JP2008083165A - Voice recognition processing program and voice recognition processing method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide technology for reducing misrecognitions in keyword recognition. <P>SOLUTION: In keywords defined as words to be recognized, the words which are not similar to each other are stored in a storage device. When the voice input is the keywords which are not similar to each other, processings defined by the keyword are performed. When the voice input are keywords which are similar to each other, information for confirming the keyword is output to an output device. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to speech recognition.

  In recent years, interfaces using voice recognition technology are becoming popular in information devices such as mobile phones and car navigation systems. Speech recognition technology includes keyword recognition technology. A keyword, which is a predetermined word, is extracted from the input voice, and processing is performed according to the extracted keyword. The keyword recognition technique is used for extracting a user's intention from a free utterance and creating an index for voice search.

  A technique for realizing this keyword recognition is described in Patent Document 1. In Patent Document 1, a keyword model generated for receiving a predetermined keyword and a garbage model for receiving a phrase other than the keyword are prepared, and obtained from the result of collation with the input speech. It describes that keywords are extracted by comparing scores.

Japanese Laid-Open Patent Publication No. 2005-92310

  At present, the recognition accuracy of keyword recognition is not yet sufficient. Keyword misrecognition can be divided into two categories. One is an error in which a voice uttered as the keyword A is erroneously recognized as the keyword B. The other is an error in which input speech that is not a keyword is accepted as a keyword.

  The former error is a part of the dictionary entry that is registered by normal speech recognition, such as extracting only “Airport” and “High School” registered as keywords for utterances of “Haneda Airport” and “Setagaya High School”, for example. Only the recognition target. For this reason, similar phrases are likely to be the target of keywords. Also, as the number of keyword entries increases, the number of similar words increases, and the error frequency increases.

  The latter error is, for example, recognizing a voice of non-keyword as a keyword, such as recognizing “Chofu Airport” as “Chofu High School” registered as a keyword. In order to prevent this error, Patent Literature 1 generates similar words. However, it cannot be solved when similar words in the keyword are included. In addition, since a word outside the keyword is added, there is a risk that the utterance of the keyword is erroneously recognized as a word outside the keyword.

  The present invention has been made in view of the above-described problems, and an object thereof is to reduce erroneous recognition in keyword recognition.

  The present invention has been made to achieve the above-mentioned object, and among the keywords defined as vocabularies to be recognized, those that are not similar to each other are stored in a storage device, and the input voices are not similar to each other. If it is a keyword, the first process is executed, and if the input speech is a keyword similar to each other, the second process is executed.

  According to the present invention, in a computer-executable speech recognition program for recognizing a keyword that is a word included in input speech, a plurality of keywords are not included in a keyword that is phonologically similar to each other. Storage means for classifying and storing the keyword group and a second keyword group that is a keyword not included in the first keyword group, and the computer includes at least one of the at least one included in the input voice A keyword recognizing step for recognizing a keyword; a first process executing step for executing a first process when the recognized keyword is included in the first keyword group; and A second process execution step for executing the second process is included in the keyword group. And butterflies.

  According to the technique of the present invention, it is possible to reduce erroneous recognition in keyword recognition.

  Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

<First Embodiment>
First, the first embodiment will be described.

  FIG. 1 is a configuration example of the speech recognition processing device 1. The voice recognition processing device 1 realizes a voice interface using the keyword recognition of the first embodiment.

  The speech recognition processing device 1 is an information processing device such as a navigation system, a portable terminal, or a CTI (Computer Telephony Integration) system. The speech recognition processing device 1 includes a CPU (Central Processing Unit) 101, a memory 102, a secondary storage device 103, an input device 104, an output device 105, a communication interface 106, and the like. The CPU 101, the memory 102, the secondary storage device 103, the input device 104, the output device 105, the communication interface 106, and the like are connected by a bus 107.

  The secondary storage device 103 stores a keyword model 161, a confusion matrix 164, a garbage model 165, and the like.

