 30 learns a translation model in advance before translation or threshold learning. That is, the translation model learning device 54 learns the translation model from the parallel translation stored in the bilingual corpus 34 and stores the obtained translation model in the translation model unit 52. In this learning process, the translation model learning device 54 generates source language phrases 120A, ..., 120Q and target language phrases 122A, ..., 122Q (see FIG. 3) from sentences included in the bilingual corpus 34. Each of the generated phrases is stored in the phrase database 56.

(Learning threshold)
Hereinafter, an operation in which the threshold learning unit 84 learns the threshold will be described. Referring to FIG. 1, when learning of a threshold is selected and an operation input 42 corresponding to the selection is given to the speech translation system 30, the selector 44 uses the learning speech 38A,. , 38P. The selected voice 46 is converted into a voice recognition result 50 by the voice recognition device 48. Further, the inspection device 60 performs an inspection on the voice recognition result 50, and an inspection result 76 (see FIG. 4B) is output from the inspection device 60. When learning the threshold value, the switch 78 outputs the inspection result 76 to the threshold value learning unit 84. Details of operations of the voice recognition device 48 and the inspection device 60 will be described later.

  Referring to FIG. 8, when test result 76 is given to correct answer giving unit 280 of threshold learning unit 84, correct answer giving unit 280 adds correct test results 40 </ b> A,..., 40 </ b> P to the given test result 76. A thing corresponding to this inspection result 76 is given. As a result, a learning test result 282 with a correct answer is generated. The generated learning test result 282 with correct answer is stored in the learning test result storage unit 284.

  When the learning test result 282 with the correct answer is stored in the learning test result storage unit 284, the correctness / incorrectness determination unit 286 compares the test result string 144 with the corresponding correct test result to determine the test result for each word. Correct / incorrect is determined and a correct / incorrect indicator string is generated. The correctness determination unit 286 forms a correctness / incorrectness determination result 288 from the correctness / incorrectness indicator string, the GWPP sequence 108 and the maximum match length Lmax sequence 140 in the learning test result 282, and stores the result in the correctness / incorrectness storage unit 290. .

  When the correctness / incorrectness determination unit 286 finishes the above processing for the learning test result 282 with correct answer in the learning test result storage unit 284, it gives an end signal to the threshold adjustment unit 294. The threshold adjustment unit 294 calculates the CER for each goodness of fit based on the correctness / incorrectness determination result 288 in the correctness / incorrectness storage unit 290, and sets the threshold so that the CER for each goodness of fit is minimized. The threshold values T0 to T4 in the table 160 are adjusted.

  When the threshold value in the threshold value table 160 is changed, the re-inspection unit 296 relates to the learning inspection result stored in the learning inspection result storage unit 284 based on the changed threshold value. A re-inspection is performed, and the inspection result stored in the learning inspection result storage unit 284 is changed. When the information in the learning test result storage unit 284 changes, the correctness / incorrectness determination unit 286 performs correctness / incorrectness determination again. If the above operation is repeated and the CER is minimized, the adjustment of the threshold value is finished. Through such a series of operations, the threshold values T0 to T4 in the threshold value table 160 become the threshold values with the least number of inspection errors when the corresponding words of the matching degree are inspected.

(voice recognition)
Referring to FIG. 1 again, when operation input 42 corresponding to speech translation for translation speech 32 is given to speech translation system 30, speech translation system 30 starts a series of operations for performing speech translation. In this case, the selector 44 selects the translation voice 32 as the voice 46 to be processed.

  When the voice 46 is given from the selector 44 to the voice recognition device 48, the voice recognition device 48 executes the series of processes described above with reference to FIG. That is, referring to FIG. 2, the speech recognition device 48 first performs speech recognition processing 100 of the given speech 46 to generate a word graph 102. The speech recognition device 48 further calculates a posterior probability for each word constituting the route network of the word graph 102, and further calculates a GWPP. Subsequently, the voice recognition device 48 performs route selection 104 to generate a word string 106 and also generates a GWPP sequence 108 corresponding to the word string 106. The speech recognition device 48 outputs a set of the generated word string 106 and the corresponding GWPP sequence 108 as a speech recognition result 50.

