JP2007183516A - Voice interactive apparatus and speech recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice interactive apparatus capable of reducing user's load at correction utterance. <P>SOLUTION: When a first collation result, obtained by collating a first utterance voice by using a first language model including a word to be subjected to speech recognition, is not adopted, a signal processor 14 in the voice interactive device collates a second utterance speech by using a second language model including other words for expressing the attribute of the word, and the first utterance voice is subjected to the voice recognition again, by using the language model restricted by the obtained second collation result, and a system response is generated according to the voice recognition result. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a voice dialogue apparatus that performs a dialogue according to spoken voice, and a voice recognition method that recognizes spoken voice.

  Conventionally, a navigation device has been proposed that recognizes the facility name of the destination spoken by the user using a voice dialogue device that performs dialogue according to the spoken voice and provides route information corresponding to the recognized destination. (See, for example, Patent Document 1).

In this type of navigation device, the user can input a destination by speaking only the facility name in the prefecture where the vehicle is present. In the navigation apparatus, for facilities in other prefectures, the user can input a destination or the like by continuously speaking the name of the attribute related to the facility such as the prefecture name and the facility name. Furthermore, in the navigation device, even if the input destination is misrecognized, the destination can be obtained by performing a so-called corrected utterance (rephrased utterance) such as when the user inputs a correction instruction and speaks again. Can be re-entered.
Japanese Patent Laid-Open No. 2002-350163

  However, in the conventional spoken dialogue apparatus, in order to correct a misrecognition of a facility name input in advance, when the user makes a correction utterance, all the facility names input in advance are canceled. For this reason, there is a problem that it is time-consuming to input the canceled facility name from the beginning. For example, in a conventional spoken dialogue apparatus, the location of the vehicle is in Tokyo, and the user speaks “Oppama Station” in Kanagawa as the destination, but this is the name of the station in Tokyo. If the user misrecognizes that it is “Okutama Station”, the user has to speak “Oppama Station in Kanagawa Prefecture” with a corrected utterance. As described above, in the conventional voice interactive apparatus, since it is assumed that the previous utterance content has not been input, there is a problem that the user has to repeat duplicate utterances.

  Therefore, the present invention has been proposed in view of the above-described circumstances, and an object of the present invention is to provide a voice interaction apparatus and a voice recognition method that can reduce the burden on the user at the time of corrected utterance.

  The speech dialogue apparatus according to the present invention includes an input unit for inputting speech speech, a speech recognition unit for recognizing speech speech input by the input unit, and generating a system response according to a speech recognition result, and the speech Output means for outputting the system response generated by the recognition means. When the first collation result obtained by collating the first uttered speech using the first language model including the vocabulary to be recognized is not adopted, the speech recognition means The second utterance speech is collated using the second language model including other vocabulary representing the attribute, and the first utterance speech is again represented using the language model limited by the obtained second collation result. The above-described problem is solved by performing speech recognition and generating a system response according to the speech recognition result.

  In addition, the speech recognition method according to the present invention includes a speech recognition step of recognizing input speech and generating a system response according to a speech recognition result, and the system response generated in the speech recognition step. An output step of outputting. In the speech recognition step, if the first collation result obtained by collating the first uttered speech using the first language model including the vocabulary to be speech-recognized is not adopted, The second utterance speech is collated using the second language model including other vocabulary representing the attribute, and the first utterance speech is again represented using the language model limited by the obtained second collation result. The above-described problem is solved by performing speech recognition and generating a system response according to the speech recognition result.

  In the spoken dialogue apparatus and the speech recognition method according to the present invention, the first utterance speech uttered by the user using the first language model is collated, and as a result, the correct utterance is It is only necessary to utter the second utterance voice including the vocabulary representing the vocabulary attribute of the utterance voice of one utterance, and the second utterance voice is collated using the second language model accordingly. By collating the first utterance voice again using the language model limited by the second collation result, it becomes possible to correctly recognize the vocabulary of the first utterance voice.

  Therefore, in the voice interaction apparatus and the voice recognition method according to the present invention, when the user utters a corrected utterance, it is not necessary to duplicate the first utterance voice as the second utterance voice, thereby reducing the burden on the user. Can do.

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

  The voice interaction device shown as this embodiment has a configuration when applied to a navigation device mounted on a vehicle, a portable terminal device or the like. When this voice interactive apparatus is applied to a navigation apparatus, various functions required by the navigation apparatus can be operated by interactive interaction between the user and the system.

  The present invention is not limited to this, and can be applied to any application that is mounted on various information processing apparatuses and can operate various functions interactively in stages.

[First Embodiment]
[Configuration of voice interactive device]
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 interactive apparatus includes a signal processing unit 1, a microphone 2, a speaker 3, an input device 4, and a display 5. Since it is not a main component according to the present invention, it is not shown in the figure, but if it is a navigation device for a portable terminal device, it may have a configuration provided with other transmission / reception means, and navigation mounted on a vehicle. If it is an apparatus, the structure provided with the transmission / reception means or the communication connection means etc. may be sufficient.

