JP2007004054A - Voice interactive device and voice understanding result generation method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve recognition rate by reducing repeated identical incorrect recognition, while leaving the possibility of being adopted as an understanding result. <P>SOLUTION: An understanding result generating part 53 generates an understood result to be the response to a voice uttered, using understood result candidates selected on the basis of understood result score provided to each recognition result candidate, from a plurality of recognition result candidates that are the recognition results of a voice recognition part 52. A cancel button 12 indicates correction to the understood result generated. The understood result score provided to the understood candidate, corresponding to the understood result to which the correction has been indicated, is corrected so in the direction of understood result candidate being less likely to be selected, when the understood result is generated by the understood result generating part 53. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  More particularly, the present invention relates to a speech dialogue apparatus and a speech understanding result generation method for improving a recognition rate of a corrected utterance that is performed once a recognition error is made. .

  2. Description of the Related Art A voice dialogue apparatus has been devised in which a voice uttered by a user is input and a system response is made according to a voice recognition result of the inputted voice, thereby performing a dialogue with the user. In such a voice interaction device, once a recognition error is made, a new voice recognition process can be executed for a corrected utterance made by pressing a cancel button or the like.

In such a voice interaction apparatus, when the voice recognition result for the input voice is erroneously recognized and canceled by the user, the same erroneous recognition is repeated by removing the voice recognition result from the voice recognition target. A technique capable of preventing this is disclosed (Patent Document 1).
JP-A-4-177756

  However, in the technique disclosed in Patent Document 1, a phrase that has been canceled in the past is immediately excluded from the speech recognition target. Therefore, even if the user clearly speaks, the phrase that has been canceled once is re-input. I can't. For example, it was not possible to cope with a case where a word / phrase to be speech-recognized was canceled accidentally, or when a canceled word / phrase needs to be input in a different scene, and lacked flexibility.

  Therefore, the present invention has been proposed in view of the above-described circumstances, and even if the utterance has been canceled once as being misrecognized, the same misrecognition remains with the possibility of being adopted as an understanding result. An object of the present invention is to provide a speech dialogue apparatus and a speech understanding result generation method capable of reducing the repetition of the above.

  The speech dialogue apparatus of the present invention includes an input means for inputting spoken speech, a speech recognition means for recognizing speech input by the input means based on a recognition target word, and a plurality of recognition results by the speech recognition means. An understanding result generating means for generating an understanding result as a response to the spoken speech using the recognition result candidate selected from the recognition result candidates based on a predetermined selection reference value given to each recognition result candidate Correction instruction means for instructing correction of the understanding result generated by the understanding result generation means, and the recognition result candidate corresponding to the understanding result instructed to be corrected by the correction instruction means. Correction means for correcting the predetermined selection reference value in a direction that makes it difficult to select the recognition result candidate when the understanding result generation means generates the understanding result. By providing, to solve the problems described above.

  The speech understanding result generation method of the present invention includes an input step for inputting spoken speech, a speech recognition step for recognizing speech input by the input step based on a recognition target word, and a recognition result by the speech recognition step. An understanding result for generating an understanding result as a response to the spoken speech using the recognition result candidate selected from a plurality of recognition result candidates based on a predetermined selection reference value given to each recognition result candidate A correction instruction step for instructing correction to the understanding result generated by the understanding result generation step; and giving to the recognition result candidate corresponding to the understanding result instructed to correct by the correction instruction step. The predetermined selection reference value is corrected in a direction that makes it difficult for the recognition result candidate to be selected when the understanding result is generated in the understanding result generation step. By providing a positive step, to solve the problems described above.

  The speech dialogue apparatus of the present invention is uttered using a recognition result candidate selected from a plurality of recognition result candidates that are recognition results by the speech recognition means based on a predetermined selection reference value given to each recognition result candidate. An understanding result is generated as a response to the voice. The recognition result candidate is not easily selected when the understanding result generation unit generates the understanding result based on the predetermined selection reference value given to the recognition result candidate corresponding to the understanding result instructed to be corrected by the correction instruction unit. Correct in the direction.

  As a result, it is possible to reduce the erroneous recognition of the speech uttered by the user over and over, and leave the possibility that the canceled recognition result is adopted as the understanding result. Therefore, it is possible to greatly improve the voice recognition rate of the voice uttered by the user.

  Further, the speech understanding result generation method of the present invention uses a recognition result candidate selected from a plurality of recognition result candidates that are recognition results by speech recognition based on a predetermined selection reference value given to each recognition result candidate. Then, an understanding result that is a response to the spoken voice is generated. Then, the predetermined selection reference value given to the recognition result candidate corresponding to the understanding result instructed to be corrected is corrected in a direction that makes it difficult to select the recognition result candidate when generating the understanding result.

  As a result, it is possible to reduce the erroneous recognition of the speech uttered by the user over and over, and leave the possibility that the canceled recognition result is adopted as the understanding result. Therefore, it is possible to greatly improve the voice recognition rate of the voice uttered by the user.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  First, the configuration of the voice interactive apparatus shown as the first embodiment of the present invention will be described with reference to FIG. The voice interaction apparatus shown in FIG. 1 as the first embodiment of the present invention has a configuration when applied to a navigation apparatus mounted on a moving body such as a vehicle. For example, when the navigation device is mounted on a vehicle that is a moving body, the navigation device detects the current position of the vehicle, displays a map corresponding to the current position of the vehicle displayed from the map data, and displays a route to a desired destination. You can give guidance.

  When this voice interactive apparatus is applied to a navigation apparatus, various functions required by the navigation apparatus can be operated interactively by the interaction between the user and the system.

  As shown in FIG. 1, the voice interaction device includes an input device 10, a microphone 20, a memory 30, a disk 40 that stores map data used for route guidance, voice data of guidance voice, and the like, and is stored in the disk 40. A disk reader 41 that reads various data, a controller 50 that recognizes voice input through the microphone 20, understands the contents of the voice recognition result, and generates a system response, and a route search result are shown. A monitor 60 such as a liquid crystal display that displays a map, a menu screen, a voice recognition result by the control device 50, and the like, and a speaker 70 that outputs a guidance voice or a system response voice in a dialog with the user are provided.

  The input device 10 includes a speech recognition start button 11 for instructing start of speech recognition processing for speech uttered by the user and input via the microphone 20 by the user pressing, and a recognition result obtained by the speech recognition processing. And a cancel button 12 for returning the system state to the state before the previous voice input and starting the voice input again when the understanding result generated based on the above is incorrect.

  The microphone 20 inputs the voice uttered by the user to a voice recognition unit 52 of the control device 50 described later. For example, the user utters words and sentences used for the operation of the navigation device, that is, operation commands and proper nouns such as place names, facility names, road names, and sentences including these phrases, and the voice is output from the microphone 20. input.

  The memory 30 is a random accessible RAM (Random Access Memory) or the like, and stores and expands a speech recognition dictionary / grammar read from the disk 40 by the disk reader 41 when a speech recognition process is executed. Area 31, storage area 32 for storing words included in the recognition result candidates obtained by the speech recognition processing and the word reliability, and understanding result and system response at the time of cancellation when the cancel button 12 is pressed And a storage area 33 to be saved as cancel information.

  The disk 40 is a storage medium that stores a speech recognition dictionary / grammar used for speech recognition, a map database, speech data of guidance speech, and the like.