(Evaluation of conformity)
Hereinafter, an operation in which the fitness evaluation unit 70 (see FIG. 1) of the inspection apparatus 60 generates the fitness evaluation result 72 (see FIG. 4A) from the speech recognition result 50 will be described. Referring to FIG. 6, when the speech recognition result 50 is input to the fitness evaluation unit 70, the speech recognition result 50 is stored in the recognition result storage unit 180. The partial sentence generation unit 190 generates a partial sentence including a partial word string and its position indicator from the word string 106 (see FIG. 2) of the speech recognition result 50 in the recognition result storage unit 180, and stores the partial sentence. Stored in the unit 192. When all the partial sentences are stored in the partial sentence storage unit 192, the search unit 194 searches for a phrase that matches each partial sentence by the following process for each partial sentence. That is, the search unit 194 first selects a partial sentence to be processed, and searches the phrase database 56 for the same phrase as the partial sentence. If the same phrase exists, this partial word string becomes one of the matching parts. In this case, the search unit 194 stores the position indicator of the partial sentence to be processed in the matching part storage unit 196. If there is no same phrase, select another partial sentence. When the search by the search unit 194 is completed, the position of each matching part in the word string 106 can be specified in the matching part storage unit 196 by the stored position indicator.

When the position markers of all the matching parts are stored in the matching part storage unit 196, the Lmax calculation unit 184 performs the next processing on the words W _{1 to} W _N in the word string 106 in the speech recognition result 50 in the recognition result storage unit 180. Then, the degree of fitness of each word is determined. That is, first, the word W _n to be processed is selected, and the position of the word W _n is specified. All the position indicators including the position of the word W _n are selected in the matched portion storage unit 196. As a result, a set of matching portions substantially relating to the word W _n is created. Based on the selected position marker, the phrase length is obtained for all matching parts including the word W _n . That is, the phrase length is obtained for all the elements of the set of matching parts related to the word W _n . The maximum matching length Lmax is obtained by searching for the largest phrase length among the obtained phrase lengths, and is output to the fitness level assigning unit 185.

The fitness level of each word determined by the above processing is given to the fitness level assigning unit 185 shown in FIG. The fitness level assigning unit 185 gives the maximum matching length Lmax of the word as the fitness level to the words W _{1 to} W _N in the recognition result storage unit 180. As a result, the fitness evaluation result 72 shown in FIG. 4A is formed in the recognition result storage unit 180. When this process is completed for the words W _{1 to} W _N , the fitness level assigning unit 185 gives an end signal to the output unit 186. In response to this, the output unit 186 reads and outputs information in the recognition result storage unit 180. As a result, a fitness evaluation result 72 is output.

(Decision of pass / fail)
Below, the operation | movement which the pass / fail determination part 74 (refer FIG. 1) test | inspects each word based on the evaluation result 72 and the reliability of a suitability is demonstrated. With reference to FIG. 7, when the fitness evaluation result 72 is input to the pass / fail determination unit 74, the fitness evaluation result 72 is stored in the storage unit 260. Word selection unit 270, the words W ₁ to W-sequentially select the word W _n to be processed from among _N, fitness of the words W _n and a GWPP respectively in the evaluation result 72 fitness, threshold setting To the unit 272 and the comparison unit 274.

When the fitness level of word W _n is given, threshold value setting unit 272 reads a threshold value corresponding to the fitness level from threshold value table 160 (see FIG. 5) and provides it to comparison unit 274. For example, the maximum match length Lmax of the _nth word Wn shown in FIG. In the threshold table 160 shown in FIG. 5, the threshold for the word with Lmax = 1 is T1. Therefore, threshold value setting unit 272 provides threshold value T 1 to comparison unit 274. The comparison unit 274 compares the GWPP of the word W _n with the threshold value. _If the GWPP of the word Wn is greater than or equal to the threshold value T1, a first value representing pass (OK) is output. Otherwise, a second value representing failure (NG) is output.

