JP4680714B2 - Speech recognition apparatus and speech recognition method - Google Patents

Description

  The present invention relates to a speech recognition apparatus that recognizes input speech using a speech recognition dictionary and performs a dialog by changing a system state based on a recognition result.

  As one of the general speech recognition methods, the most similar pattern is calculated by comparing the various acoustic patterns representing the vocabulary registered in the recognition dictionary in advance with the speech signal input from the user. There is a method in which a vocabulary in a recognition dictionary indicating a recognition result candidate is used.

  In such a general speech recognition method, speech recognition is performed using an acoustic model created in order to achieve high recognition accuracy for many users. Since it is not adapted, recognition accuracy becomes low, and a lot of misrecognitions occur.

  In addition, conventionally, for example, there is a technique for calculating a recognition reliability using a recognition score, and rejecting based on the recognition reliability even when the most similar recognition candidate is obtained and prompting re-input. By providing such a reject function, it is possible to prevent the system from malfunctioning due to erroneous recognition when, for example, noise other than speech is input.

  In such a speech recognition method, when the obtained recognition reliability is lower than a predetermined value, it is possible to prevent erroneous recognition and erroneous system operation by a function such as reject, but on the other hand, even for normal speech input Since the rejection is performed when the recognition reliability is low, a specific recognizable vocabulary may be erroneously rejected depending on the user.

  Therefore, as a countermeasure against such misrecognition / rejection, in order to adapt the acoustic model for general unspecified speakers used for speech recognition to the current user's acoustic model, the user's own speech is used again. There are proposed a method of improving recognition accuracy by learning (speaker adaptation / speaker learning) and a method of improving recognition accuracy at the time of recurrence of a user when a rejection is made.

For example, as a method of speaker adaptation / speaker learning, a method of learning an acoustic model using a small number of voices and further performing speaker learning for a misrecognized word is disclosed (for example, see Patent Document 1). Has been. On the other hand, as a method for improving the recognition accuracy at the time of recurrence voice, when it is determined that it is rephrased, a method of determining a recognition result using both the previous and current recognition candidates (see, for example, Patent Document 2), For correction utterances, a technique (for example, refer to Patent Document 3) is disclosed in which the top candidate of the previous recognition result is a recognition target vocabulary.
JP 2003-162292 A JP-A-11-149294 Japanese Patent No. 3112037

  In the conventional methods as described above, a device that can learn the user's acoustic model with a small number of learning utterances, or a device that improves recognition accuracy by changing the recognition candidates and recognition target vocabulary at the time of recurrence is made. ing.

  However, in speaker adaptation based on these learnings, a small amount of words that are not directly related to the operation are uttered by the user when learning a model suitable for the user, but there is a considerable burden on the user. In addition, the recognition accuracy improvement method at the time of re-input increases the recognition accuracy at the time of recurrence, but when the user utters the same utterance as the previous utterance again, it will be rejected again, Must be done.

  For example, in Japanese Patent Application Laid-Open No. 2003-228867, the user is first encouraged to learn speech for adaptation to a speaker, but for words that are frequently misrecognized, the user is prompted to utter the part that causes misrecognition. However, it increases the burden on the user in order to encourage the user to speak extra. Also, in Patent Document 2, when a rewording is detected, the recognition candidates to be output including the previous recognition result are adjusted. However, when the same utterance as the previously rejected utterance is input, the recognition is rejected. I have to repeat my voice every time. Further, in Patent Document 3, the next recognition rejected recognizes only the previous top candidate as a recognition target vocabulary. However, as in Patent Document 2, when the same utterance as the previously rejected utterance is input, I can't recognize it.

  Therefore, the present invention has been made to solve the above-described conventional problems, and does not require a special utterance for learning from the user, so that the user's burden is reduced and the individual adaptation of voice recognition is naturally performed. An object of the present invention is to provide a speech recognition apparatus and a speech recognition method that can be performed and that can reduce erroneous recognition.

In order to achieve the above object, a speech recognition apparatus according to the present invention recognizes an input speech and performs a dialog by changing a system state that is a system state related to a dialog with a user according to a recognition result. a speech recognition apparatus using the voice recognition dictionary, and recognizing the input speech, a speech recognition means for outputting a recognition result, the recognition result by the by transitioning the system state of the voice recognition means responsive When it is determined that the system state has escaped from the stagnation state where the system state has not progressed further and stays stagnant as a result of the current recognition, and has escaped from the stagnation state Stagnation escape determination means for determining whether or not the current recognition result is at least one of rephrasing and paraphrasing; And changes in the reject threshold as settings for, characterized by comprising a change control means for performing at least one of the newly added or changed to the speech recognition dictionary as a change of configuration for voice recognition.

  According to the speech recognition apparatus and the speech recognition method according to the present invention, a special utterance for learning is required in order to eliminate the mismatch between the features of the utterance of the user and the parameters for speech recognition of the system and the speech recognition dictionary. In addition, since learning suitable for the user is performed using the result correctly recognized by re-phrase or paraphrase once, it is possible to perform personal adaptation of voice recognition that is natural and less burdensome for the user. In addition, since personal recognition of voice recognition is performed, it is possible to recognize correctly even if the utterance is the same as the previously mistaken utterance from the next time, so that smooth voice operation can be realized by reducing the misrecognition. Can do.

  A speech recognition apparatus according to an embodiment of the present invention is a speech recognition apparatus that recognizes an input speech and performs a dialogue based on a recognition result, and recognizes the input speech using a speech recognition dictionary and recognizes the result. Is a state where the system state does not advance further and is stagnant by the current recognition result. A stagnation escape determining means for determining whether or not the vehicle has escaped from the stagnation state and determining whether or not the current recognition result is at least one of rephrasing and paraphrasing when it is determined that the vehicle has escaped from the stagnation state; And a change control means for changing at least one of a setting relating to dialogue control and a setting relating to voice recognition when it is determined that the rephrase or paraphrase is made. And features.

  As a result, user adaptation is performed as needed during normal voice operations, so that special utterance is not necessary for user adaptation, and it is possible to perform personal adaptation of voice recognition that is natural and less burdensome for the user. In addition, since personal recognition of voice recognition is performed, it is possible to recognize correctly even if the utterance is the same as the previously mistaken utterance from the next time, so that smooth voice operation can be realized by reducing the misrecognition. Can do.

  Further, the stagnation state of the system state is a state in which the same system state continues due to the rejection of the voice recognition result, and the stagnation escape judging means says when the current recognition result is the same word as the previous recognition result. If it is determined to be corrected and the current recognition result is not the same word as the previous recognition result, but is a recognition word that executes the same predetermined system operation, it may be determined to be paraphrased.

  In addition, the stagnation state of the system state is a state in which the reciprocation of two system states continues repeatedly, and the stagnation escape determination means is rephrased when the current recognition result is the same word as the previous recognition result. If the current recognition result is not the same word as the previous recognition result, but is a recognition word that executes the same predetermined system operation, it may be determined as a paraphrase.