  Generally, in a speech recognition system using a speech recognition dictionary / grammar, only input sentences using the recognition target words and grammar described in the speech recognition dictionary / grammar are accepted as speech recognition results. Can do.

  For example, assuming that the main task of the navigation device is a destination setting for route search, the user can input only words related to facilities such as “Kanagawa Prefecture” and “Yokohama Station” as input sentences input from the microphone 20 by the user. It is possible to envisage both input by sentences combining a plurality of keywords such as “Yokohama Station in Kanagawa Prefecture” and “Yokohama Station on the Tokaido Line”.

  Therefore, the speech recognition dictionary / grammar stored in the disk 40 is configured to be able to handle both such a word-only input and a document including a plurality of keywords.

  Next, the control device 50 will be described. The control device 50 includes an input control unit 51, a speech recognition unit 52, an understanding result generation unit 53, a dialogue control unit 54, a GUI display control unit 55, and a speech synthesis unit 56, and inputs via the microphone 20. Voice recognition processing is performed on the received voice, and a system response is made according to the voice recognition result.

  The input control unit 51 instructs the voice recognition unit 52 to start the voice recognition process in response to the user pressing the voice recognition start button 11. Further, when the user presses the cancel button 12 and an instruction to correct the previous understanding result is given, the input control unit 51 notifies the understanding result generation unit 53 to that effect.

  When the voice recognition process is started, a voice recognition process by the voice recognition unit 52, an understanding result generation process by the understanding result generation unit 53, and a system response output by the dialogue control unit 54 are triggered by the voice input from the microphone 20. This is repeated until one task is completed, such as destination setting and facility search, which are functions of the navigation device. During this time, the user can cancel the previous understanding result by pressing the cancel button 12.

  In response to an instruction from the input control unit 51, the speech recognition unit 52 executes speech recognition processing by capturing a speech signal uttered by a user input from the microphone 20 and digitized by an A / D converter (not shown).

  The voice recognition unit 52 performs voice recognition by matching the captured digitized voice signal with a standby sentence made up of recognition target words held in the voice recognition dictionary / grammar constructed in the storage area 31 of the memory 30. The speech recognition result is output to the understanding result generation unit 53.

  During the matching process, the speech recognition unit 52 calculates a likelihood that is the acoustic proximity between the speech feature data and each standby sentence, and recognizes a speech recognition result that has a likelihood equal to or greater than a certain value. The result is a candidate.

  The speech recognition unit 52 outputs the recognition result generation unit 53 with the top N recognition result candidates (hereinafter also referred to as N-best candidates) of the speech recognition results with the highest likelihood as the recognition result candidates. To do.

  In response to an instruction from the input control unit 51, the speech recognition unit 52 executes speech recognition processing by capturing a speech signal uttered by a user input from the microphone 20 and digitized by an A / D converter (not shown).

  The understanding result generation unit 53 calculates word reliability for each word reading method for all words included in the recognition result candidate output as the speech recognition result from the speech recognition unit 52, and calculates the calculated word reliability. Based on the degree, the correct understanding result for the speech uttered by the user is selected from the recognition result candidates and output to the dialogue control unit 54. The processing content in the understanding result generation unit 53 will be described in detail later.

  Here, the word reliability calculated by the understanding result generation unit 53 will be described. The understanding result generation unit 53 treats words having the same meaning as different words if they are read differently, and calculates the word reliability.

The word reliability is a value indicating the possibility that a word is uttered in a single utterance, and the word reliability of a certain word W is the log likelihood for each of Conf (W) and N-best candidates. If Li is Li, it can be obtained by the following equation (1).

  The word reliability calculated by the understanding result generation unit 53 is stored in the memory 30. Note that calculation of word reliability by the understanding result generation unit 53 is disclosed in Japanese Patent Application Laid-Open No. 2004-251998.

  The dialogue control unit 54 generates a response sentence based on the understanding result output from the understanding result generation unit 53 and outputs the response sentence to the GUI display control unit 55 and the speech synthesis unit 56.

  The GUI display control unit 55 controls the disk reading device 41 to read the map data stored in the disk 40 and display a map via the monitor 60 or generated by the dialogue control unit 54 as necessary. The response content corresponding to the response sentence is displayed via the monitor 60.

  The voice synthesizer 56 synthesizes a digital voice signal corresponding to the response sentence according to the response sentence generated by the dialogue control unit 54, and passes through a D / A converter and an output amplifier (not shown) included in the voice synthesizer 56. To the speaker 70.

  Next, the processing operation from when the speech recognition process by the control device 50 is started until the response sentence is output will be described using the flowchart shown in FIG.

  First, in step S 1, when the navigation device is activated, the control device 50 of the voice interactive device controls the disk reading device 41 to read the voice recognition dictionary / grammar from the disk 40 and store it in the storage area 31 of the memory 30. Let

  Then, when the user presses the voice recognition start button 11 of the input device 10, the input control unit 51 instructs the voice recognition start, and the voice recognition unit 52 enters a voice recognition enabled state. In response to this, the voice recognizing unit 52 starts taking in a voice signal that is spoken by the user and input through the microphone 20 and digitized by an A / D converter (not shown).

  The voice recognition unit 52 continues to calculate the average power of the digitized voice signal (hereinafter also simply referred to as a digital signal) until the voice recognition start button 11 is pressed. After the voice recognition start button 11 is pressed, when the instantaneous power of the digital signal becomes larger than a predetermined value compared to the average power, it is determined that the user has spoken and the digitized voice signal is captured. Be started.

  In step S <b> 2, the voice recognition unit 52 enters a user utterance standby state in response to an instruction to start voice recognition by the input control unit 51. If the user has spoken, the process proceeds to step S3, and if not spoken by the user, the process proceeds to step S6.

  In step S <b> 3, the speech recognition unit 52 waits for the captured digital speech signal and the speech recognition dictionary / grammar constructed in the storage area 31 of the memory 30 in response to the utterance by the user. The speech recognition process is executed by comparing the sentence and calculating the acoustic likelihood.

  The speech recognition unit 52 outputs the top N recognition result candidates with high acoustic likelihoods and the likelihoods to the understanding result generation unit 53 as speech recognition results.

  In step S4, the understanding result generation unit 53 executes an understanding result generation process. The understanding result generation processing by the understanding result generation unit 53 will be described in detail later.

  In step S5, in response to the completion of the understanding result generation process by the understanding result generation unit 53, the dialogue control unit 54 outputs the response sentence to the voice synthesis unit 56 based on the understanding result and the GUI display control unit 55. The response display content to be generated is generated.

  In step S <b> 6, the input control unit 51 detects whether or not the cancel button 12 of the input device 10 has been pressed in the utterance standby state by the user. When the cancel button 12 is pressed in the utterance standby state by the user, the input control unit 51 proceeds to step S7. When the cancel button 12 is not pressed, the input control unit 51 returns to step S2, and waits for the user's utterance. The standby for pressing the cancel button 12 is continued for a predetermined time.

  In step S <b> 7, the input control unit 51 notifies the understanding result generation unit 53 that the cancel button 12 has been pressed in response to the user pressing the cancel button 12. In response to this, the understanding result generation unit 53 stores the understanding result when the cancel button 12 is pressed, the response sentence generated by the dialog control unit 54 in the immediately preceding routine, and the response display content as cancellation information. 30 storage areas 33 are stored.