  When the values output from the comparison unit 274 as described above are sequentially given to the inspection result assigning unit 264 shown in FIG. 7, the inspection result assigning unit 264 uses the inspection result sequence 144 (FIG. 4) based on the values. (See (B)) and assign it to the fitness evaluation result 72 in the storage unit 260. As a result, a test result 76 shown in FIG. When this processing is completed, the inspection result giving unit 264 gives an end signal 266 to the output unit 268. In response to this, the output unit 268 reads and outputs information in the storage unit 260. As a result, the inspection result 76 is output to the switch 78 shown in FIG. When the switch 78 receives an input of the inspection result 76, the switch 78 outputs the inspection result 76 for the output corresponding to the process selected at that time. In this case, the switch 78 outputs the inspection result 76 to the preprocessing device 64.

(Pre-processing and translation)
When the preprocessing device 64 shown in FIG. 1 receives the inspection result 76 from the switch 78, the preprocessing device 64 performs preprocessing based on the inspection result 76. In the present embodiment, words whose test results are unacceptable (NG) are removed from the word strings W _{1 to} W _N of the speech recognition results, and a speech recognition result for translation is generated. The translation device 68 generates the translation result 36 by statistical translation using the phrase-based translation model stored in the translation model unit 52.

[Experiment]
In order to evaluate the word inspection performance in the speech translation system 30 according to the present embodiment, a threshold learning experiment and a speech recognition result inspection experiment using a phrase table obtained by the learning experiment are performed. I did it.

  In the learning experiment and the examination experiment, a data set for the learning experiment and a data set for the examination experiment were created by using the travel conversation basic expression corpus (Basic Travel Expression Corpus: BTEC) created by the applicant. These are hereinafter referred to as “learning set” and “test set”, respectively. The learning set is composed of 1016 kinds of BTEC utterances, and the total number of words in the entire data set is 7215 words. The test set consists of 3060 utterances of BTEC, and the total number of words in the entire data set is 21005 words.

  In this experiment, the speech recognition device 48 developed by the applicant was used. In this experiment, the phrase database 56 was also prepared in advance by the applicant. The phrases stored in the phrase database 56 are different and have about 800,000 phrases. The following table 1 shows a breakdown of the different numbers for each phrase length L.

なお、このフレーズデータベース５６において、１フレーズあたりの平均単語数は３．４８単語であった。

In this phrase database 56, the average number of words per phrase was 3.48 words.

  In the learning experiment, speech recognition was first performed for each utterance included in the learning set, Lmax was calculated for each word for each speech recognition result, and a correct answer of the test result was given. Then, the threshold value is learned for each Lmax so that the CER of the inspection result for the correct answer is minimized. The threshold value table obtained as a result of this learning experiment is shown in the following table 2.

テーブル２を参照して、Ｌmaxが大きな値になるにしたがい、しきい値が低くなった。特にＬmax＝０の単語に対するしきい値は、Ｌmax≧１のいずれの単語に対するしきい値よりもはるかに高い値になった。これにより、本実施の形態の検査装置による検査では、検査結果に次のような傾向が生じることが分かる。すなわち、一致するフレーズが長くなるにしたがいＧＷＰＰが低い単語まで合格とする傾向、及び一致するフレーズがない単語については非常に高いＧＷＰＰのもののみを許容する傾向である。これらの傾向は、Ｌmaxが大きな単語ほど翻訳に適しているという当初の予測に合致する。

Referring to Table 2, the threshold value decreased as Lmax increased. In particular, the threshold for a word with Lmax = 0 was much higher than the threshold for any word with Lmax ≧ 1. Thereby, it can be seen that the following tendency occurs in the inspection result in the inspection by the inspection apparatus of the present embodiment. That is, as the matching phrase becomes longer, it tends to pass to a word having a low GWPP, and the word having no matching phrase tends to allow only a very high GWPP. These tendencies are consistent with the initial prediction that words with larger Lmax are better suited for translation.