  As a result, the stagnation of the voice operation due to erroneous rejection and the stagnation of the voice operation due to erroneous recognition are reduced, and a smooth voice operation can be realized.

  The change control means may change a rejection threshold as a change in the setting related to the dialog control, and may newly add or change the voice recognition dictionary as a change in the setting related to the voice recognition. Thereby, the rejection accuracy and the recognition accuracy can be improved, and personal adaptation of voice recognition and smooth various voice operations can be realized with less burden on the user.

  Further, the change control means may set and change the rejection threshold for each recognition target word. Thereby, personal adaptation for each recognition target word is possible, and it is possible to realize personal adaptation of voice recognition and smooth various voice operations with less burden on the user.

  Further, the change control means may set the rejection threshold and new addition or change to the speech recognition dictionary for each user. As a result, appropriate adaptation is possible even when used by a plurality of users, and personal adaptation of voice recognition and smooth various voice operations with less burden on the user can be realized.

  In addition, when the voice recognition device further escapes from the stagnation state, the current recognition result is not the same word as the previous recognition result and is not a recognition word that executes the same predetermined system operation, An abbreviation creation means for creating an abbreviation for the current recognition result is provided, wherein the speech recognition means re-recognizes the previous recognition result using the abbreviation, and the change control means re-reads the speech recognition means. The abbreviation may be newly added to the speech recognition dictionary according to the recognition result. Accordingly, even if the user uses the abbreviation, appropriate adaptation is possible, and it is possible to realize personal adaptation of voice recognition and smooth various voice operations with less burden on the user.

  The present invention can be realized not only as such a speech recognition apparatus, but also as a speech recognition method using steps characteristic of the speech recognition apparatus, or by performing these steps as a computer. It can also be realized as a program to be executed. Needless to say, such a program can be distributed via a recording medium such as a CD-ROM or a transmission medium such as the Internet.

  Embodiments of the present invention will be described below with reference to the drawings.

(Embodiment 1)
FIG. 1 is a block diagram showing a configuration of a voice interactive information search system provided with a voice recognition device according to Embodiment 1 of the present invention.

  The voice interactive information search system is a system for searching for information while inputting a voice, and as shown in FIG. 1, the voice recognition unit 101, the voice recognition dictionary 102, the voice recognition parameter storage unit 103, the stagnation An escape determination unit 104, a dialog control unit 105, a dialog history storage unit 106, a system specification storage unit 107, a database search unit 108, a database storage unit 109, a response voice / screen output unit 110, and a timer 111 are provided.

  The speech recognition unit 101 performs speech recognition of speech input by the user using the speech recognition dictionary 102 and the speech recognition parameter storage unit 103, and outputs a recognition result. The speech recognition dictionary 102 is a dictionary in which recognition target words are registered. The voice recognition parameter storage unit 103 stores voice recognition parameters.

  The dialogue control unit 105 controls the dialogue in accordance with operation specifications determined in advance by the system developer, and determines the next system state in response to an input from the user. Specifically, the dialogue control unit 105 receives a voice recognition result input from the voice recognition unit 101, a determination result as to whether or not a stagnation escape is input from the stagnation escape determination unit 104, and a dialogue history storage unit 106. The next system state is determined from the system specification storage unit 107 using the current and past dialog history. Further, the dialogue control unit 105 requests the database search unit 108 to change the speech recognition dictionary 102 and the speech recognition parameter 103 and search the database if necessary. The system state indicates one state of the system in the system operation specification determined by the system developer.

  The stagnation escape determination unit 104 determines whether or not the system state transition has escaped from the stagnation state using information such as the current and past user recognition results input from the dialogue control unit 105. The dialogue history storage unit 106 stores various information in each system state such as a voice recognition result input from the dialogue control unit 105 and a system output (output screen information / output response information) result. The system specification storage unit 107 stores system operation specifications predetermined by the developer.

  In response to the information search request from the dialog control unit 105, the database search unit 108 searches from the database in the database storage unit 109. The database storage unit 109 stores a search target database of the database search unit 108. The response voice / screen output unit 110 outputs a screen and response voice corresponding to the system state input from the dialogue control unit 105. The timer 111 outputs the current time to the dialogue control unit 105 in response to a request from the dialogue control unit 105.

  Next, a specific operation when searching for program information in the voice interactive information search system configured as described above will be described. FIG. 2 is a flowchart showing the flow of the entire dialogue in the voice interactive information retrieval system.

  The dialog control unit 105 determines the system state of the dialog start, determines the screen and response voice in the determined system state, and outputs the response voice / screen output unit 110 to make an input request to the user. (Step S201). FIG. 3 shows a specific output screen example. Here, for example, as shown in FIG. 3, a menu screen for searching for program information is output, and the content 301 of the balloon of the agent is output as a response voice. The balloon itself may be displayed on the screen. In this example, it is assumed that the vocabulary that can be recognized in FIG. 3 is only the vocabulary surrounded by a square. For example, suppose that “1”, “1”, “program name search”, and “1. program name search” are vocabulary that can be recognized for selecting the square “1. program name search” 302.

  The voice recognition unit 101 performs a process of recognizing an input voice from a user in response to an input request by a response voice / screen from the system (step S202). More specifically, first, the dialogue control unit 105 requests the speech recognition unit 101 to notify a vocabulary that can be recognized in the current system state and to execute speech recognition processing. More specifically, in the system state shown in FIG. 3, the speech recognition unit 101 starts speech recognition processing using a vocabulary surrounded by a square as a recognition target vocabulary. Next, the voice recognition unit 101 performs a recognition process on the user input voice and outputs a recognition result to the dialogue control unit 105. Here, the output recognition result includes not only the word of the recognition target vocabulary closest to the user's utterance but also detailed information as described later regarding the recognition.

  4 and 5 are diagrams showing specific examples of the output recognition results. FIG. 4 shows the overall information of the speech recognition centering on the first recognition result, and FIG. The recognition result information including the candidates is shown. Here, the item 401 is the date and time when the recognition result is output, and the item 402 is the input speech section, that is, the section that is determined to be speech among the sections that the speech recognition unit 101 has performed the recognition process. is there. Item 403 is the word that is determined to be the closest in the recognition target vocabulary, that is, the word with the highest recognition result candidate, and item 404 is the acoustic result that is the kana character that is acoustically independent of the speech recognition dictionary. This character string is generally called the result of an audio typewriter. An item 405 is a section in which words of recognition results are matched in the input speech section. An item 406 is a recognition score indicating a recognition degree, and a higher score indicates a higher recognition degree. An item 407 indicates the recognition reliability and indicates how much the recognition result is appropriate. In general, the recognition reliability is often calculated using a difference between recognition candidate scores or a difference between a speech typewriter and a recognition candidate. An item 408 is a threshold value for rejection, and is a variable stored in the voice recognition parameter storage unit 102.