  The signal processing unit 1 converts an audio uttered by a user input from the microphone 2 into a digital audio signal and outputs the digital audio signal, and a digital audio signal output from the signal processing device 14 as a system response. A D / A converter 12 that converts and outputs an analog audio signal, an output amplifier 13 that amplifies the analog audio signal output from the D / A converter 12, a signal processing device 14, and an external storage device 15 are provided. Yes.

  The signal processing device 14 includes a CPU (Central Processing Unit) 21 and a memory 22, and performs speech recognition processing of speech uttered by a user input from the microphone 2 via the A / D converter 11. The voice recognition processing result is output. In addition, the signal processing device 14 has a barge-in function, and can perform voice recognition processing on a user's uttered voice input so as to interrupt a system response by the signal processing device 14.

  The CPU 21 is a control unit that comprehensively controls the signal processing device 14. The CPU 21 reads and executes a processing program stored in the memory 22 to control voice recognition processing and control barge-in functions.

  Normally, the barge-in function is not enabled and is not functioning. The barge-in function is effective only when it is predicted that an interrupt utterance will be made by the user in response to the system response, and a speech recognition process for the interrupt utterance is executed.

  The memory 22 stores in advance processing programs executed by the CPU 21 and various types of frequently used data. In addition, when the speech recognition process is executed by the CPU 21, the memory 22 stores a recognition target word read from the external storage device 15 and its acoustic model, thereby constructing a dictionary of recognition target words. The

  The CPU 21 of the signal processing device 14 refers to the dictionary constructed in the memory 22 and calculates the degree of coincidence between the voice feature pattern of the uttered voice spoken by the user and the voice pattern of the acoustic model of the recognition target word. The voice recognition process is performed.

  The external storage device 15 stores various data used in the navigation device, recognition target data used in speech recognition processing executed by the signal processing device 14, speech data for system response, and the like. The external storage device 15 uses, as recognition target data, a language such as a recognition target word that is a target of voice recognition in the voice recognition process, an acoustic model of the recognition target word, and a network grammar that defines the recognition target word and its connection relationship. Remember the model. The acoustic model of the recognition target word defines a partial word model having acoustic meaning.

  The acoustic model of the recognition target word is a definition of a partial word model that has acoustic meaning, and it is misrecognized when a normal utterance is spoken at a normal utterance speed and once a normal utterance is performed. After the determination, the acoustic models corresponding to the corrected utterances (rephrased utterances) uttered by the user are prepared.

  The network grammar is a rule for defining the connection relation of recognition target words, and can be represented by, for example, a hierarchical structure as shown in FIG. In the example illustrated in FIG. 2, “recognition name” “station name” is defined as a lower hierarchy B, and “prefecture name” is defined as an upper hierarchy A connected to the lower hierarchy B.

  As shown in FIG. 2, the signal processing device 14 uses a network grammar having a hierarchical structure defined for the recognition target words, so that an utterance such as “Tsurumi Station in Kanagawa Prefecture” is made by the user. Even in the case of voice recognition, voice recognition can be performed.

  In addition, if the network grammar that has “prefecture name” and “station name” is switched separately, the utterance is once completed in “Kanagawa Prefecture”, and even if “Tsurumi Station” is subsequently spoken Can recognize.

  The microphone 2 inputs the user's uttered voice to the voice interaction device. The user's utterance voice input from the microphone 2 is converted into a voice signal that is an electrical signal, converted into a digital voice signal by the A / D converter 11, and supplied to the signal processing device 14.

  The speaker 3 outputs the analog audio signal output from the signal processing device 14 as an utterance of the system, converted into an analog audio signal by the D / A converter 12 and amplified by the output amplifier 13 as audio.

  The input device 4 includes a speech switch 4a and a correction switch 4b that are pressed by the user. The utterance switch 4a is a switch for issuing a voice recognition start instruction. On the other hand, the correction switch 4b is a switch for performing correction when the voice uttered by the user is erroneously recognized in the system. If the correction switch 4b is kept pressed for a certain period, the voice recognition process can be terminated halfway.

  The display 5 is realized by, for example, an LCD (Liquid Crystal Display) or the like, and displays a navigation destination, guidance display at the time of setting search conditions, a route guidance screen, and the voice recognition processing by the signal processing device 14. A response image as a result is displayed.

[Operation of voice interactive device]
The voice interaction apparatus having such a configuration performs processing operations according to a series of procedures shown in FIG. In the figure, when a predetermined function of the navigation device is operated, a series of processing steps from when the user inputs required setting items via the voice interaction device to operate the navigation device are shown. ing.