  In the test experiment, speech recognition was performed for each utterance included in the test set, and a correct answer of the test result was assigned to each word in advance for each speech recognition result. In this inspection experiment, the following two kinds of methods were used for comparison. That is, one is the inspection method of the present embodiment, which is an inspection using the threshold value table shown in Table 2. The other is a test using a threshold common to all words regardless of Lmax. Hereinafter, these inspection methods are simply referred to as “inspection 1” and “inspection 2”, respectively. In examination 2, the common threshold for all words was set to 0.62. This value is a threshold value that is fixed so that CER is minimized regardless of Lmax. CER was calculated | required according to Lmax about the test result by two types of test methods, respectively. The CER of the entire test set was also obtained. The ratio of words in the test set and the CER for each of the inspection results obtained by the two types of inspection methods are shown for each Lmax in Table 3 below.

テーブル３を参照して、Ｌmax≧１の場合、検査１及び検査２の両方において、Ｌmaxが大きな値になるにしたがいＣＥＲが低くなった。また、検査１及び検査２の両方において、Ｌmax＝０の単語についてのＣＥＲは、Ｌmax＝１又は２の単語についてのＣＥＲより低くなった。

Referring to Table 3, in the case of Lmax ≧ 1, the CER decreased as the Lmax increased in both the inspection 1 and the inspection 2. In both Exam 1 and Exam 2, the CER for the word with Lmax = 0 was lower than the CER for the word with Lmax = 1 or 2.

  Comparing the CER for the examination results of examination 1 and examination 2, for words with Lmax> 3, Lmax = 1, and Lmax = 0, the CER for the examination result of examination 1 is lower than the CER for the result of examination 2 did. Comparing the CER for the entire test set, the CER for Exam 1 and Exam 2 was 4.3 (%) and 4.5 (%), respectively. Therefore, in the examination using GWPP as an evaluation scale, it has been clarified that the optimum threshold value as a criterion for pass / fail of each word changes according to Lmax. Furthermore, by using a threshold value optimized by learning as a criterion for pass / fail, the inspection performance can be improved in both words having a large Lmax and words having a small Lmax, and the entire word to be examined. Became clear. By inspecting the speech recognition results with the improved performance inspection method and pre-processing in this way, the linkage between speech recognition and statistical translation is strengthened, improving the overall performance of speech translation processing. Can be expected.

[Realization by computer]
The speech translation system 30 according to the present embodiment is realized by computer hardware, a program executed by the computer hardware, and data stored in the computer hardware. FIG. 9 shows the external appearance of the computer system 330, and FIG. 10 shows the internal configuration of the computer system 330.

  With reference to FIG. 9, the computer system 330 includes a computer 340, a monitor 342, a keyboard 346, a mouse 348, a microphone 370, and a speaker 372. The computer 340 includes a CD-ROM (Compact Disc Read Only Memory) drive 350 and an FD (Flexible Disc) drive 352.

  10, in addition to the CD-ROM drive 350 and the FD drive 352, the computer 340 includes a hard disk 354, a CPU (Central Processing Unit) 356, an FD drive 352, a CD-ROM drive 350, a hard disk 354, And a bus 366 connected to the CPU 356, a read-only memory (ROM) 358 connected to the bus 366 and storing a boot-up program and the like, and a bus 366 storing program instructions, system programs, work data and the like. Random access memory (RAM) 360. Computer 340 further includes a sound board 368 connected to bus 366, microphone 370, and speaker 372. Although not shown here, the computer 340 may further include a network adapter board that provides a connection to a local area network (LAN).

  A program for causing the computer system 330 to operate as the speech translation system 30 is stored in the CD-ROM 362 or FD 364 inserted into the CD-ROM drive 350 or FD drive 352 and further transferred to the hard disk 354. Alternatively, the program may be transmitted to the computer 340 through a network (not shown) and stored in the hard disk 354. The program is loaded into the RAM 360 when executed. The program may be loaded into the RAM 360 directly from the CD-ROM 362, from the FD 364, or via a network. The bilingual corpus 34 (see FIG. 1) is stored in, for example, the hard disk 354, and necessary parts thereof are appropriately read into the RAM 360 when the translation model unit 52 learns. The translation model obtained by learning and the phrases extracted in the process are also stored in the hard disk 354, for example, and necessary portions are appropriately read into the RAM 360 as necessary.