  The dialogue control unit 105 compares the rejection threshold value with the recognition reliability, and determines whether or not to accept the recognition result as a system. Specifically, when the recognition reliability is lower than the reject threshold, the dialogue control unit 105 rejects the recognition result, that is, does not process it as an input result, and prompts input in the same system state again. For example, in the example of FIG. 4, since the recognition reliability is “4.5” and the rejection threshold is “3.5”, the dialogue control unit 105 recognizes this recognition result “program name search” as an input to the system and has a dialogue. Take control. The rejection threshold may be determined in advance by the system developer or by performing an evaluation experiment. Specifically, some subjects may utter words in this dictionary set and may be determined based on the result.

  Also, in FIG. 5, item 501 is the recognition candidate rank as a result of sorting by the recognition scores of the recognition candidates, items 502 to 505 are information of each recognition candidate, and the contents thereof are the recognition described in FIG. The result is the same as the information of the first result.

  The dialogue control unit 105 outputs the speech recognition result input from the speech recognition unit 101 in step S202 and the previous recognition result stored in the dialogue history storage unit 106 to the stagnation escape determination unit 104.

  Next, the stagnation escape determination unit 104 determines whether or not the current input is a stagnation escape and outputs the result to the dialogue control unit 105 (step S203). The dialogue control unit 105 writes this result in the dialogue history storage unit 106.

  Here, the stagnation / escape determination operation in the stagnation / escape determination unit 104 will be described in detail with an example of stagnation due to erroneous recognition of speech recognition. FIG. 6 is a flowchart showing the flow of the stagnation / escape determination operation in the stagnation / escape determination unit 104.

  First, the stagnation escape determination unit 104 acquires the current speech recognition result and the previous recognition result (step S601). Then, based on the voice recognition result, it is determined whether or not it is rejected (step S602). As a result of this determination, when it is determined to be rejected (YES in step S602), the stagnation escape determination unit 104 outputs a determination result that the stagnation escape is not performed. This is because reject means that the reliability of the recognition result is low and is not adopted as the recognition result. In this case, the state does not advance to the next system state, that is, does not escape from the stagnation.

  On the other hand, if it is determined not to be rejected (NO in step S602), it is determined from the dialog history whether the previous utterance was rejected (step S603). As a result of this determination, if it is determined that the previous utterance is not rejected (NO in step S603), no stagnation has occurred in the previous utterance, so that the stagnation escape determination unit 104 determines that the current utterance is not in stagnation. The judgment result that there is no is output.

  On the other hand, when it is determined that the previous utterance is rejected (YES in step S603), the stagnation escape determination unit 104 can escape from the stagnation state by the current utterance to indicate that the system is in a stagnation state due to the previous utterance. It is determined whether or not it is rephrased (step S604). Retouching here means that the previous utterance and the current utterance are the same. For example, the user may say “Search program name” on the output screen as shown in FIG. 3 and say “Search program name” again when the user is rejected and prompted to input again. This re-statement determination is performed by comparing the previous recognition result with the current recognition result. If it is determined that the re-statement is made (YES in step S604), the stagnation escape determination unit 104 escapes the stagnation due to re-statement. The determination result that is is output.

  On the other hand, when it is determined that it is not rephrasing (NO in step S604), the stagnation escape determination unit 104 determines whether or not it is paraphrasing (step S604). The paraphrase here means a utterance in which the utterance vocabulary is different between the previous utterance and the current utterance, but the utterance contents are the same, that is, the system operation by the utterance is the same. For example, the user may say “Search program name” on the output screen as shown in FIG. 3 and say “No. 1” when rejected and prompted to re-input. This paraphrasing determination is performed by comparing the previous recognition result with the current recognition result in the same manner as the rewording determination. More specifically, if the vocabulary between the previous recognition result and the current recognition result is different and the vocabulary in which the recognition result performs the same operation as the system specification, it is determined that the words are paraphrased. Whether the recognition result is the same operation as the system specification is determined by each system specification defined in the system specification storage unit 107. Specifically, as shown in FIG. 7, the system specification storage unit 107 stores a vocabulary to be accepted as a recognition result and a state to be changed to when the vocabulary is accepted, and can be selected here. The word corresponding to the item is treated as a paraphrase target word.

  As a result of this determination, when it is determined that the paraphrase is made (YES in step S605), the stagnation escape determination unit 104 outputs a determination result that the stagnation escape is due to paraphrase. On the other hand, when it is determined that it is not paraphrasing (NO in step S605), the stagnation escape determination unit 104 outputs a determination result that is not stagnation escape.

As described above, the stagnation escape determination unit 104 performs the stagnation escape determination operation.
Next, the dialogue control unit 105 changes the speech recognition parameters such as the speech recognition dictionary, the rejection threshold, and the acoustic model based on the speech recognition result and the stagnation escape determination result obtained until the stagnation escape determination process (step S203). Is performed (step S204).

  Next, the dialogue control unit 105 determines the next system state and the response voice and screen output in this system state based on the recognition result, and outputs them to the response voice / screen output unit 110 (step S205). If necessary, the dialog control unit 105 outputs a result of requesting the database search unit 108 to search for data from the database storage unit 109 to the response voice / screen output unit 110.

  Then, the dialogue control unit 105 determines whether or not the dialogue is ended according to the system specification defined in the system specification storage unit 107. As a result, if the dialogue is not finished (NO in step S206), the above steps are repeated again from the input speech recognition process (step S202). If the dialogue is finished (YES in step S206), the dialogue is performed. finish.

  Next, a specific operation example of the system will be described with reference to a specific example of the system output screen of the system and dialog history data stored in the dialog history storage unit 106.

  FIG. 8 is a diagram illustrating a specific example of dialogue history data targeted in the operation example. Item 801 is a step number assigned to centrally manage changes in system status, item 802 is a system status indicating the type of system status, item 803 is a response output start time indicating the date and time when the system outputs a response, and item 804 is the first candidate word of the speech recognition result obtained from the speech recognition unit 101, item 805 is also the recognition reliability obtained from the speech recognition result, and item 806 is a dialogue based on the speech recognition result from the speech recognition unit 101 Reject determination result determined by the control unit 105, item 807 is a rephrase determined by the stagnation escape determination unit 104 based on the previous recognition result stored in the conversation history storage unit 106 and the recognition result output by the current voice recognition unit 101 The result of determination of stagnation escape by means of item 808 is based on the paraphrase determined by the stagnation escape determination unit 104 in the same manner as item 807. Stagnation Escape determination result, item 809 are stored in the speech recognition parameter storage unit 103, a reject threshold that can be acquired from the recognition result. Although not shown in this figure, the detailed information of the recognition result as shown in FIG. 4 at each step, the information about the display screen as shown in FIG. 5, specifically the displayed words and Other information such as an output response sentence indicating what kind of response sentence the system has output may be stored in the dialogue history storage unit 106.