  First, in step S1, the signal processing device 14 in the voice interaction device starts voice recognition processing for the spoken voice in response to an instruction to start utterance by an operation of the utterance switch 4a by the user.

  Subsequently, in step S <b> 2, the signal processing device 14 reads a grammar (language model) that is awaited as a recognition target word in an initial state and sets it in the memory 22 to initialize the vocabulary. That is, the signal processing device 14 recognizes a recognition target word that is a target of speech recognition in the speech recognition processing stored in the external storage device 15 as recognition target data, an acoustic model of the recognition target word, and a recognition target word. A language model such as a network grammar that defines the connection relationship is read out and set in the memory 22.

  Here, for example, the signal processing device 14 recognizes addresses and facility names as shown in FIGS. 4 to 7.

  Specifically, as shown in FIG. 4, the signal processing device 14, based on the grammar of an address including “prefecture name”, “city name”, “city name”, “Yokosuka City Summer, Kanagawa Prefecture” It is possible to recognize continuous speech utterances of addresses such as “Shimamachi” and speech utterances divided into words such as “Kanagawa Prefecture”, “Yokosuka City”, and “Natsushima Town”. Further, as shown in FIG. 5, the signal processing device 14, based on the grammar of the facility composed of “prefecture name”, “station name”, and the like, continuously voice utterances of facilities such as “Oppama Station in Kanagawa Prefecture” Speech utterances divided for each word such as “prefecture” and “Oppama station” can be recognized.

  Furthermore, as shown in FIGS. 6 and 7, the signal processing device 14 can also use a grammar (language model) that is dynamically constructed by collecting facility names existing in the vicinity of the host vehicle position O. FIG. 6 shows an example in which facilities existing in the area A are registered in the grammar, and facilities existing in the areas other than the area A are not registered. Here, the region A is usually a circular region having a radius of several tens of kilometers. On the other hand, FIG. 7 shows an example in which a facility existing in the region C containing the region A is registered in the grammar and a facility existing in the region B is not registered. The area A is an area where a grammar is constructed by extracting up to a facility name having a higher degree of detail than the area C. Usually, the area A is a circular area having a radius of several tens of kilometers, and the area C is a circular area having a radius of several hundred kilometers. The dynamically constructed grammar area may be a prefectural area where the vehicle is present. In this embodiment, it is assumed that the position of the host vehicle is in Tokyo, and the area A includes several thousand nearby facility names.

  The signal processing device 14 may recognize a navigation operation command (not shown) as a recognition target.

  When the signal processing device 14 initializes the vocabulary in this manner, in step S3 in FIG. 3, the signal processing device 14 reads out the notification voice data stored in the external storage device 15, and the D / A converter 12, the output amplifier 13, By outputting through the speaker 3, the user is notified that the process has been started, and an utterance request is made.

  That is, the user starts uttering the recognition target word included in the recognition target data in response to listening to the notification voice that is output via the speaker 3 and informing that the processing by the signal processing device 14 has started. To do. The voice uttered by the user and input via the microphone 2 is converted into a digital voice signal by the A / D converter 11 and output to the signal processing device 14.

  Subsequently, in step S4, the signal processing device 14 starts capturing the voice uttered by the user.

  Normally, the signal processing device 14 calculates the average power of the output (digital audio signal) of the D / A converter 12 until the speech switch 4a is operated. When the speech switch 4a is operated in step S1 described above, the signal processing device 14 compares the calculated average power with the instantaneous power of the input digital audio signal. Then, when the input digital audio signal is greater than the calculated average power by a predetermined value or more, the signal processing device 14 determines that it is the audio section in which the user uttered and captures the audio. Start. Thereafter, the signal processing device 14 continues to calculate the average power, and determines that the user's utterance has ended when the average power becomes smaller than a predetermined value.

  Subsequently, in step S <b> 5, the signal processing device 14 starts a degree-of-match calculation between the captured utterance and the recognition target word read from the external storage device 15 into the memory 22. The degree of coincidence indicates how similar the digital speech signal in the speech section part is to each recognition target word as a score. For example, the signal processing device 14 evaluates that the recognition target word having a larger score value has a higher matching degree. Note that the signal processing device 14 continues to execute voice capturing by parallel processing while executing the coincidence calculation.

  Then, in step S6, the signal processing device 14 determines that the user's utterance has ended in response to the instantaneous power of the digital audio signal being not less than a predetermined time and not more than a predetermined value, and ends the audio capturing.

  After completing the voice capturing, the signal processing device 14 waits in step S7 until the coincidence calculation in step S5 is completed, and determines whether or not the recognition target word is confirmed as the speech recognition result. Specifically, the signal processing device 14 calculates the reliability of the speech recognition result, and determines the speech recognition result when the reliability is equal to or higher than a predetermined threshold. The calculation of the reliability of the speech recognition result is described in “Komatani, Kawahara,“ Usage of the reliability of the speech recognition result in the speech dialogue system ”, Japanese Acoustical Society Proceedings 3-5-2, pp. 73-74, 2000 "and the like.