  The keyword model 161 is an acoustic model sequence of the keyword speech to be extracted. As a feature of the present embodiment, the keyword model 161 is divided into two, a keyword class A162 and a keyword class B163. The keyword class A162 includes keywords that are not similar to each other. The keyword class B163 includes keywords that are not included in the keyword class A162 among the keywords in the keyword model 161. Therefore, keywords similar to each other are included in the keyword class B163.

  The data structure of keyword class A162 and keyword class B163 is arbitrary. For example, the keyword class A 162 and the keyword class B 163 may be stored in different tables or databases, and the keyword class A 162 and the keyword class B 163 are stored in the same table or database and assigned to each keyword. It may be classified into keyword class A162 and keyword class B163 by a flag or the like.

  The confusion matrix 164 stores, for each of the plurality of sounds, a degree of similarity that is a degree recognized as matching the same sound and another sound.

  The garbage model 165 stores an acoustic model sequence of speech other than keywords. The acoustic model sequence in the garbage model 165 is the same as that used in general speech recognition.

  The CPU 101 implements a class classification unit 151, an analysis unit 152, a collation unit 153, a determination unit 154, a process execution unit 155, a process execution unit A156, a process execution unit B157, and the like by executing a program (not shown).

  The class classification unit 151 generates a keyword class A 162 and a keyword class B 163 from the confusion matrix 164. The analysis unit 152 converts the input speech waveform into feature parameters. The collation unit 153 collates the feature parameter series of the input speech converted by the analysis unit 152 with the keyword model 161 and the garbage model 165, and obtains a model having the maximum score. The collation performed by the collation unit 153 is the same as that used in general speech recognition. The determination unit 154 determines a keyword included in the input voice based on the result obtained by the collation unit 153. In addition, the determination unit 154 determines whether the keyword is a keyword corresponding to either the keyword class A 162 or the keyword class B 163.

  The process execution unit 155 executes a process based on the keyword detected by the determination unit 154. The process execution unit 155 includes a process execution unit A156, a process execution unit B157, and the like. If the keyword belongs to keyword class A162, process execution unit A156 executes the process. Specifically, the process execution unit A156 executes a process according to the recognized keyword. If the keyword belongs to the keyword class B163, the process execution unit B157 executes the process. Specifically, the process execution unit B157 outputs information for confirming the keyword.

  The input device 104 is, for example, a microphone, a keyboard, a mouse, a scanner, or the like. The output device 105 is, for example, a display, a speaker, a printer, or the like. The speech recognition processing device 1 is connected to another communication terminal (not shown) or the like via the communication interface 106 and a communication network (not shown).

  Next, information in the secondary storage device 103 will be described.

  First, the keyword model 161 will be described.

  The keyword model 161 is an acoustic model sequence generated from one or more keywords registered in advance. The keyword used here is a phrase that is registered in association with processing in the processing execution unit 155 described later, such as a search query or a command. Although the acoustic model is not particularly limited, for example, a conventional HMM (Hidden Markov Model) may be used. The HMM is a probability model composed of an internal state that transitions according to a Markov model and an appearance probability distribution of an observation signal in the internal state.

  Here, an example of the acoustic model will be specifically described with reference to FIG.

  FIG. 2 is an example of an acoustic model series for the keyword “station”. In this example, a tri-state triphone HMM is used as an acoustic model. In FIG. 2, an HMM model 201 is a triphone “* / e / k” model. The HMM model 202 is a triphone “e / k / i” model. The HMM model 203 is a triphone “k / i / *” model. By connecting these three HMM models 201 to 203, an acoustic model having the keyword “station” is formed.

  As described above, in the present embodiment, the keywords registered in the model are classified into the keyword class A 162 and the keyword class B 163. Only keywords that are not acoustically similar are registered in the keyword class A162, and keywords that are not registered in the keyword class A162 are registered in the keyword class B163.

  Next, the confusion matrix 164 will be described with reference to FIGS.

  FIG. 3 is an example of the confusion matrix 164 in units of words. In FIG. 3, the confusion matrix 164 includes a vertical axis 311, a horizontal axis 312, and a matrix unit 313. The vertical axis 311 and the horizontal axis 312 are utterances of each keyword. The matrix unit 313 is the probability of recognizing the utterance keyword on the vertical axis 311 as the utterance keyword on the horizontal axis 312.