  For example, assume that the user utters “program name search” on the menu screen shown in FIG. Since the recognition reliability (0.47) of the recognition result is higher than the rejection threshold (0.35), the dialogue control unit 105 determines the next system state, and outputs a screen transition and a response sentence (FIG. 8). Step = 1). Specifically, the response voice / screen output unit 110 outputs a screen as shown in FIG. 9 and a response voice “Please specify the initial of the program name” from the system. Next, the user utters “A line”, which, like the previous utterance, is recognized correctly because the recognition certainty (0.36) is higher than the reject threshold (0.35) (step in FIG. 8). = 2). From the system, the response voice / screen output unit 110 outputs a screen as shown in FIG. 10 and a response “Which program?”. Next, the user utters “next screen” because there is no program to be viewed in the list, but in the recognition result for this utterance, the recognition reliability (0.33) is lower than the rejection threshold (0.35). It is determined to be a rejection (step = 3 in FIG. 8). If it is determined to be rejected, the dialogue control unit 105 again prompts input in the system state (in this case, dialogue = state 3). This operation is a malfunction of the system in which the dialog control unit 105 rejects the user correctly while the user is uttering correctly, and is an operation that occurs because the reject threshold is not set correctly for the user.

  In the same system state again, the user who is prompted to input by the system speaks again “next screen”, and the recognition reliability (0.38) in the result of the speech recognition is higher than the rejection threshold (0.35). The dialogue control unit 105 accepts the result (step = 4 in FIG. 8). Here, in this step, the stagnation / escape determination unit 104 determines that “the previous utterance is rejected” and “the current utterance is re-stated”, so that “stagnation escape due to re-statement” is determined, and the information is stored in the item 807. The Furthermore, in this step, the dialogue control unit 105 performs personal adaptation so as to accept the first utterance that has malfunctioned from the next by using the detected malfunction and the correct action. That is, the speech recognition parameter, in this example, the rejection threshold is changed, and speech recognition is performed using this value from the next step. Specifically, the current rejection threshold value “0.35” is changed to “0.30” so that it can be correctly recognized even with the reliability in the previous erroneously rejected utterance. This threshold value may be changed at a fixed rate preset by the system developer. Further, the threshold value may be changed by calculation using the current rejection threshold value and the recognition reliability when the rejection is erroneously performed. More specifically, if the difference between the current rejection threshold and the recognition reliability when erroneously rejected is within a certain value, the rejection threshold is set to the recognition reliability when erroneously rejected, and the difference is If the value is equal to or greater than a certain value, the reject threshold value may be set using a weighted average value between the current reject threshold value and the recognition reliability when the error is rejected by mistake. Further, the threshold value may be changed using the recognition reliability of the word correctly recognized as the rejected word. Specifically, it is determined by the same method as the calculation method using the current rejection threshold and the recognition reliability when erroneously rejected.

  Upon receiving the restatement result, the dialogue control unit 105 determines the next system state, and outputs a screen transition and a response sentence. Specifically, the response voice / screen output unit 110 outputs a screen as shown in FIG. 11 and a response voice “Which program?” From the system. Since there is no program that the user wants to watch on this screen, the user further says “next screen”. The recognition reliability in the recognition result of this utterance is “0.33”, which is the same as in step 3. This recognition reliability is the value rejected in step 3, but the dialogue control unit 105 compares the recognition reliability “0.33” with the rejection threshold “0.3” adapted in the previous step. In this step, this utterance is not rejected, the next system state is determined, and screen transitions and response sentences are output. Specifically, the response voice / screen output unit 110 outputs a screen as shown in FIG. 12 and a response voice “Which program?” From the system (step = 5 in FIG. 8). The user finds the program he / she wants to see on this screen, and utters “I want to do” to select the program (step = 6 in FIG. 8). FIG. 13 is a diagram summarizing the series of operations described above, and corresponds to step = 1 to step = 6 in FIG. 8 in order from the top.

  Next, a reject threshold value changing operation example using paraphrasing will be described using a specific example of dialogue history data. FIG. 14 is a diagram illustrating a specific example of dialogue history data targeted in the operation example. The items of the dialogue history data are the same as those in FIG. Furthermore, since the difference in utterance from the example of the operation of changing the rejection threshold due to the above rewording is only step 3 to step 5, only the example of operation from step 3 to step 5 in FIG. 14 will be described.

  A response voice / screen output unit 110 outputs a screen as shown in FIG. 10 and a response “Which program?” From the system. The user utters “next screen” because there is no program to view in the list. However, in the recognition result for this utterance, the recognition reliability (0.33) is lower than the rejection threshold (0.35), so the rejection is made. It is determined that there is (step = 3 in FIG. 14). If it is determined to be rejected, the dialogue control unit 105 prompts input again in the system state (in this case, dialogue = state 3).

  Again, in the same system state, the user who is prompted to input by the system utters “No. 5”, which is a command for performing the same system operation as the “next screen”. Since the recognition reliability (0.38) in the voice recognition result is higher than the reject threshold (0.35), the dialogue control unit 105 accepts the result (step = 4 in FIG. 14). Here, in this step, the stagnation / escape determination unit 104 determines that “the previous utterance is rejected” and “the current utterance is paraphrased”, so it is determined as “stagnation escape due to paraphrasing”, and the information is stored in the item 1408. Further, in this step, the dialogue control unit 105 performs personal adaptation using the detected malfunction and the correct action to accept the first utterance that has malfunctioned from the next. That is, the speech recognition parameter, in this example, the rejection threshold is changed, and speech recognition is performed using this value from the next step. Specifically, the current rejection threshold “0.35” is changed to “0.3” so that the reliability can be correctly recognized even in the previous erroneously rejected utterance. Subsequent operations are the same as in the example of changing the reject threshold by the above-mentioned rephrasing, and are therefore omitted.

  In the above specific example, the rejection threshold was changed at the stage where "stagnation escape by rephrasing or paraphrasing" was detected once, but the number of times of stagnation escape detection as a reference for changing the speech recognition parameter is variably set. You may be able to do it. For example, if it is set to 3 times, the rejection threshold value is changed when “stagnation escape by rephrase or paraphrase” is detected 3 times. In this case, for example, the rejection threshold value may be changed using the recognition reliability in the recognition results for three times. More specifically, it is determined by the minimum value, average value, and weighted average value of the reliability in the recognition results for three times.

  Also, in the above specific example, an example having only one reject threshold has been described, but even if there is a reject threshold for each word, “stagnation escape by rephrase or paraphrase” is detected for each word, and the threshold is changed. Good. Specifically, for example, data as shown in FIG. 15 is stored in the speech recognition parameter storage unit 103. Here, the item 1501 is a word whose reject threshold is changed due to stagnation escape, and the item 1502 is the reject threshold of the word. A word that is not in this list uses a default value, for example, “0.35” in the above specific example.