  In step S8, the voice interactive apparatus returns the system state of the voice interactive apparatus to the previous state. As a result, the content of the immediately preceding utterance is canceled by the user.

  In step S9, the voice interaction apparatus receives the process of step S5 or step S8 and outputs a system response.

  In step S5, the GUI display control unit 55 causes the monitor 60 to display the response display content generated by the dialogue control unit 54. The voice synthesizer 56 synthesizes a digital voice signal corresponding to the response sentence in accordance with the response sentence generated by the dialogue control unit 54, and passes through a D / A converter and an output amplifier (not shown) included in the voice synthesizer 56. To the speaker 70.

  In addition, when step S <b> 8 is performed, the dialogue control unit 54 reads the response sentence and the response display content stored as the cancellation information in the storage area 33 of the memory 30, and outputs them to the speech synthesis unit 56 and the GUI display control unit 55, respectively. .

  The GUI display control unit 55 causes the monitor 60 to display the response display content read from the memory 30 by the dialogue control unit 54. The voice synthesis unit 56 synthesizes a digital voice signal corresponding to the response sentence in accordance with the response sentence read from the memory 30 by the dialogue control unit 54, and a D / A converter (not shown) included in the voice synthesis unit 56, It outputs to the speaker 70 via an output amplifier.

  In step S10, the dialogue control unit 54 determines whether or not the tasks such as facility search and destination setting started in response to the voice recognition start button 11 being pressed are completed. When all tasks are completed, the dialogue control unit 54 ends the speech recognition process, and when the task is continuing, returns to step S1 and resumes voice capture.

  After step S6 to step S8, since the task is not completed, the process returns to step S1 and waits for the next utterance from the user.

  In this way, the voice interaction device executes a process for recognizing the voice uttered by the user, and a system response corresponding to the understanding result generated from the recognition result is output. At this time, by pressing the cancel button 12, the output system response, that is, the understanding result can be canceled.

  Subsequently, an understanding result generation process by the understanding result generation unit 53 will be described. In explaining the understanding result generation processing by the understanding result generation unit 53, an example of dialogue between the user and the voice dialogue device shown below is used as “dialogue example 1”.

Dialogue example 1
First system utterance: “Destination please”
First user utterance: “X × Shinagawa Station of Railway”
Second system utterance: "Are you sure you want to go to Kitagawa Station on the X Train?"
First user operation: Press the cancel button 12 Third system utterance: “Please go to your destination”
Second user utterance: “X × Shinagawa Station of Railway”
Fourth system utterance: “Are you sure you want to go to Shinagawa Station on the X Train?”

  First, the “dialogue example 1” will be described. In the “dialogue example 1”, in response to the question “please have a destination” output as the first system utterance, the user uses “○ × Shinagawa station of the railway” as the first user utterance in order to set the destination. ".

  The voice interactive device misrecognized “○ × Railway Kitagawa Station” in response to the first user utterance, and responded with the second system utterance “Are you sure you want at Kitagawa Station?”

  Therefore, the user presses the cancel button 12 by the first user operation, and cancels the recognition result “○ × Railway Kitagawa Station” by the voice interaction device. As a result, the voice interaction device makes the same response as the first system utterance “Please Destination” as the third system utterance, so that the system state of the apparent voice interaction device is changed to the first user. It will return to the state before the utterance was input.

  And it is an example that the user repeats the same utterance as the first user utterance as “○ × Shinagawa Station of Railway” as the second user utterance.

  FIG. 3 shows N-best which is a recognition result candidate obtained by the speech recognition unit 52 of the speech dialogue apparatus for “○ × Shinagawa Station of Railway” which is the first user utterance in “Dialogue Example 1”. . Further, FIG. 4 shows N-best, which is a recognition result candidate obtained by the speech recognition unit 52 of the speech interactive apparatus for “○ × Shinagawa station of the railway” which is the second user utterance in “dialogue example 1”. Indicates. As shown in FIG. 3 and FIG. 4, the first recognition result candidate in both recognition result candidates is “○ × Kitakawa Station of Railway”.

  As described above, the speech recognition results of the speech recognition unit 52 for the first user utterance and the second user utterance both have the first candidate “○ × Kitakawa Station of Railway”, and the same erroneous recognition is repeated. . In the voice interaction apparatus shown as the embodiment of the present invention, it is possible to avoid such a state of being erroneously recognized repeatedly and to generate an accurate understanding result.

  Based on this “dialogue example 1”, an understanding result generation process by the understanding result generation unit 53 that is the process in step S4 of the flowchart shown in FIG. 2 will be described using the flowchart shown in FIG.

  First, in step S <b> 21, upon receiving the recognition result candidate N-best from the speech recognition unit 52, the understanding result generation unit 53 checks whether the cancel button 12 has been pressed immediately before the current utterance. The understanding result generation unit 53 proceeds to step S22 if the cancel button 12 is not pressed immediately before the current utterance, and proceeds to step S24 if the cancel button 12 is pressed.

  In step S <b> 22, in response to the cancel button 12 not being pressed immediately before the current utterance, the understanding result generation unit 53 starts from the N-best recognition result candidate output from the speech recognition unit 52. The recognition result candidate of the position is taken as the understanding result.

  In step S <b> 23, the understanding result generation unit 53 deletes a past recognition result if it is stored in the storage area 32 of the memory 30. In addition, when the cancel button 12 is pressed before the previous utterance and an understanding result at the time of cancellation is stored in the storage area 33 of the memory 30 as cancel information, the cancel information is also deleted.

  In “Interaction Example 1”, the understanding result generation unit 53 has the highest likelihood in FIG. 3 because the cancel button 12 has not been pressed before the generation of the understanding result for the first user utterance. A high recognition result candidate “X Train Kitagawa Station” is selected as an understanding result.

  In the “dialogue example 1”, the system presses the cancel button 12 in response to the second system utterance “Are you sure you want to do at Kitagawa station on the railway?” The previous understanding result “○ × Railway + Kitakawa Station” is stored as cancellation information in the storage area 33 of the memory 30 (FIG. 2: step S7), and the system state is returned to the previous state (FIG. 2: step S8). Become. Here, the cancel information stored in the storage area 33 of the memory 30 is used when the next recognition result is understood.

  In step S24, the understanding result generation unit 53 determines whether the cancel button 12 has been pressed again immediately before the cancel button 12 is pressed in response to the cancel button 12 being pressed immediately before the current utterance, that is, It is checked whether or not the cancel button 12 is pressed twice in succession.

  In this step S24, when the cancel button 12 is pressed, and the cancel button 12 is pressed twice immediately before the current utterance, the step is performed even immediately after the cancel button 12 is pressed. It progresses to S22 and step S23.

  As described above, when the cancel button 12 is pressed twice in succession without interposing an utterance, the understanding result generation unit 53 determines that the user has reset, and executes when the cancel button 12 is pressed. The processing to be performed is invalidated.

  In addition, when the cancel button 12 is pressed only once instead of twice in succession, the understanding result generation unit 53 saves it in the storage area 33 of the memory 30 by pressing the cancel button 12 in the processing after step S25. The optimum understanding result is derived from the current recognition result while considering the canceled information.

  In the case of accepting the second user utterance in the “dialogue example 1”, since the cancel button 12 is pressed only once immediately before, the processing of steps S25 to S35 is applied.