  Here, an example in the case of the utterance of the keyword “station” will be described. In the confusion matrix 164 of FIG. 3, the probability that the horizontal axis 312 “station” is correctly recognized for the utterance of the vertical axis 311 “station” is the matrix portion 313 “98.2” percent. The probability that the horizontal axis 312 “high school” is erroneously recognized is the matrix portion 313 “0.1” percent.

  An example in the case of the utterance of the keyword “high school” will be described. In the confusion matrix 164 of FIG. 3, the probability that the horizontal axis 312 “high school” is correctly recognized with respect to the utterance of the vertical axis 311 “high school” is the matrix portion 313 “90.1” percent. The probability that the horizontal axis 312 “airport” is erroneously recognized is the matrix portion 313 “5.1” percent.

  Next, an example of another confusion matrix 164 will be described.

  FIG. 4 shows an example of a phoneme unit confusion matrix. In FIG. 4, the confusion matrix 164 includes a vertical axis 411, a horizontal axis 412, and a matrix unit 413. A vertical axis 411 and a horizontal axis 412 are utterances of each phoneme. The matrix unit 413 is the probability of recognizing the phoneme of the utterance on the vertical axis 411 as the phoneme of the utterance on the horizontal axis 412.

  Next, the garbage model 165 will be described.

  The garbage model 165 is an acoustic model of speech other than the keywords registered in the keyword model 161. The acoustic model is not particularly limited, but, for example, a conventional HMM may be used as in the keyword model 161 described above.

  Here, an example of the acoustic model will be specifically described with reference to FIG.

  In FIG. 5, an HMM model 501 is a model whose central phoneme is “a”. The HMM model 502 is a model whose central phoneme is “i”. The HMM model 503 is a model whose central phoneme is “N”. In this way, by creating a loop (HMM model 506) from the end node (HMM model 505) to the start end (HMM model 504) for an HMM model in which models corresponding to all phonemes are arranged in parallel, all phonemes are created. A series combination is converted into an HMM model.

  Next, an operation example of the speech recognition processing apparatus 1 will be described with reference to FIG.

  The voice data is input to the voice recognition processing apparatus 1 from the input device 104 such as a microphone and the communication interface 106. When receiving the input of voice data (S601), the class classification unit 151 of the voice recognition processing device 1 classifies the keywords of the keyword model 161 into the keyword class A162 and the keyword class B163 (S602). For this purpose, for example, the class classification unit 151 selects one keyword in the keyword model 161, sets a keyword class A162 that has a rate that the keyword is erroneously recognized as another keyword to be less than a predetermined threshold, Those that are equal to or greater than the threshold are defined as keyword class B163.

  A specific example will be described by taking the case of FIG. 3 as an example. Here, an example in which the threshold value of the criterion is 4% will be described.

  For example, the case of the keyword “station” will be described. In the case of the confusion matrix 164 in FIG. 3, the probability of misrecognizing the keyword “station” as another word is 4% or less. Accordingly, the class classification unit 151 classifies the keyword “station” into the keyword class A162.

  An example of the keyword “high school” will be described. In the case of the confusion matrix 164 in FIG. 3, the possibility that the keyword “high school” is erroneously recognized as the keyword “airport” is 5.1%. Accordingly, the class classification unit 151 classifies the keyword “high school” into the keyword class B163.

  In the case of the confusion matrix 164 shown in FIG. 4, the class classification unit 151 performs the same processing as in the case of FIG. 3 described above for each phoneme constituting the keyword, and thus detailed description thereof is omitted.

  In FIG. 6, the analysis unit 152 converts the input speech waveform into a feature parameter (S603). The conversion process by the analysis unit 152 is the same as the conventional speech recognition. That is, the feature parameter is a multidimensional vector quantity obtained by dividing the audio signal every short period (several tens of ms) and converting the signal in the section into MFCC (Mel frequency cepstrum coefficient) or the like. Therefore, the analysis unit 152 acquires time-series data of feature vectors represented as multidimensional vectors.