  FIG. 16 is a diagram showing an example of a dialogue sequence when the present embodiment is used and when it is not used. In the example shown in FIG. 16, when the present embodiment is used, the user's utterance can be reduced once. In this example, the user decides to select a program on the second page. However, the effectiveness of the present embodiment appears remarkably in a dialog having many steps as in the case of viewing more pages. It ’s easy to understand. Also, when the search is finished once and the same program is selected again from the beginning, using this embodiment, the utterance method rejected last time can be recognized from the beginning.

  Thus, according to the present embodiment, it is possible to appropriately change the speech recognition parameter by detecting a malfunction and a correct operation in a series of dialogue sequences. As a result, the system can operate correctly even if the next malfunctioning utterance is performed, so that it is not necessary to repeat it again and again, and a smooth conversation that does not burden the user can be realized. Moreover, since the change of the speech recognition parameter according to the present embodiment does not prompt special utterance for the change, the burden on the user is small.

  As shown in FIG. 17, the present embodiment includes an EPG receiving unit 201 that receives an EPG (Electronic Program Guide) in addition to the above-described configuration, and performs voice recognition for EPG as a target to search for information. The present invention can also be applied to a type information retrieval system. In this case, the EPG received by the EPG receiving unit 201 is stored in the database storage unit 109. The dialogue control unit 105 creates the speech recognition dictionary 102 using the EPG stored in the database storage unit 109. The voice recognition unit 101 performs voice recognition of the voice input by the user using the voice recognition dictionary 102 created using EPG. In addition, the database search unit 108 performs a search for an EPG or the like stored in the database storage unit 109.

(Embodiment 2)
According to the first embodiment, it is possible to personally adapt voice recognition parameters by detecting a malfunction and a correct operation, and to realize personal adaptation with less burden on the user. The same adaptation is referred to as addition of a voice recognition dictionary. You can also do it. In this embodiment, a method of detecting a malfunction and a correct operation and changing or newly registering a speech recognition dictionary will be described.

  The present embodiment is different from the above-described first embodiment in that individuals to be personally adapted are based on the stagnation escape determination result in the dialogue control unit 105 in FIG. Therefore, basically, reference is made to FIGS. Hereinafter, the operation of the dialogue control unit 105 in the present embodiment and the voice recognition dictionary change processing not described in the previous embodiment will be described.

  An operation example of personal adaptation by dictionary change / registration in the present embodiment will be described using a specific example of dialogue history data stored in the dialogue history storage unit 106.

  FIG. 18 is a diagram illustrating a specific example of dialogue history data stored in the dialogue history storage unit 106. The example of the dialog history data shown in FIG. 18 is the same history as the example of the dialog history data (FIG. 8) in the dialog sequence in the first embodiment, and the item “recognition result 2” not shown in FIG. 8 is the same as FIG. 8 except that the item “response output start time” shown in FIG. 8 is omitted. Note that as described above, the recognition result 2 is obtained by outputting an acoustically close kana character string as the speech recognition result without using the speech recognition dictionary, and the speech recognition result 2 in FIG. 4 showing an example of the recognition result. Is the same.

Hereinafter, specific operations will be described in order using the item “step” in FIG.
In step 3, the speech recognition unit 101 outputs a recognition result 2 of “sui no damen”, a recognition reliability of “0.33”, and a rejection threshold of “0.35” for the user's utterance “next screen”. Since the recognition reliability is lower than the rejection threshold, the dialogue control unit 105 determines that the recognition is rejected, and prompts input again in the system state. In step 4, the voice recognition unit 101 outputs the recognition result 2 “Tsurino Gamen”, the recognition reliability “0.38”, and the rejection threshold “0.35” in response to the user's recurrent voice “next screen”. The determination unit 104 outputs a determination “stagnation escape due to rephrasing”. In response to these results, the dialogue control unit 105 performs personal adaptation so that the utterance in step 3 that has malfunctioned is correctly recognized from the next time. That is, “Suri no damen” of the speech recognition result 2 for the utterance rejected in step 3 is newly registered in the speech recognition dictionary 102 in correspondence with the command “next screen” correctly recognized in step 4.

  FIG. 19 is a diagram showing a specific example of a speech recognition dictionary. The item 1801 is a word number uniquely assigned to each word, the item 1802 is a meaning number assigned with the same number as the same meaning in the system specification, the item 1803 is a word notation, and the item 1804 is a word reading is there. Here, in the above example, the word number 130 in FIG. 19 is newly registered.

  In step 5, the user speaks “next screen”. As in the case of step 3, the speech recognition unit 101 outputs the result “sui no damen” as the speech recognition result 2, but at the time of the speech recognition at this time, “sino damen” is displayed as “next screen” in the speech recognition dictionary 102. Since it is registered in a corresponding state, a high certainty factor (0.45 in this case) is output as a result. As described above, the utterance in step 5 is correctly recognized in spite of the utterance similar to that in step 3 due to personal adaptation.

  In the above specific example, the recognition dictionary was changed when the rephrasing “stagnation escape by rephrasing” was detected once, but the number of times of stagnation escape detection as a reference for changing the recognition dictionary is variable. It may be settable. For example, if it is set to 3 times, the recognition dictionary will be changed when “stagnation escape by rephrasing” is detected 3 times. Here, all the recognition results in the recognition results for three times may be registered, or a character string created in combination may be registered. Specifically, “Tsugino Damen” obtained by changing only “Da” part that is common to “Shinoda Damen”, “Tsugino Damen”, and “Sugino Damen” may be registered. Furthermore, the kana characters that have been changed are stored, and when it is determined that “ga” is often mistaken for “da”, “ga” may be changed to “da” for other words. Specifically, the reading “Maenodamen” may be assigned to the “previous screen” and additionally registered in the speech recognition dictionary.

  In the present embodiment, only the example of adding / changing the speech recognition dictionary based on the determination of rephrasing stagnation has been described. However, in the same manner as in the first embodiment, addition / change of the speech recognition dictionary is also performed in the case of paraphrasing. It can be carried out.

  As described above, according to the present embodiment, it is possible to appropriately change not only the speech recognition parameters but also the speech recognition dictionary by detecting a malfunction and a correct operation in a series of dialogue sequences. As a result, the system can operate correctly even if the next malfunctioning utterance is performed, so that it is not necessary to repeat it again and again, and a smooth conversation that does not burden the user can be realized. In addition, the change of the speech recognition dictionary according to the present embodiment does not prompt special utterances to increase the recognition rate, and raises the recognition rate from natural conversation, so the burden on the user is small.

  It should be noted that the addition / change to the voice recognition dictionary in the present embodiment and the change in the voice recognition parameter in the first embodiment may be combined.