  In step S25, the understanding result generation unit 53 extracts all words necessary for understanding the meaning excluding particles etc. from the words included in all the recognition result candidates received from the speech recognition unit 52, and their likelihoods. Using the degree, the word reliability for the current recognition result is calculated from the above-described equation (1).

  FIG. 6 shows, as an example, the result of calculating the word reliability of a word included in the recognition result candidate for “○ × Shinagawa Station of Railway” as the second user utterance in “Dialogue Example 1”.

  In step S <b> 26, the understanding result generation unit 53 calculates the word reliability, and then stores words included in the past recognition result candidates and the word reliability as past recognition results in the storage area 32 of the memory 30. Find out if you have.

  At this time, the understanding result generation unit 53 checks whether the cancel button 12 has been pressed immediately before the previous speech recognition, that is, during the previous speech recognition, and whether word reliability has been calculated.

  The understanding result generation unit 53 stores the past saved in the storage area 32 of the memory 30 because the cancel button 12 was not pressed immediately before the previous speech recognition or the cancel button 12 was pressed twice in succession. If the word and the word reliability included in the recognition result candidate are cleared, the process proceeds to step S27.

  Further, when the past recognition result stored in the storage area 32 of the memory 30 remains without being cleared, the understanding result generation unit 53 stores the past recognition result candidate stored in the storage area 32 of the memory 30. The word and its word reliability are set as the word necessary for generating the understanding result in the previous recognition result candidates and its reliability, and the process proceeds to step S28.

  In step S <b> 27, the understanding result generation unit 53 determines that a word included in a past recognition result candidate stored in the storage area 32 of the memory 30 and its word reliability are once cleared, and will be described later in step S <b> 35. , The word reliability of the word included in the previous recognition result candidate saved after clearing the storage area 32 of the memory 30 is calculated.

  In order to calculate the word reliability from the word included in the previous recognition result candidate, the equation (1) described above is used in the same manner as in the case of calculating the word reliability of the word included in the current recognition result candidate in step S25. Use to calculate.

  Then, the understanding result generation unit 53 sets the previous recognition result candidate and the word reliability as the word necessary for generating the understanding result included in the previous recognition result candidate and the word reliability.

  During the understanding result generation process for the first user utterance “XX train Shinagawa station” in “dialog example 1”, the word reliability was not calculated because the cancel button 12 was not pressed. Therefore, in response to the cancel button 12 being pressed by the first user operation, the word reliability of the word newly included in the recognition result candidate of the first user utterance is newly calculated in this step. FIG. 7 shows words included in the recognition result candidates of the first user utterance and the calculated word reliability of the words.

  In step S28, the understanding result generation unit 53 corrects the word reliability of all words included in the obtained recognition result candidates up to the previous time.

  When correcting the word reliability, the understanding result generation unit 53 has a lower word reliability of words included in the previous recognition result candidates than the word reliability of words included in the current recognition result candidates. As described above, correction is performed in such a way that the word reliability of the words included in the recognition result candidates until the previous time is reduced by a certain ratio.

  FIG. 7 described above also shows a correction result obtained by correcting the word reliability of the words included in the recognition result candidates up to the previous time in “dialogue example 1”. Here, correction is made to reduce the word reliability to 60%.

  In step S29, the understanding result generation unit 53, in response to the completion of the word reliability correction, the words included in the recognition result candidates up to the previous time corrected in step S28, and the current time obtained in step S25. The recognition result word list is generated by merging the words included in the recognition result candidates.

  When generating the recognition result word list, the understanding result generation unit 53 can obtain the word reliability of the words that are included in both the previous recognition result candidate and the current recognition result candidate from the previous recognition result. And the word reliability obtained from the current recognition result. For other non-overlapping words, the word reliability of each word is used as it is.

  In step S30, the understanding result generation unit 53 uses the result obtained in step S29 as “a word included in the recognition result candidate up to the previous time and its word reliability” in the next understanding result generation process. The generated recognition result word list is stored in the memory 30.

  FIG. 8 shows a recognition result word list obtained by merging words included in the recognition result candidates of the first user utterance and the second user utterance and their word reliability in “dialogue example 1”.

  In step S31, the understanding result generation unit 53 assigns a new word reliability to the word included in the current recognition result candidate obtained in step S3 of the flowchart shown in FIG. 2, and determines the assigned word reliability. A score for selecting the optimum understanding result from the recognition result candidates is added.

  Specifically, the understanding result generation unit 53 first searches the recognition result word list generated in step S29 for the same word as the word included in the current recognition result candidate. And the word reliability matched with the word searched from the recognition result word list | wrist is acquired, and this word is allocated with respect to the word contained in this recognition result candidate.

  Furthermore, after assigning the word reliability acquired from the recognition result word list to the words included in the current recognition result candidate, the understanding result generation unit 53 adds the word reliability assigned to each recognition result candidate. By doing so, the above-mentioned score is calculated. Hereinafter, this score is set as the understanding result score, and the recognition result candidate is set as the understanding result candidate when the understanding result score is obtained.

  FIG. 9 shows an example of an understanding result candidate and an understanding result score in the second user utterance of “conversation example 1”. For example, as shown in FIG. 9, if the understanding result candidate is “◯ × railway + Kitakawa station”, the word reliability of 1.32 is assigned to “鉄 道 × railway” from the recognition result word list shown in FIG. It is possible to set a word reliability of 0.74 for “Kitakawa Station” and add them together to obtain an understanding result score of 2.06.

  In step S <b> 32, the understanding result generation unit 53 compares each of the understanding result scores with different word counts, so that the understanding result scores can be compared with each other according to the number of words included in each understanding result candidate. Correct. Various correction methods are conceivable for correcting the understanding result score, and there is a method of dividing the understanding result score by a correction value corresponding to the number of words.

  For example, when the understanding result candidate includes two words in the understanding result candidate, the word reliability for two words is added to the understanding result score. Therefore, by dividing the understanding result score by the correction value 1.6, it becomes possible to compare with the understanding result score of the understanding result candidate of only one word.

  FIG. 9 described above also shows a corrected score obtained by correcting the understanding result score of each understanding result candidate. The corrected score shown in FIG. 9 is a value obtained by dividing each understanding result score by the correction value 1.6.

  Such an understanding result candidate including the understanding result score and the corrected score of the understanding result score is collectively used as an understanding result candidate list.

  In step S33, the understanding result generation unit 53 corrects the corrected score of the understanding result candidate according to the understanding result canceled in the past to obtain a corrected score.

  For example, the understanding result generation unit 53 searches the understanding result candidate list to determine whether there is an understanding result candidate that matches the information canceled in the past, and corrects the corrected score of the understanding result candidate. Is corrected so as to lower the value of the corrected score, and the corrected score is obtained. The process of correcting the corrected score and obtaining the corrected score will be described in detail later.

  In step S34, the understanding result generation unit 53 selects the understanding result candidate having the maximum corrected score as the optimum understanding result.

  In step S35, the understanding result generation unit 53 stores the current recognition result candidate obtained by the speech recognition unit 52 in the storage area 32 of the memory 30 in step S3 of the flowchart shown in FIG. finish.

  At this time, when passing through steps S22 and S23, only the current recognition result candidate is stored in the storage area 32 of the memory 30 as the past recognition result. Further, when passing through steps S25 to S34, the current recognition result candidate and its word reliability are additionally stored in the storage area 32 of the memory 30 as past recognition results.