  The collation unit 153 collates the feature parameter series of the input speech converted by the analysis unit 152 with the keyword model 161 and the garbage model 165, and obtains a model having the maximum score (S604). The collation process performed by the collation unit 153 is the same as the conventional voice recognition. For example, the matching unit 153 calculates the similarity between the feature parameter series and the model in the keyword model 161 as a score. In addition, the matching unit 153 calculates the similarity between the feature parameter series and the model in the garbage model 165 as a score, and selects one having a high score. In general, in the collation including the garbage model, it is common to adjust the score balance between the keyword model and the garbage model by penalizing the score of the garbage model.

  Based on the result obtained by the collation unit 153, the determination unit 154 determines whether or not the model having the maximum score is the keyword class A162 (S605).

  As a result of the determination in S605, when the model having the maximum score is the model of the keyword in the keyword class A162, the determination unit 154 instructs the process execution unit 155 to execute the process determined by the keyword. The process execution unit 155 executes a process based on the keyword (S606). Specifically, for example, the process execution unit A156 of the process execution unit 155 executes a process corresponding to the recognized keyword. In this process, for example, a table (not shown) in which the recognized keyword is associated with the process executed when the keyword is recognized or an application or the like is stored in advance in the stored speech recognition processing apparatus 102. The process execution unit 155 may determine a process to be executed or an application to be executed with reference to the table.

  Specifically, for example, when the keyword “mail” is recognized, the process execution unit A156 executes the “mail application” associated with the keyword “mail” in advance. In addition, the nationwide facility name search task requires a large amount of data to be searched with limited computational resources, so it is often possible to improve processing efficiency by narrowing down the search target by classifying the facility name into categories. . In this case, when the keyword “high school” is recognized, the search target category is narrowed down to “high school”.

  On the other hand, as a result of the determination in S605, if the model having the maximum score is not the keyword model in the keyword class A162, the determination unit 154 determines whether the model having the maximum score is the keyword model in the keyword class B163. (S607).

  As a result of the determination in S608, when the model having the maximum score is not a keyword model in the keyword class B163, the determination unit 154 ends the process.

  As a result of the determination in S608, when the model having the maximum score is a keyword model in the keyword class B163, the determination unit 154 instructs the process execution unit 155 to execute the process. The process execution unit 155 confirms whether the recognized keyword is correct (S608). For this purpose, the process execution unit B157 of the process execution unit 155 outputs the keyword of the model having the maximum score to the communication terminal or the like from the output device 105 such as a display or a speaker, or the communication interface 106, and confirms with the user. Request. Using the input device 104 or an input device (not shown) of the communication terminal, the user inputs information indicating whether or not the keyword is correct to the voice recognition processing device 1.

  Here, an example of a screen displayed on the display to confirm the correctness of the recognized keyword will be described with reference to FIG. In FIG. 7, a screen 701 is an example displayed based on information output from the speech recognition processing device 1 in order to confirm a recognized keyword. The user uses the input device to check one of the radio button 711 and the radio button 712 to indicate whether the displayed keyword is correct. In the example of the screen 701, when the displayed keyword is not correct, the user inputs the correct keyword in the area 713. When the user presses the button 714, information indicating whether the keyword is correct, correct keywords, and the like are input to the speech recognition processing apparatus 1.

  Note that the process execution unit B157 may present a plurality of keywords similar to the recognized keyword and allow the user to select, or may calculate a reliability score for the recognition result. In order to present a plurality of keywords similar to the recognized keyword, the process execution unit B157 selects a word having a probability that it is erroneously recognized as the recognized keyword from the confusion matrix 164, and recognizes this word. It may be output together with the keyword. Further, in order to calculate a score for the recognition result, a reliability score is calculated from the score for each model calculated by the above-described processing of the matching unit 153 and an arbitrary mathematical expression. The process execution unit B157 may select an arbitrary number of keywords from the top of the reliability score, and may output the reliability score and the keyword.