(Embodiment 3)
According to Embodiment 1 and Embodiment 2 described above, the speech recognition parameters and the recognition dictionary are changed to those suitable for the user by detecting malfunctions and correct operations in a series of dialogue sequences. In the first embodiment and the second embodiment described above, the system specification storage unit determines whether the word that was previously rejected for “paraphrase” is a word that performs the same system operation as the word that was correctly recognized this time. Judgment is made using data such as However, there are various forms of “paraphrasing”, and there are cases where registration in the system specification in advance is not possible. In particular, in a program search system using EPG, it is necessary to recognize a program name that is updated every day, and the system developer cannot register the paraphrase in advance. The present embodiment addresses such a case.

  FIG. 20 is a block diagram showing a configuration of a voice interactive information retrieval system including a voice recognition device according to Embodiment 3 of the present invention.

  The difference between the third embodiment and the first and second embodiments is that the operation of the stagnation / escape determination unit 302 due to the addition of the abbreviation creation unit 301 and the user utterance storage unit 304 is different. The other operations are the same as those in the first and second embodiments. Therefore, in the present embodiment, only the paraphrase determination operation of the stagnation escape determination unit 302 having different operations will be described.

  The stagnation escape determination unit 302 performs stagnation escape determination by rephrasing and paraphrasing according to the flowchart of FIG. 6 in the same manner as in the first and second embodiments, but the processing in step S605 of FIG. The process for determining whether the utterance is paraphrasing is different. FIG. 21 is a flowchart showing the flow of the paraphrase determination operation in the present embodiment.

  First, it is determined whether or not the current utterance is a recognized word that performs the same system operation as the previous utterance (step S2001). As a result of this determination, if the recognition word is the recognition word that performs the same system operation as the previous utterance (YES in step S2001), the operation is the same as in the previous embodiments, and it is determined that the stagnation escape due to paraphrase (FIG. 6). (Step S605: YES) On the other hand, if it is not a recognized word that performs the same system operation as the previous utterance (NO in step S2001), an abbreviation is created from the current recognition target vocabulary (step S2002). The abbreviation creation is performed in the abbreviation creation unit 301 using the current recognition target vocabulary.

  The abbreviation creation unit 301 receives the current recognition target vocabulary and creates abbreviations based on predefined rules. As an abbreviation creation method, the current recognition target vocabulary is decomposed into morphemes using a morpheme analysis tool, and the abbreviation is created based on the decomposition results. For example, one morpheme may be an abbreviation, or a plurality of morphemes may be connected to form an abbreviation. More specifically, for example, an abbreviation such as “Excavation”, “A certain”, “Korenji”, “A certain Hiroji” can be created for the word “Excavation with a certain broad word”, or “Winter Requiem” To create an abbreviation such as “Winter Lek”. The abbreviation created by the abbreviation creation unit 301 is held in the dialogue control unit 303 via the stagnation escape determination unit 302.

  Next, the speech recognition unit 101 uses the abbreviation created by the abbreviation creation unit 301 held in the dialogue control unit 303 to perform the previously rejected utterance stored in the user utterance storage unit 304. Recognition is performed again (step S2003).

  Then, the stagnation escape determination unit 302 compares the reliability of the re-recognition result with the rejection threshold (step S2004). Here, when the reliability of the re-recognition result is higher than the reject threshold (YES in step S2004), the dialogue control unit 303 sets the abbreviation of the recognition candidate first place as a word that performs the same operation as the currently recognized word. It registers in the memory | storage part 107 and the speech recognition dictionary 102 (step S2005), and determines with the stagnation escape by paraphrase (determination of YES in step S605 of FIG. 6). On the other hand, when the reliability of the re-recognition result is lower than the rejection threshold (NO in step S2004), the stagnation escape determination unit 302 determines that the stagnation escape is not due to paraphrasing (NO determination in step S605 of FIG. 6).

  As a result of the above operation, even if the user utters an abbreviation that is not registered in the speech recognition dictionary 102 in the system specifications and is rejected, if the utterance is expressed with the correct expression in the next utterance and is recognized, the abbreviation uttered last time Since the word is newly registered, it can be recognized from the next time. As a result, the user who utters the abbreviation is not rejected many times, and a smooth dialogue that does not burden the user can be realized. Furthermore, since the creation of abbreviations according to the present embodiment does not prompt special utterance, the burden on the user is small.

(Embodiment 4)
According to the first to third embodiments, the voice recognition parameters are changed and the recognition dictionary is changed by detecting a malfunction and a correct operation in a series of dialogue sequences, thereby enabling personal adaptation. However, since it is not assumed that a plurality of users will use it, when it is used by a plurality of users, it will not be possible to personally adapt correctly. The present embodiment addresses such a case.

  FIG. 22 is a block diagram showing a configuration of a voice interactive information retrieval system including a voice recognition device according to Embodiment 4 of the present invention.

  The difference between the present embodiment and the third embodiment is that the personal adaptation process in the dialog control unit 403 is different due to the addition of the user input unit 401 and the user information storage unit 402. This is the same as in Embodiment 1 to Embodiment 3. Therefore, in the present embodiment, the operation of the dialogue control unit 403 when used by a plurality of users will be described.

  When a user name is input from the user input unit 401, the dialog control unit 403 determines whether or not a speech recognition parameter and a recognition target dictionary adapted to the input user name are registered from the user information storage unit 402. Confirm. If there is no speech recognition parameter or recognition target vocabulary adapted to the input user name, the speech recognition parameter and speech recognition dictionary operate the system using the initial values. If a user who is not adapted to the user detects a malfunction and a correct operation sequence by the stagnation / escape determination unit 302 while using the system, the voice recognition parameter or voice as described in the first to third embodiments is used. When the recognition target vocabulary needs to be changed, the dialogue control unit 403 stores various information about the new user's user name, voice recognition parameters, and words whose dictionary has been changed in the user information storage unit 402.

  On the other hand, when the speech recognition parameter and the recognition target dictionary adapted to the user name input from the user input unit 401 are registered in the user information storage unit 402, the dialogue control unit 403 is previously registered from the user information storage unit 402. The voice recognition parameter after personal adaptation of the user name and the new recognition dictionary registered word are extracted, and the information is registered in the voice recognition parameter storage unit 103 and the voice recognition dictionary 102.

  FIG. 23 is a diagram showing specific examples of various information stored in the user information storage unit 402. Note that the example of FIG. 23 shows a specific example in the case where each word has a rejection threshold. Item 2201 is a user name, item 2202 is a stagnation word, that is, a word whose speech recognition parameter or dictionary is changed, item 2203 indicates the number of stagnation escapes, that is, how many times stagnation escape is detected, and item 2204 is changed The reject threshold, which is the voice recognition parameter, and the item 2205 are newly registered readings added to the voice recognition dictionary.