  Here, when the recognition target word is determined as the speech recognition result, the signal processing device 14 proceeds to step S8, whereas when the recognition target word is not determined as the speech recognition result, the signal processing device 14 proceeds to step S12. In order to notify the user that he / she wants to speak again, a re-utterance request is made and the processing from step S4 is repeated.

  Here, the user's purpose is that the vehicle position is in Tokyo, and the facility names included in the dictionary dynamically constructed according to the vehicle position are only the facility names existing in the vicinity of Tokyo. It is assumed that despite being uttered “Oppama Station” in Kanagawa Prefecture as a place, it is misrecognized as “Okutama Station” as a result of collating this with the station name existing in Tokyo. In this case, “Okutama Station” is very similar in sound to “Oppama Station” and has a sufficiently high reliability. Therefore, the signal processing device 14 uses “Okutama Station” as a speech recognition result. After confirming, the process proceeds to step S8.

  In step S8, the signal processing device 14 determines whether or not the vocabulary that has been speech-recognized includes the lowest vocabulary as in the grammar shown in FIG. When the signal processing device 14 determines that the speech-recognized vocabulary does not include the lowest-level vocabulary, the signal processing device 14 determines that the speech-recognized vocabulary includes the lowest-level vocabulary while proceeding to step S13. In that case, the process proceeds to step S9. In this case, the signal processing device 14 shifts the process to step S9 because the vocabulary “Okutama Station” recognized by speech is the vocabulary at the lowest level having point information.

  Subsequently, the signal processing device 14 generates and outputs a system response in step S9. Specifically, the signal processing device 14 converts “Okutama Station”, which is a speech recognition result, into a speech signal using a speech synthesis processing function (not shown). This audio signal is converted into an analog audio signal by the D / A converter 12, amplified by the output amplifier 13, and then output as audio through the speaker 3.

  In step S10, the signal processing device 14 waits for a predetermined time whether or not the correction switch 4b is pressed by the user. Here, if the correction switch 4b is not pressed, the signal processing device 14 determines that the user has accepted the speech recognition result, determines the speech recognition result in step S11, and responds to the speech recognition result. Process. For example, in a voice interaction device applied to a navigation device, an address of a voice recognition result is set as a destination, and a series of processing ends. On the other hand, if the correction switch 4b is pressed, the signal processing device 14 determines that the user has denied the voice recognition result, and proceeds to step S12.

  Here, since the system response is “Okutama Station” for the speech voice “Oppama Station” by the user, the user presses the correction switch 4b, shifts the processing to Step S12, and again. In order to notify the user that he / she wants to speak, a re-utterance request is made and the processing from step S4 is repeated.

  In step S4 to step S7, the signal processing device 14 performs voice recognition processing of the second uttered voice by the user. Here, it is assumed that the user corrects utterance “Kanagawa” in the second utterance in response to the misrecognition of “Okutama Station” existing in Tokyo with respect to the utterance voice “Oppama Station”. . The voice uttered by the user and input via the microphone 2 is converted into a digital voice signal by the A / D converter 11 and output to the signal processing device 14. In response to this, it is assumed that the signal processing device 14 performs voice recognition processing and obtains a correct voice recognition result “Kanagawa”.

  In this case, the signal processing device 14 shifts the processing to step S13 because the vocabulary “Kanagawa prefecture” recognized in step S8 is the vocabulary of the upper hierarchy. In step S13, the signal processing device 14 again evaluates the reliability of the speech recognition result “Kanagawa” in the same manner as in step S7. If the reliability is equal to or higher than a predetermined threshold, the signal processing device 14 Determine. In this reliability evaluation, a threshold value higher than the threshold value used in step S7 is used. Only when the signal processing device 14 is sufficiently reliable by using such a threshold value, the signal processing device 14 determines the recognition target word as a speech recognition result, and proceeds to step S14. On the other hand, if the recognition target word is not confirmed as the speech recognition result, the signal processing device 14 gives up the use of the first utterance speech “Oppama Station” and proceeds to step S21. Here, it is assumed that the speech recognition result is confirmed and the process proceeds to step S14.

  Subsequently, in step S14, the signal processing device 14 determines whether or not there is a fact that the utterance as the previous system response has been denied by the user. Here, if the previous utterance exists and is denied, the signal processing device 14 determines that the user's intention to correct the misrecognition is present, and proceeds to step S15. On the other hand, if not, the process proceeds to step S21. In addition, because the speech utterance “Oppama Station” was misrecognized as “Okutama Station” in Tokyo, the correct utterance was “Kanagawa” as the second utterance speech. The process proceeds to step S15.