  Returning to FIG. 6, when the process execution unit B157 acquires a correct keyword, the process execution unit B157 performs a process using the keyword (S609). This process may be arbitrary, and may be a process determined according to the recognized keyword as in S606 described above, or may be a process designated by the user at the time of the keyword confirmation described above.

  The process execution unit B157 may perform a process corresponding to the recognized keyword class instead of performing a process corresponding to the recognized keyword. A specific example will be described in the case of the search task described above. In the search task described above, in the process corresponding to the keyword, the search target category is narrowed down based on the recognized keyword. For example, when the keyword “high school” is recognized, the search category is narrowed down to “high school”. At this time, if the recognized keyword “high school” is mistakenly recognized as “airport”, the search category narrowing down fails. Therefore, when the recognized keyword belongs to the keyword class B163, the search category based on the recognized keyword is not narrowed down, and the search category is narrowed down by the logical sum (or) of all the keywords belonging to the keyword class B163. If there are too many search targets due to too narrow restrictions, narrow down the search target again by asking questions other than the genre set by the keyword, such as “Please specify prefectures”. Is also possible.

As described above, since there are no entries similar to each other for the keywords belonging to the keyword class A162, the possibility of recognizing an incorrect keyword is low. Therefore, even if the recognition result is trusted and the process is executed without confirmation to the user, there is little risk that the process contrary to the user's intention is executed. On the other hand, for keywords belonging to the keyword class B163, there may be entries similar to each other. Therefore, the process execution unit B157 is contrary to the user's intention even if there is an error in keyword recognition by checking the recognition result or performing a search in all keyword genres regardless of the recognized keyword. The process is not executed.
<Second Embodiment>
Next, a second embodiment will be described. The second embodiment is different from the first embodiment described above in that a subset dictionary including a predetermined keyword is created and a vocabulary including a recognized keyword is selected from the subset dictionary. Here, a word including one or more words is called a vocabulary.

  Since the second embodiment described below is partially the same as the first embodiment described above, the same reference numerals are given to the same components, the description thereof is omitted, and only different components will be described in detail. .

  A configuration example of the speech recognition processing device 801 will be described with reference to FIG.

  The speech recognition processing device 801 according to the second embodiment includes a speech re-recognition unit 811 instead of the process execution unit 155 of the speech recognition processing device 1 according to the first embodiment. Further, it further includes a main dictionary 851, a subset dictionary A852, a subset dictionary B853, a scale model 854, and the like.

  The main dictionary 851 stores the recognition target speech of the speech recognition processing device 801. The subset dictionary A852 and the subset dictionary B853 are generated by processing to be described later. The scale model 854 stores an acoustic model of the vocabulary in the main dictionary 851.

  The speech re-recognition unit 811 performs speech recognition on the feature vector series generated by the analysis unit 152 based on the keywords recognized by the analysis unit 152, the matching unit 153, and the determination unit 154. The voice re-recognition unit 811 includes a re-recognition unit A812, a re-recognition unit B813, and the like.

  When the recognized keyword is a keyword of a model belonging to the keyword class A162, the re-recognition unit A812 recognizes a word including the keyword. The re-recognition unit A812 includes a dictionary generation unit 821, a collation unit 822, a determination unit 823, and the like. The dictionary generation unit 821 generates a subset to be stored in the subset dictionary A852 from the keywords recognized by the analysis unit 152, the collation unit 153, and the determination unit 154 and the main dictionary 851. The matching unit 822 uses the subset dictionary A852 and the scale model 854 to collate with the input feature vector series and output a score. The determination unit 823 obtains the recognition result of the input speech by searching for a hypothesis that maximizes the matching score.

  When the recognized keyword is a model keyword belonging to the keyword class B163, the re-recognition unit B813 recognizes a vocabulary including the keyword. The re-recognition unit B 813 includes a dictionary generation unit 831, a collation unit 832, a determination unit 833, and the like. The dictionary generation unit 831 generates a subset to be stored in the subset dictionary B853 from the keyword class B163 and the main dictionary 851. Using the subset dictionary B853 and the scale model 854, the collation unit 832 collates the input feature vector series and outputs a score. The determination unit 833 obtains a recognition result of the input speech by searching for a hypothesis that maximizes the matching score.