  When the data shown in FIG. 23 is stored in the user information storage unit 402, when the user name A is input from the user input unit 401, the dialogue control unit 403 displays “next screen” as personal adaptation information of the user name A. "Reject threshold" 3.4 "for new reading registration," Turi no damen "for new reading registration, Reject threshold" 3.5 "for" previous screen ", and" Maeno damen "for new reading registration Registration is performed in the recognition parameter storage unit 103 and the speech recognition dictionary 102.

  From the above operations, it is possible not only to enable individual adaptation of speech recognition parameters and speech recognition dictionaries by detecting malfunctions and correct operations in a series of dialogue sequences, but also when multiple users use the system. Personal adaptation is possible correctly, and personal adaptation and smooth dialogue with less burden on the user can be realized.

  In this embodiment, user identification is performed based on input from the user input unit, and personal adaptation corresponding to a plurality of users is performed. However, speaker identification and speaker identification technology currently generally exist. Therefore, the user may be determined using those techniques.

(Embodiment 5)
In the first to fourth embodiments, the stagnation state caused by an erroneous rejection is targeted as the stagnation state of the system. However, the stagnation state that occurs when the system transitions to the wrong system state due to erroneous recognition is not described. Therefore, this embodiment deals with such a stagnation state.

  As a specific example of the stagnation state of the system that occurs when a transition to the wrong system state due to misrecognition, the user uttered “time search”, but the system recognizes this as “genre search” and the user There is a case where the system state transitions to a system state different from the system state that was expected. At this time, the user utters a command for returning to the original state such as “return” in order to return the erroneous state transition of the system due to the erroneous recognition. When the system status returns, the user speaks “time search” again. This series of operations is a state in which the reciprocation of two system states continues repeatedly, and can be said to be one stagnation state.

  In the present embodiment, there is no change in the system configuration compared to the above-described fourth embodiment, and the difference is the stagnation determination operation processing (flowchart in FIG. 6) in the stagnation escape determination unit 302, and the other is the fourth embodiment. It is the same.

  The operation process of the stagnation / escape determination unit 302 in the present embodiment will be described. FIG. 24 is a flowchart showing a flow of operation of the stagnation escape determination unit 302 in the present embodiment. The past recognition results in the following description are used by referring to the data stored in the dialogue history storage unit 106, and the rephrasing and paraphrase determination are the methods described in the first to fourth embodiments. Do the same.

  First, the stagnation escape determination unit 302 acquires the current speech recognition result (step S2301). Next, it is determined whether or not the voice recognition result is rejected (step S2302). As a result of this determination, if it is determined to be rejected (YES in step S2302), it is determined that it is not stagnation escape, and the process is terminated. On the other hand, if it is determined not to be rejected (NO in step S2302), it is determined whether or not the previous utterance was utterance that returns the state ("return" in the above example) (step S2303). As a result of this determination, if the previous utterance is not an utterance that returns the state (NO in step S2303), it is determined that it is not a stagnation escape, and the process ends. On the other hand, if the previous utterance is an utterance that returns the state (YES in step S2303), it is determined whether or not the current utterance is a restatement of the previous utterance (step S2304). If the result of this determination is rephrasing (YES in step S2304), it is determined that stagnation escape has occurred due to reworking, and the process ends. On the other hand, if it is not rephrased (NO in step S2304), it is determined whether the current utterance is a paraphrase of the previous utterance (step S2305). If the result of this determination is paraphrasing (YES in step S2305), it is determined that the stagnation escape is due to paraphrasing, and the process ends. On the other hand, if it is not a paraphrase (NO in step S2305), it is determined that it is not a stagnation escape, and the process ends. Note that the stagnation escape caused by re-phrase or paraphrase detected in this way is stored in the dialogue history storage unit 106 in distinction from the escape from stagnation caused by an erroneous reject described in the first to fourth embodiments. To do.

  In this way, the determination of escape from stagnation due to misrecognition is performed, and the parameters for speech recognition and the time of recognition are changed. Specifically, for example, the recognition dictionary is changed as described in the second embodiment. More specifically, among the speech recognition results for the previous misrecognized utterances, a result of outputting a character string that is acoustically close without using the speech recognition dictionary (for example, FIG. 4). The recognition result 2) is added to the speech recognition dictionary in association with the correctly recognized word obtained this time.

  Based on the above operations, not only system status stagnation due to erroneous rejection but also personal adaptation using system status stagnation due to misrecognition, so that stagnation due to misrecognition will occur even if the previous malfunctioned utterance is made Therefore, it is possible to realize a smooth dialogue that does not burden the user. In addition, the change of the speech recognition parameters and the speech recognition dictionary according to the present embodiment does not prompt special special utterances, and raises the recognition rate from natural dialogue, so that the burden on the user is small.

  In each of the above embodiments, the speech recognition unit is the speech recognition unit, the stagnation escape determination unit is the stagnation escape determination unit, the dialogue control unit is the dialogue control unit and the change control unit, and the abbreviation creation unit is the abbreviation creation. Corresponds to the means.

  The voice recognition apparatus and the voice recognition method according to the present invention can be used for many systems having a voice interactive interface. For example, in a home information search system, a car navigation system, and information search from a portable terminal. It is useful and its availability is very large.

It is a block diagram which shows the structure of the speech interactive information retrieval system provided with the speech recognition apparatus which concerns on Embodiment 1 of this invention. It is a flowchart which shows the flow of the process of the whole dialog in this invention. It is a figure which shows the example of an output screen of the voice interactive information retrieval system in this invention. It is a figure which shows the example of a recognition result output and stored from the speech recognition part in Embodiment 1 of this invention. It is a figure which shows the example of a recognition result output and stored from the speech recognition part in Embodiment 1 of this invention. It is a flowchart which shows the flow of a process in the stagnation escape determination part in Embodiment 1 of this invention. It is a figure which shows the system operation specification example in the system specification memory | storage part in Embodiment 1 of this invention. It is a figure which shows the example of dialog log | history data memorize | stored in the dialog log | history memory | storage part in Embodiment 1 of this invention. It is a figure which shows the example of an output screen of the voice interactive information retrieval system in this invention. It is a figure which shows the example of an output screen of the voice interactive information retrieval system in this invention. It is a figure which shows the example of an output screen of the voice interactive information retrieval system in this invention. It is a figure which shows the example of an output screen of the voice interactive information retrieval system in this invention. It is a figure which shows the system operation | movement outline in the example of interaction | dialogue of Embodiment 1 of this invention. It is a figure which shows the example of dialog log | history data memorize | stored in the dialog log | history memory | storage part in Embodiment 1 of this invention. It is a figure which shows the example of data which has set the rejection threshold value for every word in Embodiment 1 of this invention. It is a figure which shows the comparison with the case where this technique is not used with the dialog sequence in Embodiment 1 of this invention, when not using. It is a block diagram which shows the other structure of the voice interactive information retrieval system which concerns on Embodiment 1 of this invention. It is a figure which shows the example of dialog log | history data memorize | stored in the dialog log | history memory | storage part in Embodiment 2 of this invention. It is a figure which shows the example of the recognition object vocabulary memorize | stored in the speech recognition dictionary in Embodiment 2 of this invention. It is a block diagram which shows the structure of the speech interactive information retrieval system provided with the speech recognition apparatus which concerns on Embodiment 3 of this invention. It is a flowchart which shows the flow of a process of the paraphrase determination operation | movement by Embodiment 3 of this invention. It is a block diagram which shows the structure of the speech interactive information retrieval system provided with the speech recognition apparatus which concerns on Embodiment 4 of this invention. It is a figure which shows the example of user information data memorize | stored in the user information storage part in Embodiment 4 of this invention. It is a flowchart which shows the process in the stagnation escape determination part in Embodiment 5 of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 101 Voice recognition part 102 Voice recognition dictionary part 103 Voice recognition parameter memory | storage part 104,302 Stagnation escape determination part 105,303,403 Dialog control part 106 Dialog history memory | storage part 107 System specification memory | storage part 108 Database search part 109 Database memory part 110 Response Voice / screen output unit 111 Timer 201 EPG reception unit 301 Abbreviation creation unit 304 User utterance storage unit 401 User input unit 402 User information storage unit