  Subsequently, in step S15, the signal processing device 14 trusts “Kanagawa Prefecture”, which is the voice recognition result, and sets a facility in Kanagawa Prefecture as a standby language. That is, the signal processing device 14 narrows down only the facility name vocabulary located in the lower hierarchy of Kanagawa Prefecture shown in FIG. 5 as the standby range, and the grammar shown in FIG. 4, the other part of FIG. 5, FIG. The grammar shown in FIG. 7 is out of the standby range.

  Subsequently, in step S16, the signal processing device 14 performs voice recognition again by collating the first utterance voice by the user saved at the time of the previous voice recognition with the grammar narrowed down in step S15. Process.

  Then, in step S17, the signal processing device 14 evaluates the reliability of the speech recognition result “Oppama station” again in the same manner as in step S7, and if the reliability is equal to or higher than a predetermined threshold, the speech recognition Confirm as a result. In this reliability evaluation, a threshold value higher than the threshold value used in step S7 is used. Only when the signal processing device 14 is sufficiently reliable by using such a threshold value, the signal processing device 14 determines the recognition target word as the speech recognition result, and proceeds to step S18. On the other hand, if the recognition target word is not determined as the speech recognition result, the signal processing device 14 gives up the use of the first utterance speech, discards the first utterance speech in step S20, and then proceeds to step S21. And migrate the process.

  If the recognition target word is confirmed as a speech recognition result, the signal processing device 14 converts the speech recognition result “Oppama Station in Kanagawa Prefecture” into a speech signal using a speech synthesis processing function (not shown) in step S18. To do. This audio signal is converted into an analog audio signal by the D / A converter 12, amplified by the output amplifier 13, and then output as audio through the speaker 3.

  In step S19, the signal processing device 14 waits for a predetermined time whether or not the user has pressed the correction switch 4b. Here, when the correction switch 4b is not pressed, the signal processing device 14 determines that the user has accepted the speech recognition result, determines the speech recognition result in step S11, and as described above, A process according to the voice recognition result is performed, and a series of processes is terminated. On the other hand, if the correction switch 4b is pressed, the signal processing device 14 determines that the user has denied the voice recognition result, and discards the first utterance voice in step S20, and then proceeds to step S21. And move the process.

  In addition, since the signal processing device 14 is in a state of accepting only the utterance voice “Kanagawa” when the process proceeds to step S21, the user presses the correction switch 4b in step S22. It waits for a predetermined time. Here, if the correction switch 4b is not pressed, the signal processing device 14 determines that the user has accepted the speech recognition result, and in step S23, restricts the grammar to addresses and facilities in Kanagawa Prefecture, The processing from step S4 is repeated. On the other hand, if the correction switch 4b is pressed, the signal processing device 14 determines that the user has denied the voice recognition result, and notifies the user that he / she wants to speak again in step S12. Then, after making a re-utterance request, the processing from step S4 is repeated.

  By performing the processing operation according to such a series of procedures, the voice interactive apparatus can correctly recognize the second utterance voice even if the first utterance voice is erroneously recognized. The ground can be set and a predetermined function of the navigation device can be operated.

[Effect of the first embodiment]
As described above in detail, in the voice interactive apparatus shown as the first embodiment, the signal processing apparatus 14 uses the first language model including the vocabulary to be recognized by the first speech. If the first collation result obtained by collating the speech is not adopted, the second speech model is collated using the second language model including another vocabulary representing the attribute of the vocabulary, and the obtained first utterance is obtained. The first utterance speech is recognized again using a language model limited by the collation result of 2, and a system response corresponding to the speech recognition result is generated.

  As a result, in this voice interactive apparatus, when the user utters a corrected utterance, it is not necessary to duplicate the first utterance as the second utterance, and the burden on the user can be reduced.

  Specifically, in the voice interaction device, the signal processing device 14 uses the same language model as that limited by the second collation result as the first utterance speech included in the point indicated by the second utterance speech. Refine to language models that contain vocabulary representing attributes. For example, in the spoken dialogue apparatus, when the vehicle position is in Tokyo, the user utters “Oppama Station” as the destination, but the result of matching this with the name of the facility existing in Tokyo is “ Even if it is misrecognized as "Okutama Station", only "Kanagawa Prefecture" representing the prefecture as an attribute of "Oppama Station" should be uttered as a corrected utterance. In response to this, the voice dialogue device outputs the correct speech recognition result “Oppama Station” by collating the name of the facility existing in Kanagawa Prefecture with “Oppama Station” as the first utterance voice. Is possible. In the spoken dialogue apparatus, as the corrected utterance, for example, an utterance including an attribute other than the prefecture such as “Yokosuka City in Kanagawa Prefecture” or “Keihin Express Main Line” may be used as the second utterance speech.