  Next, an operation example will be described with reference to FIG.

  The voice data is input to the voice recognition processing apparatus 1 from the input device 104 such as a microphone and the communication interface 106. When receiving the input of voice data (S901), the class classification unit 151 of the voice recognition processing device 801 classifies the keywords of the model 161 into the keyword class A162 and the keyword class B163 (S902). This process is the same as S601 and S602 described above.

  When receiving the input of voice data, the analysis unit 152 of the voice recognition processing device 801 converts the input voice waveform into a feature parameter (S903). The conversion process by the analysis unit 152 is the same as S603 described above.

  The collation unit 153 collates the feature parameter series of the input speech converted by the analysis unit 152 with the keyword model 161 and the garbage model 165, and obtains a model having the maximum score (S904). This process is the same as S604 described above.

  Based on the result obtained by the collation unit 153, the determination unit 154 determines whether or not the model with the maximum score is the keyword class A162 (S905). This process is the same as S605 described above.

  If the model having the maximum score is the model of the keyword in the keyword class A162 as a result of the determination in S905, the determination unit 154 instructs the re-recognition unit A812 to execute the process. The dictionary generation unit 821 of the re-recognition unit A812 generates a subset dictionary A852 from the recognized keyword and the main dictionary 851 (S906). Specifically, for example, the dictionary generation unit 821 extracts a vocabulary including the recognized keyword from the main dictionary 851 and stores it in the subset dictionary A852.

  Here, an example of the processing of S906 will be specifically described with reference to FIG.

  In FIG. 10, the main dictionary 851 includes “Atsugi City Hall”, “Atsugi High School”, “Inokashira Park”, “Sakai Park”, “Kyoto University”, “Kokubunji Station”, “Kokubunji City Hall”, “Kusatsu Onsen”, Fourteen vocabularies are registered: “Setagaya Park”, “Setagaya Art Museum”, “Setagaya High School”, “Tokyo Station”, “Shinagawa Station”, and “Haneda Airport”. Here, an example in which “park” is recognized as a keyword will be described.

  The dictionary generation unit 821 extracts the words “Inokashira Park” and “Nagi Park” including “park” from the words registered in the main dictionary 851, and stores them as a subset dictionary A852. At this time, the dictionary generation unit 821 may select not only the notation “park” but also a vocabulary including the phoneme string “Kouen” (for example, “performance”, “support”, “lecture”, etc.).

  In FIG. 9, the matching unit 822 uses the subset dictionary A852 and the scale model 854 to collate with the input feature vector series and calculate a score (S907). This process is the same as S604 described above.

  The determination unit 823 determines a vocabulary that maximizes the matching score (S908). This process is the same as S604 described above.

  On the other hand, as a result of the determination in S905, if the model having the maximum score is not the model of the keyword in the keyword class A162, the determination unit 154 determines whether or not the model having the maximum score is the keyword class B163 (S909). ). This process is the same as S608 described above.

  As a result of the determination in S909, if the model having the maximum score is not the model of the keyword in the keyword class B163, the process ends.

  As a result of the determination in S909, when the model having the maximum score is the model of the keyword in the keyword class B163, the determination unit 154 instructs the re-recognition unit B813 to execute the process. The dictionary generation unit 831 of the re-recognition unit B 813 generates a subset dictionary B 853 from the keyword class B 163 and the main dictionary 851 (S910). Specifically, for example, the dictionary generation unit 821 extracts a vocabulary including a keyword in the keyword class B163 from the main dictionary 851, and stores it in the subset dictionary B853.

  A specific example of S910 will be described with reference to FIG.

  In FIG. 11, the main dictionary 851 includes “Atsugi City Hall”, “Atsugi High School”, “Inokashira Park”, “Sakai Park”, “Kyoto University”, “Kokubunji Station”, “Kokubunji City Hall”, “Kusatsu Onsen”, Fourteen vocabularies are registered: “Setagaya Park”, “Setagaya Art Museum”, “Setagaya High School”, “Tokyo Station”, “Shinagawa Station”, and “Haneda Airport”. Further, it is assumed that the vocabulary “high school”, “hot spring”, and “airport” is registered in the keyword class B163.