  In this way, in the voice interactive apparatus shown as the embodiment of the present invention, when the word included in the recognition result once canceled by pressing the cancel button 12 is included in the understanding result candidate, the user speaks It is modified so that the value of the corrected score is lowered because it is unlikely that it matches the received voice, and the possibility of being selected as the final understanding result is reduced.

  As a result, it is possible to reduce the erroneous recognition of the speech uttered by the user over and over, and leave the possibility that the canceled recognition result is adopted as the understanding result. Therefore, the voice recognition rate of the voice uttered by the user can be greatly improved.

  Next, several techniques for correcting the corrected score of the canceled understanding result candidate in step S33 of the flowchart shown in FIG. 5 will be described.

(Correction of corrected score: correction according to the degree of coincidence)
First, a method for correcting the corrected score according to the degree of coincidence between past understanding results stored in the storage area 33 of the memory 30 as cancellation information and the understanding result candidates in the understanding result candidate list will be described.

  The understanding result generation unit 53 reads past understanding results stored as cancellation information in the storage area 33 of the memory 30, and is greater than 0 and greater than 1 according to the degree of coincidence with the understanding result candidates in the understanding result candidate list. A correction coefficient is selected by selecting a small correction coefficient (0 <correction coefficient <1) and multiplying the corrected score by the selected correction coefficient.

  The understanding result generation unit 53 compares the current understanding result candidate with the previous understanding result at the time of cancellation, and, for example, when both are completely matched (complete matching), the understanding result candidate is Cases are classified into three cases: when they are included in an understanding result canceled in the past (correction⊃understanding result) and when an understanding result canceled in the past is included in an understanding result candidate (correction⊂understanding result). Depending on the case, different correction factors are selected.

  This means that it is unlikely that the content that completely matches the understanding canceled in the past will be re-entered, that is, the “complete match” is the lowest, followed by a portion of the understanding that was canceled in the past. The content that is to be entered again, that is, the “correction⊃understanding result” is unlikely to be entered, and the content that includes all of the previously canceled understanding content is entered again, that is, the “correction⊂understanding result” is possible. Is based on the highest.

  For example, as shown in FIG. 10, different correction coefficients are used depending on the degree of coincidence between the current understanding result candidate and the past understanding result at the time of cancellation. FIG. 11 shows an example of a combination of the understanding result at the time of cancellation and the current understanding result, which are “complete match”, “correction / understanding result”, and “correction / understanding result” in “dialogue example 1”.

  Next, the processing operation for correcting the corrected score of the canceled understanding result candidate in step S33 of the flowchart shown in FIG. 5 will be described using the flowchart shown in FIG.

  First, in step S41, the understanding result generation unit 53 extracts one understanding result candidate from the understanding result candidate list.

  In step S42, the understanding result generation unit 53 reads the past understanding result stored as the cancellation information in the storage area 33 of the memory 30, and determines whether or not it completely matches the understanding result candidate extracted from the understanding result candidate list. Investigate. The understanding result generation unit 53 proceeds to step S43 if they match, and proceeds to step S44 if they do not match.

  In step S43, in response to the fact that the understanding result candidate extracted from the understanding result candidate list and the past understanding result completely match, the understanding result generation unit 53 sets “complete match” to the corrected score of this understanding result candidate. The corrected score a is multiplied to calculate a corrected score.

  As shown in FIG. 11, in “Dialogue Example 1”, when the understanding result at the time of cancellation is “◯ × Railway + Kitakawa Station” and the current understanding result candidate is also “◯ × Railway + Kitakawa Station”, This understanding result candidate is considered to be “perfectly matched” with the understanding result at the time of cancellation.

  As shown in FIG. 10, in the case of “perfect match” in this way, 0.5 is selected as the correction coefficient. Therefore, the corrected score of the current understanding result candidate “○ × Railway + Kitagawa Station” The corrected score can be obtained by multiplying the correction coefficient 0.5.

  In “Dialogue Example 1”, if “Kitakawa Station” exists only as a subcategory of “XX Railway”, “XX Railway + Kitakawa Station” and “Kitakawa Station” are semantically identical. It is assumed that it is “perfect match”.

  In the “dialogue example 1”, the understanding result when canceled by the first user operation is “◯ × railway + Kitakawa station”. As for the current understanding result candidate shown in FIG. 9, the understanding result at the time of cancellation and the understanding result at this time completely coincide with each other. Therefore, the corrected score of “◯ × Railway + Kitagawa Station” is 0.65 by multiplying the corrected score 1.29 by the correction coefficient 0.5 by cancellation. As shown in FIG. 9, the corrected score is described in the understanding result candidate list.

  In step S44, the understanding result generation unit 53 checks whether or not the understanding result at the time of cancellation includes all the understanding result candidates extracted from the understanding result candidate list (correction understanding result).

  In step S45, the understanding result generation unit 53 determines that the correction result of the understanding result candidate is “corrected” when the understanding result at the time of cancellation includes all the understanding result candidates extracted from the understanding result candidate list. The corrected score b is multiplied by the “understanding result” to calculate a corrected score.

  As shown in FIG. 11, when the understanding result at the time of cancellation is “○ × railway + Kitakawa Station” and the current understanding result candidate is “○ × railway”, the understanding result at the time of cancellation is all the understanding result candidates. Because it is included, it is regarded as “correction ⊃ understanding result”.

  In this way, in the case of “correction ⊃ understanding result”, as shown in FIG. 10, 0.7 is selected as the correction coefficient, so the corrected score of the current understanding result candidate “○ × Railway” is A corrected score can be obtained by multiplying the correction coefficient 0.7.

  In step S46, the understanding result generation unit 53 checks whether the understanding result candidates extracted from the understanding result candidate list include all of the understanding results at the time of cancellation (correction habit understanding results).

  In step S47, the understanding result generation unit 53 adds “correction” to the corrected score of the understanding result candidate according to the fact that the understanding result candidate extracted from the understanding result candidate list includes all the understanding results at the time of cancellation. The corrected score is calculated by multiplying the correction coefficient c in the case of “⊂ understanding result”.

  As shown in FIG. 11, when the understanding result at the time of cancellation is “○ × Railway” and the current understanding result candidate is “○ × Railway + Kitakawa Station”, the understanding result candidate is the understanding result at the time of cancellation. Because it includes everything, it is considered to be a “correction / understanding result”.

  Thus, in the case of “correction ⊂ understanding result”, as shown in FIG. 10, 0.9 is selected as the correction coefficient, so that the current understanding result candidate “○ × railway + Kitagawa station” is corrected. The score can be corrected by multiplying by a correction coefficient of 0.9 to obtain a corrected score.

  In step S48, the understanding result generation unit 53 ends the score correction process in response to the completion of checking the degree of coincidence with the understanding result at the time of cancellation for all the understanding result candidates.

  Among the understanding result candidates in the understanding result candidate list shown in FIG. 9, only “○ × railway + Kitagawa station” matches the understanding result at the time of cancellation shown in FIG. The corrected score of the result candidate is not corrected, and the corrected score becomes the corrected score as it is.