  Further, in the voice interaction device, the signal processing device 14 recognizes the first utterance voice again only when the reliability of the voice recognition result of the second utterance voice is equal to or higher than a predetermined threshold. As a result, in the voice interaction device, when the second collation result is not reliable, the dictionary is not switched based on the second collation result, and the voice recognition is not performed again, and the re-recognition result is erroneously recognized. It can be avoided. Further, in the voice interactive apparatus, it is possible to avoid an increase in calculation amount due to wasteful re-recognition.

  Further, in the voice interactive apparatus, the signal processing apparatus 14 generates a system response only when the reliability of the re-speech recognition result of the first uttered voice is equal to or higher than a predetermined threshold. Thereby, in the voice interaction apparatus, it is possible to avoid a situation in which the re-recognition result is erroneously recognized.

  Furthermore, in the voice interaction device, the signal processing device 14 uses the threshold value used for the reliability calculation of the voice recognition result of the second utterance voice or the reliability calculation of the re-voice recognition result of the first utterance voice. As the threshold value, a threshold value higher than the threshold value used for the reliability calculation of the speech recognition result of the first uttered speech is used. Thereby, in the voice interaction apparatus, the adoption standard of the result of performing the voice recognition again by switching the dictionary can be increased, and the situation where the re-recognition result is erroneously recognized can be surely avoided.

[Second Embodiment]
Next, a voice interactive apparatus shown as the second embodiment of the present invention will be described.

  The voice interactive apparatus shown as the second embodiment has the same configuration as the voice interactive apparatus shown as the first embodiment described with reference to FIG. Therefore, about the structure of the voice interactive apparatus shown as 2nd Embodiment, the description is abbreviate | omitted by attaching | subjecting the same code | symbol.

  In addition, the voice interaction apparatus shown as the second embodiment, like the voice interaction apparatus shown as the first embodiment described above, uses a language model to be used at the time of correct utterance when there is a misrecognition. By limiting, it is possible to prevent the user from repeating repeated utterances, but the language model to be used is limited based on the distance from the point indicated by the user's uttered speech. .

  Therefore, the processing operation of the voice interaction apparatus shown as the second embodiment is only a part of the flowchart described with reference to FIG. 3 as the processing operation of the voice interaction apparatus shown as the first embodiment. The same processing contents are given the same step numbers, and the description thereof is omitted.

[Operation of voice interactive device]
The voice interactive apparatus shown as the second embodiment performs processing operations according to a series of procedures shown in FIG. In the figure, when a predetermined function of the navigation device is operated, a series of processing steps from when the user inputs required setting items via the voice interaction device to operate the navigation device are shown. ing.

  When the signal processing device 14 in the voice interactive device performs the processing of Steps S1 to S6 and finishes the voice capturing, in Step S7, the reliability of the speech recognition result is calculated, and the reliability is equal to or higher than a predetermined threshold. If it is detected, the result is confirmed as a voice recognition result.

  Here, the user's vehicle position is in Tokyo, and the facility name included in the dictionary that is dynamically constructed according to the vehicle position is only the facility name that exists in the vicinity of Tokyo. Despite having spoken with the “Japan Agency for Marine-Earth Science and Technology” in Kanagawa Prefecture as a destination, the result of collating this with the name of a facility existing in Tokyo was misunderstood as “Kasai Rinkai Park”. To do. In this case, since the pronunciation of “Kasai Rinkai Park” is very similar to that of “Japan Agency for Marine-Earth Science and Technology” and the reliability is sufficiently large, the signal processing device 14 changes the name of “Kasai Rinkai Park”. The result is confirmed as a voice recognition result, and the process proceeds to step S8.

  In step S8, the signal processing device 14 determines whether or not the vocabulary that has been speech-recognized includes the lowest vocabulary as in the grammar shown in FIG. If the signal processing device 14 determines that the speech-recognized vocabulary does not include the lowest-level vocabulary, the signal processing device 14 determines that the speech-recognized vocabulary includes the lowest-level vocabulary while proceeding to step S21. In that case, the process proceeds to step S31. In this case, since the vocabulary “Kasai Rinkai Park” recognized by speech is the lowest vocabulary having point information, the signal processing device 14 shifts the processing to step S31.

  Subsequently, in step S31, the signal processing device 14 determines whether or not there is a fact that the utterance as the previous system response has been denied by the user. Here, if the previous utterance exists and is denied, the signal processing device 14 determines that the user's intention to correct the misrecognition is present, and proceeds to step S32. On the other hand, if not, the process proceeds to step S9. Here, since only the first utterance voice (first utterance voice) has been input, the process proceeds to step S9.

  Subsequently, the signal processing device 14 generates and outputs a system response in step S9. Specifically, the signal processing device 14 converts “Kasai Rinkai Park”, which is a speech recognition result, into a speech signal using a speech synthesis processing function (not shown). This audio signal is converted into an analog audio signal by the D / A converter 12, amplified by the output amplifier 13, and then output as audio through the speaker 3.