  The dictionary generation unit 831 registers a word including any keyword in the keyword class B163 from the words in the main dictionary 851 in the subset dictionary B853. In the example of FIG. 11, the words registered in the subset dictionary are four words “Atsugi High School”, “Kusatsu Onsen”, “Setagaya High School”, and “Haneda Airport”.

  In FIG. 9, the collation unit 832 collates the input feature vector series using the subset dictionary B853 and the scale model 854, and calculates a score (S911). This process is the same as S604 described above. Specifically, for example, in the case of FIG. 11 described above, the collation unit 832 determines that “Atsugi High School”, “Kusatsu Hot Spring”, regardless of which keyword “high school”, “hot spring”, or “airport” is recognized. ”,“ Setagaya High School ”and“ Haneda Airport ”are used to recognize speech using a subset dictionary B853.

  The determination unit 833 determines the vocabulary with the highest matching score (S912). This process is the same as S604 described above.

  The determination unit 823 and the determination unit 833 output the vocabulary with the maximum matching score to the output device 105 and the communication interface 106 (S913). Note that the vocabulary with the highest matching score may be used for processing by a program (not shown). This process may be the same as the process performed by the process execution unit A156, the process execution unit B157, and the like described in the first embodiment.

  In the second embodiment, as described above, by performing speech recognition using a subset dictionary based on keywords, it is possible to perform processing with a small amount of calculation and memory compared to speech recognition using a main dictionary. It becomes.

  The keyword class A162 is designed so that keywords with similar phoneme sequences are not mixed in advance, whereas the keyword class B163 has similar phoneme sequences such as “high school” and “airport”. Therefore, there is a possibility that a keyword that is easy to recognize incorrectly is included. However, in the keyword class B163, since the subset dictionary B853 is created regardless of the recognized keyword, it is possible to reduce erroneous recognition. That is, for example, even if the keyword is mistakenly recognized as “airport” for the utterance of “Atsugi high school”, the subset dictionary B853 includes an entry of “Atsugi high school”, so the correct recognition result Can be obtained.

  In addition, since the keyword class A 162 does not include phonologically similar keywords, there is little possibility of generating a subset dictionary based on an incorrect keyword. In addition, regarding the keyword of keyword class B, even if the keyword recognition is incorrect, it is possible to obtain a correct recognition result by using a subset dictionary that does not depend on the keyword.

  Also, a subset dictionary is created, and the speech to be recognized is specified from this dictionary. Therefore, it is particularly effective when the speech to be recognized is composed of a combination of a general noun and a noun in which various patterns such as place names and personal names are considered, such as “Atsugi High School”.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes design changes and the like within a scope not departing from the gist of the present invention.

  For example, in the above-described embodiment, the speech recognition processing apparatus classifies similar vocabularies from the confusion matrix, but the present invention is not limited to this. For example, the classification into the keyword class A and the keyword class B may be arbitrarily performed by the designer, or may be performed based on an arbitrary determination criterion.

  Also, the generation timing of the subset dictionary is arbitrary. For example, it may be generated and stored in advance, or may be created every time a keyword is recognized, every predetermined time, or when an instruction from a designer is input. A plurality of subset dictionaries may be created in advance, and an appropriate dictionary may be selected from the created subset dictionaries according to a given keyword.

  As another method of creating a subset dictionary, a subset dictionary may be created for words that do not include any keywords belonging to keyword class A162. An example of this will be described with reference to FIG. In FIG. 12, there are three keywords 1201 belonging to the keyword class A162: “station”, “park”, and “city hall”. If the vocabulary 1202 including these keywords is excluded from the vocabulary of the main dictionary 851, the remaining words are “Atsugi High School”, “Kyoto University”, “Kusatsu Onsen”, “Setagaya Art Museum”, “Setagaya High School”, “ 6 words “Haneda Airport”. These six words are registered in the subset dictionary B853. The subsequent processing is the same as described above.