Claims (10)

A speech recognition device for recognizing an input voice and changing a system state, which is a system state related to a dialogue with a user, according to a recognition result, and performing a dialogue.
Voice recognition means for recognizing input voice using a voice recognition dictionary and outputting a recognition result;
Dialog control means for making a response by changing the system state according to a recognition result of the voice recognition means;
If it is determined whether or not the system state has escaped from the stagnation state, which is a state where the system state does not proceed further and is stagnant, and the current recognition result indicates that the system state has escaped from the stagnation state. A stagnation escape determination means for determining whether or not it is at least one of correction and paraphrasing;
When it is determined that the rephrasing or paraphrasing, a change that changes at least one of a rejection threshold as a setting related to dialogue control and a new addition or change to the voice recognition dictionary as a change of setting related to voice recognition A speech recognition apparatus comprising: a control unit. A speech recognition device for recognizing an input voice and changing a system state, which is a system state related to a dialogue with a user, according to a recognition result, and performing a dialogue.
Voice recognition means for recognizing input voice using a voice recognition dictionary and outputting a recognition result;
Dialog control means for making a response by changing the system state according to a recognition result of the voice recognition means;
When it is determined from the current recognition result whether the system state has escaped from the stagnation state that is the same as the system state according to the previous recognition result, and it is determined that the system state has escaped from the stagnation state, Stagnation escape determination means for determining whether the recognition result is at least one of rephrasing and paraphrasing;
When it is determined that the rephrasing or paraphrasing, a change that changes at least one of a rejection threshold as a setting related to dialogue control and a new addition or change to the voice recognition dictionary as a change of setting related to voice recognition A speech recognition apparatus comprising: a control unit. The stagnation state of the system state is a state in which the same system state continues due to rejection of the voice recognition result,
The stagnation escape determination means determines that the current recognition result is the same word as the previous recognition result, and determines that the current recognition result is not the same word as the previous recognition result, but is determined in advance. The speech recognition apparatus according to claim 1, wherein when the recognition word is a recognition word that executes the same system operation, it is determined as a paraphrase. The stagnation state of the system state is a state in which the round trip between two system states continues repeatedly,
The stagnation escape determination means determines that the current recognition result is the same word as the previous recognition result, and determines that the current recognition result is not the same word as the previous recognition result, but is determined in advance. The speech recognition apparatus according to claim 1, wherein when the recognition word is a recognition word that executes the same system operation, it is determined as a paraphrase. The change control means, the speech recognition apparatus according to claim 1 or 2, characterized in that to change and set the threshold value of the reject each recognition target word. The change control means, the reject threshold, and speech recognition apparatus according to newly added or changed to the speech recognition dictionary, to claim 1 or 2, characterized in that to set for each user. The voice recognition device further includes:
When the current recognition result is not the same word as the previous recognition result and is not a recognition word that performs the same predetermined system operation when exiting from the stagnation state, an abbreviation of the current recognition target vocabulary is created Abbreviation creation means to
The voice recognition means re-recognizes the previous recognition result using the abbreviation,
The speech recognition apparatus according to claim 1, wherein the change control unit newly adds the abbreviation to the speech recognition dictionary according to a re-recognition result of the speech recognition unit. An electronic program guide voice recognition apparatus for recognizing an input voice related to an electronic program guide, performing a dialog by changing a system state that is a system state related to a dialog with a user according to a recognition result,
Voice recognition means for recognizing the voice related to the input electronic program guide using a voice recognition dictionary corresponding to the electronic program guide, and outputting a recognition result;
Dialog control means for making a response by changing the system state according to a recognition result of the voice recognition means;
If it is determined whether or not the system state has escaped from the stagnation state, which is a state where the system state does not proceed further and is stagnant, and the current recognition result indicates that the system state has escaped from the stagnation state. A stagnation escape determination means for determining whether or not it is at least one of correction and paraphrasing;
When it is determined that the rephrasing or paraphrasing, a change that changes at least one of a rejection threshold as a setting related to dialogue control and a new addition or change to the voice recognition dictionary as a change of setting related to voice recognition And an electronic program guide voice recognition device. A speech recognition method for recognizing an input voice, changing a system state that is a system state related to a dialog with a user according to a recognition result, and performing a dialog,
A speech recognition step of recognizing input speech using a speech recognition dictionary and outputting a recognition result;
A dialog control step of making a response by changing the system state according to a recognition result in the voice recognition step;
If it is determined whether or not the system state has escaped from the stagnation state, which is a state where the system state does not proceed further and is stagnant, and the current recognition result indicates that the system state has escaped from the stagnation state. A stagnation escape determination step of determining whether or not it is at least one of correction and paraphrasing;
When it is determined that the rephrasing or paraphrasing, a change that changes at least one of a rejection threshold as a setting related to dialogue control and a new addition or change to the voice recognition dictionary as a change of setting related to voice recognition A speech recognition method comprising: a control step. A program for recognizing an input voice, changing a system state that is a state of a system related to a dialogue with a user according to a recognition result, and performing a dialogue,
A speech recognition step of recognizing input speech using a speech recognition dictionary and outputting a recognition result;
A dialog control step of making a response by changing the system state according to a recognition result in the voice recognition step;
If it is determined whether or not the system state has escaped from the stagnation state, which is a state where the system state does not proceed further and is stagnant, and the current recognition result indicates that the system state has escaped from the stagnation state. A stagnation escape determination step of determining whether or not it is at least one of correction and paraphrasing;
When it is determined that the rephrasing or paraphrasing, a change that changes at least one of a rejection threshold as a setting related to dialogue control and a new addition or change to the voice recognition dictionary as a change of setting related to voice recognition A program characterized by causing a computer to execute control steps.

Images