  In step S10, the signal processing device 14 waits for a predetermined time whether or not the correction switch 4b is pressed by the user. Here, if the correction switch 4b is not pressed, the signal processing device 14 determines that the user has accepted the speech recognition result, determines the speech recognition result in step S11, and responds to the speech recognition result. Process. For example, in a voice interaction device applied to a navigation device, an address of a voice recognition result is set as a destination, and a series of processing ends. On the other hand, if the correction switch 4b is pressed, the signal processing device 14 determines that the user has denied the voice recognition result, and proceeds to step S12.

  Here, since the system response is “Kasai Rinkai Park” with respect to the speech voice “Marine Research and Development Organization” by the user, the user presses the correction switch 4b, and the process proceeds to step S12. In order to notify the user that he / she wants to speak again, a re-utterance request is made and the processing from step S4 is repeated.

  In step S4 to step S8, the signal processing device 14 performs voice recognition processing of the second uttered voice by the user. Here, in response to the misrecognition of the utterance voice “Marine Research and Development Organization” as “Kasai Rinkai Park” existing in Tokyo, the user uses “Marine Research and Development Organization” as the second utterance voice. Assume that you have made a correct utterance as "Oppama Station in Kanagawa Prefecture" as an alternative facility. The voice uttered by the user and input via the microphone 2 is converted into a digital voice signal by the A / D converter 11 and output to the signal processing device 14. In response to this, it is assumed that the signal processing device 14 performs voice recognition processing and obtains a correct voice recognition result “Oppama Station in Kanagawa”.

  In this case, since the vocabulary “Oppama Station in Kanagawa Prefecture” uttered by the user in step S8 is the lowest vocabulary, the signal processing device 14 shifts the processing to step S31 and moves to the previous system. It is determined whether or not there was a fact that the utterance as a response was denied by the user. In addition, here, in response to the misrecognition of the utterance voice “Japan Marine Research and Development Organization” as “Kasai Rinkai Park” in Tokyo, the second utterance voice is “Oppama Station in Kanagawa”. Since the corrected speech has been made, the process proceeds to step S32.

  In step S32, the signal processing device 14 trusts the voice recognition result “Oppama Station in Kanagawa” and sets a nearby facility existing within a predetermined distance from Oppama Station as a standby word. That is, as shown in FIG. 9, for example, the signal processing device 14 collects nearby facility names existing within a predetermined distance from Oppama Station A and dynamically constructs a grammar (language model), and uses only the facility name vocabulary. We narrow down as waiting range. At this time, the signal processing device 14 extracts a facility name having a high degree of detail as a facility in the vicinity of Oppama station A, and incorporates the facility name into the grammar.

  Subsequently, in step S33, the signal processing device 14 performs voice recognition again by collating the first utterance voice by the user saved at the time of the previous voice recognition with the grammar narrowed down in step S32. Process. In this case, it is assumed that “Japan Agency for Marine-Earth Science and Technology” is obtained as a result of speech recognition.

  Then, in step S34, the signal processing device 14 again evaluates the reliability of the speech recognition result “Marine Research and Development Organization” as in step S7, and when the reliability is equal to or higher than a predetermined threshold value. Confirmed as a speech recognition result. In this reliability evaluation, a threshold value higher than the threshold value used in step S7 is used. Only when the signal processing device 14 is sufficiently reliable by using such a threshold, the signal processing device 14 determines the recognition target word as a speech recognition result, and proceeds to step S35. On the other hand, if the recognition target word is not determined as the speech recognition result, the signal processing device 14 gives up the use of the first utterance speech, discards the first utterance speech in step S37, and then proceeds to step S38. And move the process.

  If the recognition target word is confirmed as the speech recognition result, the signal processing device 14 uses the speech synthesis processing function (not shown) to obtain the speech recognition result “Ocean Research and Development Organization near Oppama Station in Kanagawa Prefecture” in step S35. Is converted into an audio signal. This audio signal is converted into an analog audio signal by the D / A converter 12, amplified by the output amplifier 13, and then output as audio through the speaker 3.

  In step S36, the signal processing device 14 waits for a predetermined time whether or not the user has pressed the correction switch 4b. Here, when the correction switch 4b is not pressed, the signal processing device 14 determines that the user has accepted the speech recognition result, determines the speech recognition result in step S11, and as described above, A process according to the voice recognition result is performed, and a series of processes is terminated. On the other hand, if the correction switch 4b is pressed, the signal processing device 14 determines that the user has denied the speech recognition result, and discards the first uttered speech in step S37, and then proceeds to step S38. And move the process.

  Further, when the process proceeds to step S38, the signal processing device 14 does not illustrate “Oppama station in Kanagawa”, which is the current speech recognition result, because the first utterance speech is discarded. The voice signal is converted into a voice signal using the voice synthesis processing function. This audio signal is converted into an analog audio signal by the D / A converter 12, amplified by the output amplifier 13, and then output as audio through the speaker 3.