  As a result, the corrected scores of “○ × Railway + Kitakawa Station” and “○ × Railway + Shinagawa Station”, which are candidates for understanding, are reversed from the corrected score shown in FIG. 9, and “○ × Railway + Shinagawa Station” is reversed. The corrected score is maximized. Therefore, in step S34 in the flowchart shown in FIG. 5, the understanding result generation unit 53 selects “選 択 × railway + Shinagawa station”, which is the maximum corrected score, from the understanding result candidates as the understanding result.

  As a result, the voice recognition unit 52 does not recognize the same error for the first user utterance and the second user utterance in the “dialogue example 1”, but “○ × railway” in the fourth system utterance. Is it OK at Shinagawa Station? "

  In this way, the canceled understanding result can be correctly reflected in the correction coefficient by changing the correction coefficient according to the degree of coincidence when the current understanding result candidate and the understanding result at the time of cancellation are compared. The corrected score can be corrected accurately.

  In the above-described example, by preparing three different correction coefficients according to the degree of coincidence when the current understanding result candidate and the understanding result at the time of cancellation are compared, the canceled understanding result is reflected in the correction coefficient. It was. Furthermore, from the viewpoint other than the degree of coincidence, the understanding result at the time of cancellation can be reflected in the correction coefficient.

(Correction score after correction: Correction based on the number of times the understanding result was canceled)
First, a method of correcting the corrected score according to the number of cancellations for the understanding result will be described.

  For example, when the understanding result is canceled a plurality of times, the correction coefficient can be changed according to how many times the target understanding result has been canceled before. This method can also be used in combination with the above-described method of changing the correction coefficient according to the degree of coincidence.

  FIG. 13 shows a correction coefficient when a parameter indicating how many times the understanding result was canceled in addition to the degree of coincidence between the understanding result candidate and the understanding result at the time of cancellation is added as a parameter for changing the correction coefficient. An example is shown.

  For example, when the correction coefficient shown in FIG. 13 is used, if the understanding result when the cancel button 12 is pressed once and the current understanding result candidate completely match, the corrected score is multiplied by 0.5. Make corrections. If the understanding result when the cancel button 12 is pressed twice before and the current understanding result candidate completely coincide, the corrected score is multiplied by 0.6 and corrected. Further, if the understanding result when the cancel button 12 is pressed three times before and the current understanding result candidate completely match, the corrected score is corrected by 0.7.

  Thus, the older the information that the understanding result is “cancelled”, that is, in accordance with the decrease in the influence on the understanding result due to the cancellation, 0 <d (d: correction coefficient) <1 In the range, increase the correction coefficient and decrease the rate of lowering the corrected score.

  Specifically, in the storage area 33 of the memory 30, information obtained by counting the number of times each time another understanding result is canceled is stored together with the past understanding result stored as cancellation information.

  When comparing the understanding result candidate in the understanding result candidate list with the understanding result at the time of cancellation, reference is made to the information obtained by counting the number of cancellations added to the understanding result at the time of cancellation. If the influence is small and the number of cancellations is small, it can be determined that the influence on the understanding result candidate is large.

  Therefore, by changing the correction coefficient according to this, the canceled understanding result can be reflected in the correction coefficient, so that the corrected score can be corrected accurately.

(Correction score correction process: correction according to the elapsed time since the understanding result was canceled)
Next, a method for correcting the corrected score according to the time required from the input of the information that the information is canceled until the input of the utterance that is currently under the understanding result generation process will be described.

  This method can be realized by changing the correction coefficient according to the time required from the input of the information indicating that the cancellation has been performed until the input of the utterance currently being processed for the generation of the understanding result.

  This will be described with reference to a timing chart showing the timing of the user's utterance and the system response by the system shown in FIG. In the timing chart shown in FIG. 14, the user speaks “Shinagawa Station” at time T1, and the system misrecognizes that it is “Kitakawa Station”. Is the system response?

  In response to this, the user presses the cancel button 12 at time T3, and re-enters “○ × Shinagawa station of railway” at time T5. Here, the system recognizes the input speech using the recognition result and the cancellation information so far, but misrecognizes it again. At time T6, the system response “Are you sure you want to be at Tachikawa Station on the railway?” Is going.

  Therefore, the user presses the cancel button 12 again at time T7 and inputs “Shinagawa station” again at time T9. In response to this, the system derives an optimal understanding result using the recognition result and the cancellation information so far, and performs a system response “Are you sure at Shinagawa Station?” At time T10.

  In such a dialogue example between the user and the system shown in FIG. 14, for the user utterance input at time T9, the understanding result generation unit 53 causes the understanding result “north side” by the system canceled at time T3. It is necessary to determine whether or not the station “and the understanding result“ ○ × Tachikawa Station of the railway ”by the system canceled at time T7 matches the current understanding result candidate, and correct the corrected score.

  At this time, the understanding result generation unit 53 changes the correction coefficient according to the time required from the input of the canceled utterance to the input of the utterance for which the current understanding result is generated.

  For example, in the example illustrated in FIG. 14, the time required from the input of the canceled utterance to the input of the utterance for which the current understanding result generation is performed is the time at which the canceled understanding result “Kitakawa At time Tx1 from time T1 when the input of the utterance leading to “station” is started to time T9 when the input of the utterance currently under the understanding result generation process is started, and at time T7, the understanding result “○ The time Tx2 from the time T5 when the input of the utterance leading to “Tachikawa station of the railway” is started to the time T9 when the input of the utterance currently being processed for the understanding result generation is started.

  The understanding result generation unit 53 is in proportion to the time required from the input of the canceled utterance to the input of the utterance currently being processed for the understanding result generation, 0 <d (d: correction coefficient) In the range of <1, the correction coefficient is increased. That is, the longer the time from the input of the canceled utterance to the input of the utterance that is currently undergoing the understanding result generation process, the larger the correction coefficient, the smaller the influence.

Here, when the correction coefficient is d and the time from when the canceled utterance is input until the current input is started is T, the correction coefficient can be expressed by the following equation (3).

  When Tx1 and Tx2 are set to 40 seconds and 20 seconds, respectively, and the expression (3) is applied to the interactive example shown in FIG. 14, the correction coefficient d1 and the correction coefficient d2 corresponding to Tx1 and Tx2 are as follows: Can be sought.

d1 = 0.02 × 40 = 0.8
d2 = 0.02 × 20 = 0.4
When the corrected score is corrected and the corrected score is obtained, the correction coefficient is multiplied by the corrected score. Therefore, the smaller the correction coefficient is, the larger the ratio of correction to the corrected score is.

  Therefore, the current understanding result candidate is the understanding result obtained from the utterance input at time T5, rather than the case where the understanding result candidate matches the “Kitakawa station” which is the understanding result obtained from the utterance input at time T1. The ratio of correcting the corrected score is greater when it matches “Ox Train Tachikawa Station”.

  FIG. 15 shows an understanding result candidate for the third user utterance in the example of interaction shown in FIG. 14, its understanding result score, a corrected score, a correction coefficient, and a corrected score.

  Of the understanding result candidates shown in FIG. 15, the first candidate “○ × Railway + Tachikawa Station” and the second candidate “Kitakawa Station” are misrecognized as shown in the example of FIG. “Shinagawa Station” is a correct result that matches the utterance input by the user.