  The program includes a plurality of instructions that cause the computer 340 to operate as the speech translation system 30 of the present embodiment. Some of the basic functions necessary to perform this operation are provided by operating system (OS) or third party programs running on the computer 340 or various toolkit modules installed on the computer 340. Therefore, this program does not necessarily include all functions necessary for realizing the system and method of the present embodiment. This program includes only instructions for executing the operation as the speech translation system 30 described above by calling an appropriate function or “tool” in a controlled manner so as to obtain a desired result. It only has to be. The operation of computer system 330 is well known and will not be repeated here.

  In the above embodiment, GWPP is used as the reliability for each word. However, the reliability that can be used for inspection by the inspection apparatus is not limited to GWPP. If each word in the word string of the speech recognition result is given a reliability with a predetermined evaluation scale, each word can be inspected using the given reliability. However, in that case, in order to perform the inspection using the assigned reliability, it is necessary to perform threshold value learning in advance for each Lmax. Alternatively, it is necessary to set a threshold value table determined for each Lmax by some learning as a threshold value table.

  The phrase database stores pairs of phrases in the source language and the target language. However, the present invention is not limited to such an embodiment. The inspection device 60 evaluates the suitability using the source language phrase. Therefore, a database in which only the source language phrases are stored in advance may be prepared, and the suitability may be evaluated using the database.

  In the present embodiment, partial sentences are created with all combinations from the speech recognition result as the target of collation with the phrase. However, the present invention is not limited to such an embodiment. For example, if probability information such as the appearance frequency of each phrase can be obtained in the phrase database, segmentation in units of phrases may be performed on the speech recognition result based on the probability information. By doing so, it is possible to efficiently obtain Lmax while avoiding phrase overlap. In this embodiment, Lmax is used as the fitness level. However, the fitness level may be obtained in consideration of the probability information.

In the embodiment described above, a set of matching parts is obtained for each word of the speech recognition result, and the fitness is determined based on the maximum phrase length (maximum matching length L _max ) of the matching parts included in the set. ing. However, the method for determining the degree of matching between each word of the speech recognition result and the subsequent phrase-based natural language processing is not limited to this. When a set of phrases is given, if a set of partial words that include a word with a speech recognition result and is included in the set of phrases is determined, the fitness is determined as a function of the set. Can do.

  In the above embodiment, this function is “the phrase length having the largest phrase length among the matching portions constituting the set”. In addition, the fitness can be calculated based on, for example, the number of elements in the set of partial word strings, the average phrase length, the sum of the phrase lengths of each element, and the like.

  In the above-described embodiment, the preprocessing device 64 generates a word string for translation by excluding rejected words from the word string of the speech recognition result based on the inspection result. However, the pre-processing using the inspection result by the inspection apparatus 60 of the present embodiment is not limited to such a process. For example, the speech recognition result itself may be rejected if there is even one word whose test result fails. In this case, it is convenient to issue a predetermined error signal in response to rejection of the speech recognition result. Further, for example, the speech recognition device outputs the N-best speech recognition result, and the pre-processing device selects the optimum speech recognition result for translation based on the inspection result for each of the N-best speech recognition results. It may be. Conversely, it is also possible to evaluate whether the translation model used for translation and the domain of the bilingual corpus used for learning thereof match the language and domain of the speech to be recognized. Furthermore, by selecting an appropriate bilingual corpus and translation model based on the evaluation, it can be expected to improve speech translation performance for speech to be recognized.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim in the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