It is a figure which shows the structural example of the speech recognition processing apparatus of 1st Embodiment. It is a figure for demonstrating an audio | voice model. In the same embodiment, it is a figure for showing an example of a confusion matrix. In the same embodiment, it is a figure for showing an example of a confusion matrix. It is a figure for demonstrating an audio | voice model. FIG. 6 is a diagram showing an operation example in the same embodiment. In the same embodiment, it is a figure which shows the example of a screen. It is a figure which shows the structural example of the speech recognition processing apparatus of 2nd Embodiment. FIG. 6 is a diagram showing an operation example in the same embodiment. FIG. 10 is a diagram for describing an operation example for generating a subset dictionary in the embodiment. FIG. 10 is a diagram for describing an operation example for generating a subset dictionary in the embodiment. FIG. 10 is a diagram for describing an operation example for generating a subset dictionary in the embodiment.

Explanation of symbols

  1: speech recognition processing device, 101: CPU, 102: memory, 103: secondary storage device, 104: input device, 105: output device, 106: communication interface, 151: class classification unit, 152: analysis unit, 153: Collation unit, 154: determination unit, 155: processing execution unit, 156: processing execution unit A, 157: processing execution unit B, 161: keyword model, 162: keyword class A, 163: keyword class B, 164: confusion matrix 165: Garbage model, 801: Speech recognition processing device, 811: Speech re-recognition unit, 812: Re-recognition unit A, 813: Re-recognition unit B, 821: Dictionary generation unit, 822: Verification unit, 823: Determination unit, 831: Dictionary generation unit, 832: Verification unit, 833: Determination unit, 851: Main dictionary, 852: Subset dictionary A, 853: Support Set dictionary B, 854: scale model

Claims (5)

In a computer-executable speech recognition program for recognizing a keyword that is a word included in input speech,
Storage means for classifying and storing a plurality of keywords into a first keyword group that does not include phonologically similar keywords and a second keyword group that is a keyword that is not included in the first keyword group The computer having
A keyword recognition step for recognizing at least one of the keywords included in the input voice;
When the recognized keyword is included in the first keyword group, a first process execution step of executing a first process;
When the recognized keyword is included in the second keyword group, a second process execution step for executing a second process is executed. A speech recognition processing program according to claim 1,
In the computer further storing a confusion matrix including a degree of similarity indicating a degree recognized as matching the same voice and another voice for each of a plurality of voices in the storage means,
From the confusion matrix and the keyword, a word whose similarity is less than a predetermined threshold is selected from the keywords, the word is classified as the first keyword group, and keywords other than the selected keyword are A classification step for classifying as a second keyword group;
A speech recognition processing program characterized by further executing A speech recognition processing program according to claim 1 or 2,
The first process is a process determined according to the recognized keyword,
The voice recognition processing program, wherein the second process is a specific process. A speech recognition processing program according to claim 3,
The voice recognition processing program characterized in that the second processing outputs information for confirming the recognized keyword to an output means. In a computer-executable speech recognition program for recognizing a keyword that is a word included in input speech,
The plurality of keywords include a first keyword group that does not include phonologically similar keywords, a second keyword group that is a keyword that is not included in the first keyword group, and one or more words. A storage means for storing a dictionary including a plurality of vocabularies,
A keyword recognition step for recognizing at least one of the keywords included in the input voice;
When the recognized keyword is included in the first keyword group, a vocabulary including the recognized keyword is selected from the dictionary, and the selected vocabulary is stored in the storage unit as a first subset dictionary. A first vocabulary recognition step for extracting from the input speech a vocabulary included in the first subset dictionary and a vocabulary including the recognized keyword;
When the recognized keyword is included in the second keyword group, the vocabulary including the plurality of keywords is selected from the dictionary, and the selected vocabulary is stored in the storage unit as a second subset dictionary. A second vocabulary recognition step of extracting, from the input speech, a vocabulary included in the second subset dictionary and a vocabulary including the recognized keyword;
And a step of outputting the extracted vocabulary.

Images