  The understanding result score and the corrected score in FIG. 15 are values obtained respectively in step S31 and step S32 of the flowchart of FIG. 5 described above. Also, among the understanding result candidates shown in FIG. 15, the correction coefficient d2 = 0.4 and the correction coefficient d1 = 0 described above for “○ × railway + Tachikawa station” and “Kitakawa station” canceled in the past, respectively. Multiply by .8 to get a modified score.

  As shown in FIG. 15, the corrected score, which is the score before correction, is larger for “○ × Railway + Tachikawa Station” than for “Kitakawa Station”. It can be seen that “○ × Railway + Tachikawa Station” which was canceled at has a smaller score than “Kitakawa Station”. However, since the corrected score of “Shinagawa Station” that was not corrected by cancellation is larger than the corrected score of other understanding result candidates, “Shinagawa Station” is selected as the final optimal understanding result. It will be.

  In this way, the correction coefficient is changed according to the time required from the time when the voice that leads to the canceled understanding result is uttered and inputted to the time when the voice that is currently undergoing the understanding result generation process is spoken and inputted. Since the canceled understanding result can be correctly reflected in the correction coefficient, the corrected score can be corrected accurately.

(Correction score after correction: correction by task sequence)
It is also possible to change the correction coefficient depending on where the current system state is in the task sequence.

  For example, in the case of destination setting that searches for facilities by following the tree structure hierarchically, the final item is selected after selecting the upper item and confirming its contents for several upper items. There are cases.

  For example, when a restaurant is selected near the current location, it is assumed that there are “Japanese food” and “Chinese” as selectable genres. At that time, once select “Japanese food” and confirm the contents, cancel “Japanese food” and select “Chinese”, then confirm the contents of “Chinese”, but also cancel “Chinese” after all There is a case where “Japanese food” is selected again.

  In such a case, “Japanese food” has been canceled once, but it was not a misrecognition. In this way, when an item in the middle of a hierarchically tree structure is canceled in the past, it is necessary to consider that it may have been canceled even if it was a correct answer rather than a misrecognition.

  However, if you select the end of the tree structure, for example, a store name such as “XX Sushi”, it is very unlikely that it will be canceled once it is selected. It is necessary to think that there was a high possibility.

  That is, when the spoken dialogue apparatus executes a task sequence that requests utterances in stages and realizes a desired task, the correction coefficient when the end of the tree structure in the task sequence is canceled is 0 <d (d : Correction coefficient) In a range of <1, correction is performed in such a way that the correction factor is increased so that the item in the middle of the tree structure is smaller than that in the case where the item is canceled, that is, the correction ratio is increased to make it difficult to select again. .

  As a result, the canceled understanding result can be correctly reflected in the correction coefficient, so that the corrected score can be corrected accurately.

  For example, in a task sequence that realizes a task by following a menu of a tree structure, when selecting an upper item having a lower selection item, the influence of cancellation is reduced, and when selecting a lower lower selection item Can increase the influence of cancellation. Therefore, in the menu operation of the tree structure, in the case of selecting the upper item after confirming what the lower item is under each upper item, if it is a higher item even if it is canceled once, It can be easily recognized again at the next confirmation.

(Correction of corrected score: correction according to the time required until cancel operation)
It is also possible to correct the correction coefficient according to the time required from the input of the spoken voice to the generation of an understanding result for the voice and the canceling operation.

  Specifically, when the time from the input of the canceled utterance to the cancel operation is short, the correction coefficient is corrected so that the influence of the canceled understanding result is increased, and the canceled utterance is input. If the time until the cancel operation is long, the correction coefficient is corrected so that the influence of the canceled understanding result is reduced.

  As a result, in the tree-structured menu as described above, when the lower item is selected after confirming what lower item is under each upper item, the user does not show interest in the lower item. If the cancellation is immediately performed without taking time after the understanding result is output, the higher-level item is less likely to be recognized again at the next recognition. On the contrary, if it takes time to display the lower item for a long time and confirm the contents, the upper item is likely to be recognized again at the next recognition.

(Correction score correction processing: correction according to the difference in input method of cancellation instructions)
Further, in the voice interactive apparatus shown as the embodiment of the present invention, when canceling the voice recognition result, the cancel button 12 is pressed to input the cancellation, but for example, voice input such as a negative word is input. To cancel the speech recognition result. In addition, the user may input that the speech recognition result is canceled by pressing the cancel button 12, inputting by voice, or using another method.

  As described above, when a plurality of means for notifying that the speech recognition result is canceled are provided, the correction coefficient described above can be changed according to the difference of the canceling means.

  For example, misrecognition due to voice input occurs more frequently than erroneous input due to a mistake in pressing at the time of switch input. Therefore, if the input of canceling the voice recognition result is input by voice input, the reliability is considered to be low. .

  Therefore, in the case of inputting cancellation to be performed by voice recognition by voice input, the range of 0 <d (d: correction coefficient) <1 is greater than in the case of inputting cancellation of the voice recognition result by pressing the cancel button 12. Thus, by increasing the correction coefficient and decreasing the rate of correcting the corrected score, even uncertain cancellation information by voice input can be used for generating an understanding result.

  As a result, the canceled understanding result can be correctly reflected in the correction coefficient, so that the corrected score can be corrected accurately.

  As described above, the correction score correction processing method described above can accurately reflect the canceled understanding result in the correction coefficient, regardless of whether they are used independently or in any combination. It becomes possible to correct the corrected score accurately. Therefore, since the speech uttered by the user is recognized at a high recognition rate, a more accurate understanding result can be generated.

  In the above description, the word reliability of the word included in the recognition result candidate is obtained, and an understanding result score that is a criterion for selecting an optimum understanding result from the understanding result candidate is calculated from the word reliability. When the understanding result is canceled, the corrected result score is corrected based on the number of words of the understanding result candidate. The selected understanding result reflects the influence of the canceled understanding result.

  In the present invention, in selecting a final understanding result from recognition result candidates that are speech recognition results by the speech recognition unit 52, it is not always necessary to obtain an understanding result score as a reference value for selection, and words included in the recognition result candidate An understanding result may be obtained using the word reliability as a reference value for selection, or an understanding result may be obtained using the acoustic likelihood of a recognition result candidate as the reference value for selection.

  Therefore, in such a case, in order to reflect the influence of the canceled understanding result on the finally selected understanding result, not the score correction to the understanding result score but the word reliability correction or the acoustic likelihood. A correction will be made.

  Specifically, the word reliability obtained for the word included in the recognition result candidate when obtaining the above-described understanding result score is less likely to be selected as the final understanding result according to the canceled understanding result. Will be corrected in the direction. Then, the word having the highest word reliability corrected according to the canceled understanding result is selected as the final understanding result.

  In the case of acoustic likelihood, the likelihood of the recognition result candidate obtained by the speech recognition processing by the speech recognition unit 52 is less likely to be selected as the final understanding result according to the canceled understanding result. Make corrections in the direction. And the recognition result candidate with the highest acoustic likelihood corrected according to the canceled understanding result is selected as the final understanding result.

  As described above, in the case of generating the understanding result based on the word reliability, in the case of generating the understanding result based on the acoustic likelihood, in any case, the canceled understanding in step S33 of the flowchart shown in FIG. All methods of correcting the understanding result score for the result candidate can be applied, and by using this method to correct the word reliability or acoustic likelihood, the canceled understanding result can be reflected more accurately. Can produce final understanding results.