1 is a block diagram showing the overall configuration of a speech translation system 30. FIG. It is a schematic diagram which shows the outline | summary of the process by the speech recognition apparatus 48, and the data structure of the speech recognition result 50. FIG. 3 is a schematic diagram illustrating a configuration of a phrase database 56. FIG. It is a figure which shows the data structure of the speech recognition result 72 with the evaluation result of a fitness, and the test result 76. FIG. It is a figure which shows the data structure of the threshold value table. 3 is a block diagram illustrating a functional configuration of a fitness evaluation unit 70. FIG. 3 is a block diagram illustrating a functional configuration of a pass / fail determination unit 74. FIG. 3 is a block diagram showing a functional configuration of a threshold learning unit 84. FIG. 1 is an external view of a computer system that implements a speech translation system 30 according to an embodiment of the present invention. FIG. 10 is a block diagram of the computer shown in FIG. 9.

Explanation of symbols

DESCRIPTION OF SYMBOLS 30 Speech translation system 34 Bilingual corpus 44 Selector 48 Speech recognition apparatus 52 Translation model part 54 Translation model learning apparatus 56 Phrase database 60 Inspection apparatus 64 Preprocessing apparatus 68 Translation apparatus 70 Conformity evaluation part 74 Pass / fail decision part 78 Switch 82 Threshold value Table section 84 Threshold learning section 160 Threshold table 180 Recognition result storage section 182 Collation section 184 Lmax calculation section 185 Conformity assignment section 186, 268 Output section 190 Partial sentence generation section 192 Partial sentence storage section 194 Search section 196 Match Partial storage unit 260 Storage unit 264 Test result assignment unit 270 Word selection unit 272 Threshold setting unit 274 Comparison unit 280 Correct answer assignment unit 284 Learning test result storage unit 286 Correct / incorrect determination unit 290 Correct / incorrect storage unit 294 Threshold adjustment unit 296 Re-inspection department

Claims (4)

A word sequence of a speech recognition result generated from a predetermined input speech by speech recognition processing, and a word sequence of the speech recognition result as a target of a predetermined phrase-based statistical natural language processing subsequent to the speech recognition processing A speech recognition result inspection device for inspecting whether or not the words constituting the word should be accepted,
The speech recognition result inspection device is used together with a set of phrases extracted by a predetermined extraction method from a corpus for statistical model learning used in the natural language processing,
Each word constituting the word sequence of the speech recognition result is given a reliability in advance by the speech recognition process,
The voice recognition result inspection apparatus comprises:
For each word constituting the word string of the speech recognition result, a word having a phrase that matches the phrase set in the word string that includes the word and forms a partial word string of the voice recognition result A fitness level giving means for giving a fitness level to the natural language processing as a function of a set of columns;
For each word constituting the word string of the speech recognition result, a comparison is made between the threshold value determined according to the fitness level given to the word by the fitness level giving means and the reliability level given to the word Accordingly, it viewed including a determination means for determining whether or not to accept the word,
The fitness level giving means is
Collating means for detecting a word string having a matching phrase in the phrase set by comparing the phrase set and a word string forming a partial word string of the speech recognition result;
For each word constituting the word string of the speech recognition result, a phrase of the word string included in the set based on a set of word strings including the word among the word strings detected by the matching unit A speech recognition result inspection apparatus comprising: a maximum value, an average value, or a sum of lengths, or means for giving the fitness based on the number of elements included in the set . The natural language process includes a phrase-based statistical translation process between the language of the input speech and a predetermined target language,
The set of phrases includes a set of phrases extracted in a process of learning a translation model for the statistical translation process from a bilingual corpus of a language of the input speech and the predetermined target language. Inspection device for the voice recognition result described. The determining means includes
Means for associating and holding the fitness and the threshold;
For each word constituting the word string of the speech recognition result, based on the fitness given to the word, the fitness for the word according to the fitness and the threshold held by the means for holding Means for setting a threshold;
Whether to accept the word by comparing the threshold value set by the setting means and the reliability assigned to the word for each word constituting the word string of the speech recognition result The voice recognition result inspection apparatus according to claim 1, further comprising: a comparison unit for determining whether or not . A computer program that, when executed by a computer, causes the computer to operate as the speech recognition result inspection apparatus according to any one of claims 1 to 3 .

Images