  However, when the word reliability is obtained, and further, when the understanding result score is obtained, it is possible to determine the utterance possibility of each word based on the same criterion, so that a more accurate understanding result is generated. There is an advantage in that it can be.

  The above-described embodiment is an example of the present invention. For this reason, the present invention is not limited to the above-described embodiment, and various modifications can be made depending on the design and the like as long as the technical idea according to the present invention is not deviated from this embodiment. Of course, it is possible to change.

It is a figure for demonstrating the structure of the voice interactive apparatus shown as embodiment of this invention. 6 is a flowchart for explaining a processing operation from the start of speech recognition processing to the output of a response sentence in the voice interactive device. It is the figure which showed an example of the recognition result candidate, and its likelihood. It is the figure which showed an example of the recognition result candidate, and its likelihood. It is a flowchart for demonstrating an understanding result production | generation process. It is the figure which showed an example of the word contained in a recognition result candidate, and its word reliability. It is the figure which showed an example of the word contained in a recognition result candidate, its word reliability, and corrected word reliability. It is the figure which showed an example of the recognition result word list. It is a figure showing an example of an understanding result candidate and its understanding result score, a corrected score, and a corrected score. It is the figure which showed the relationship between a coincidence degree and a correction coefficient. It is the figure which showed an example of the degree of agreement with the understanding result at the time of cancellation, and this understanding result. It is a flowchart for demonstrating the processing operation | movement which corrects the score after correction of the canceled understanding result candidate according to a coincidence degree. It is the figure which showed the correction coefficient which changes according to a coincidence degree and how many times it was canceled before. It is a figure for demonstrating correction of the score after correction | amendment according to time. It is the figure which showed an example of an understanding result candidate based on the example of interaction shown in FIG. 14, its understanding result score, a corrected score, a correction coefficient, and a corrected score.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Input apparatus 12 Cancel button 20 Microphone 30 Memory 50 Control apparatus 51 Input control part 52 Speech recognition part 53 Understanding result generation part 54 Dialog control part

Claims (11)

Input means for inputting spoken speech;
Speech recognition means for recognizing speech input by the input means based on recognition target words;
A response to the spoken speech using the recognition result candidate selected from a plurality of recognition result candidates as recognition results by the speech recognition means based on a predetermined selection reference value given to each recognition result candidate; An understanding result generating means for generating an understanding result,
Correction instruction means for instructing correction for the understanding result generated by the understanding result generation means;
When the understanding result generation unit generates the understanding result, the recognition result candidate corresponding to the predetermined recognition criterion value given to the recognition result candidate corresponding to the understanding result instructed to be corrected by the correction instruction unit. And a correcting means for correcting in a direction that makes it difficult to select. The correction unit is configured to recognize the recognition result candidate according to an acoustic coincidence between the understanding result instructed to be corrected by the correction instruction unit and the plurality of recognition result candidates that are recognition results by the speech recognition unit. The spoken dialogue apparatus according to claim 1, wherein the predetermined reference value is corrected. The correction means corresponds to the understanding result instructed to be corrected by the correction instructing means depending on how many times the understanding result instructed to correct by the correction instructing means The spoken dialogue apparatus according to claim 1, wherein a predetermined reference value of the recognition result candidate is corrected. The correcting means utters a voice that leads the understanding result instructed to be corrected by the correction instructing means and is input to the input means, and then speaks a voice that is currently being processed to generate an understanding result and is input by the input means The voice according to claim 1, wherein a predetermined reference value of the recognition result candidate corresponding to the understanding result instructed to be corrected by the correction instructing unit is corrected according to a time required until the correction is performed. Interactive device. When the spoken dialogue device executes a task sequence that requests utterances step by step and realizes a desired task,
The correction means determines the predetermined one of the recognition result candidates corresponding to the understanding result instructed to be corrected by the correction instruction means depending on at which stage of the task sequence correction is instructed by the correction instruction means. The spoken dialogue apparatus according to claim 1, wherein the reference value is corrected. The correction means is necessary until the understanding result generated by the speech recognition means and the understanding result generation means is instructed to be corrected by the correction instruction means after the speech is spoken and input to the input means. 2. The spoken dialogue apparatus according to claim 1, wherein a predetermined reference value of the recognition result candidate corresponding to the understanding result instructed to be corrected by the correction instructing unit is corrected according to the corrected time. A plurality of the correction instruction means different in the way of inputting the correction instruction by the user,
The correction means corrects a predetermined reference value of the recognition result candidate corresponding to the understanding result instructed to be corrected by the correction instruction means in accordance with a method of inputting the correction instruction to the correction instruction means by a user. The spoken dialogue apparatus according to claim 1, wherein: A word reliability calculation means for calculating the word reliability that indicates the possibility that the word included in the recognition result candidate is spoken and becomes the predetermined selection reference value;
The understanding result generation means is the recognition result selected based on the word reliability calculated by the word reliability calculation means from the words included in a plurality of recognition result candidates that are recognition results by the speech recognition means. Using the words included in the candidates, generate an understanding result that is a response to the spoken speech,
The correction means generates the word reliability of the word included in the recognition result candidate corresponding to the understanding result instructed to be corrected by the correction instruction means when the understanding result generation means generates the understanding result. The spoken dialogue apparatus according to any one of claims 1 to 7, wherein the word included in the recognition result candidate is corrected in a direction in which the word is difficult to be selected. A new value is given to the word included in the recognition result candidate based on the word reliability calculated by the word reliability calculation means and the word reliability given to the word included in the past recognition result candidate. Score calculation means for calculating a score of the recognition result candidate that is the predetermined selection reference value from the given word reliability,
The understanding result generation means is uttered by using the recognition result candidate selected based on the score calculated by the score calculation means from a plurality of recognition result candidates that are recognition results by the voice recognition means. Generate an understanding result that is a response to the voice,
The correction means selects the score of the recognition result candidate corresponding to the understanding result instructed to be corrected by the correction instructing means, when the recognition result candidate is selected by the understanding result generating means. The spoken dialogue apparatus according to claim 8, wherein correction is made in a direction in which it is difficult to be performed. A likelihood calculating means for calculating a likelihood indicating acoustic closeness to the recognition target word of the recognition result candidate;
The understanding result generation means uses the recognition result candidate selected based on the likelihood calculated by the likelihood calculation means from a plurality of recognition result candidates that are recognition results by the voice recognition means, and uses the recognition result candidate. Generate an understanding result that is a response to
The correction means is configured to determine the likelihood of the recognition result candidate corresponding to the understanding result instructed to be corrected by the correction instructing means when the recognition result candidate is generated when the understanding result generating means generates the understanding result. The spoken dialogue apparatus according to any one of claims 1 to 7, wherein correction is made in a direction in which selection becomes difficult. An input process for inputting spoken voice;
A speech recognition step for recognizing the speech input by the input step based on a recognition target word;
A response to the spoken speech using the recognition result candidates selected from a plurality of recognition result candidates that are recognition results by the speech recognition step based on a predetermined selection reference value given to each recognition result candidate; An understanding result generation step for generating an understanding result,
A correction instruction step for instructing correction to the understanding result generated by the understanding result generation step;
When the predetermined selection reference value given to the recognition result candidate corresponding to the understanding result instructed to be corrected by the correction instruction step is generated in the understanding result generation step, the recognition result candidate A speech understanding result generation method comprising: a correction step of correcting in a direction that makes it difficult to select.

Images