  In step S39, the signal processing device 14 waits for a predetermined time whether or not the correction switch 4b is pressed by the user. Here, if the correction switch 4b is not pressed, the signal processing device 14 determines that the user has accepted the speech recognition result, determines the speech recognition result in step S11, and as described above, A process according to the voice recognition result is performed, and a series of processes is terminated. On the other hand, if the correction switch 4b is pressed, the signal processing device 14 determines that the user has denied the voice recognition result, proceeds to step S12, and indicates that the user wants to speak again. In order to notify the user, after making a re-utterance request, the processing from step S4 is repeated.

  By performing the processing operation according to such a series of procedures, the voice interactive apparatus can correctly recognize the second utterance voice even if the first utterance voice is erroneously recognized. The ground can be set and a predetermined function of the navigation device can be operated.

[Effect of the second embodiment]
As described above in detail, in the voice interactive device shown as the second embodiment, the signal processing device 14 instructs the second spoken voice as a language model limited by the second collation result. To a language model including a vocabulary that represents the same attribute as the first uttered speech existing within a predetermined distance from the point where

  As a result, in this voice interactive apparatus, when the user utters a corrected utterance, it is not necessary to duplicate the first utterance as the second utterance, and the burden on the user can be reduced. For example, in a spoken dialogue device, when the vehicle position is in Tokyo, the user spoke with “Japan Agency for Marine-Earth Science and Technology” as the destination, but this was compared with the name of the facility existing in Tokyo. Even if the result is misunderstood as “Kasai Rinkai Park”, “Oppama Station in Kanagawa” can be used as a corrected utterance to speak out a facility that is an alternative to the target facility. In response to this, the spoken dialogue device collates the name of the facility existing within a predetermined distance from Oppama Station in Kanagawa Prefecture with the “Ocean Research and Development Organization” as the first utterance speech, It is possible to output a correct speech recognition result “mechanism”.

  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 even if it is a form other than this embodiment, as long as it does not depart from the technical idea according to the present invention, the design and the like Of course, various modifications are possible.

It is a block diagram explaining the structure of the voice interactive apparatus shown as embodiment of this invention. It is a figure for demonstrating network grammar. It is a flowchart explaining the processing operation of the voice interactive apparatus shown as the first embodiment of the present invention. It is a figure for demonstrating the grammar of an address. It is a figure for demonstrating the grammar of a facility. It is a figure for demonstrating the grammar which collects the facility names which exist in the vicinity of the own vehicle position, and is built dynamically. It is a figure for demonstrating the grammar which collects the facility names which exist in the vicinity of the own vehicle position, and is built dynamically, and is a figure for demonstrating the grammar different from FIG. It is a flowchart explaining the processing operation of the voice interactive apparatus shown as the 2nd Embodiment of this invention. It is a figure for demonstrating the grammar constructed | assembled dynamically by collecting the name of the nearby facility which exists within the predetermined distance from an alternative facility.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Signal processing unit 2 Microphone 3 Speaker 4 Input device 4a Speech switch 4b Correction switch 5 Display 11 A / D converter 12 D / A converter 13 Output amplifier 14 Signal processing device 15 External storage device 21 CPU
22 memory

Claims (7)

An input means for inputting speech voice;
Voice recognition means for recognizing speech speech input by the input means and generating a system response according to the voice recognition result;
Output means for outputting the system response generated by the voice recognition means,
If the first collation result obtained by collating the first uttered speech using the first language model including the vocabulary to be recognized is not adopted, the speech recognition means sets the attribute of the vocabulary. The second utterance speech is collated using the second language model including the other vocabulary to represent, and the first utterance speech is recognized again using the language model limited by the obtained second collation result. And generating a system response according to the voice recognition result. The speech recognition means includes, as a language model limited by the second collation result, a language model including a vocabulary representing the same attribute as that of the first utterance speech included in a point indicated by the second utterance speech The voice interaction device according to claim 1, wherein the voice interaction device is narrowed down to The speech recognition means, as a language model limited by the second collation result, a vocabulary representing the same attribute as the first speech speech existing within a predetermined distance from a point indicated by the second speech speech The spoken dialogue apparatus according to claim 1, wherein the language model is narrowed down to language models including Comprising reliability calculation means for calculating the reliability of the voice recognition result by the voice recognition means;
The voice recognition means recognizes the first utterance voice again only when the reliability of the voice recognition result of the second utterance voice obtained by the reliability calculation means is equal to or greater than a predetermined threshold. The spoken dialogue apparatus according to claim 1, wherein: The speech recognition means generates a system response only when the reliability of the re-speech recognition result of the first uttered speech obtained by the reliability calculation means is equal to or greater than a predetermined threshold value. The voice interactive apparatus according to claim 4. The reliability calculation means uses the first threshold as a threshold used for the reliability calculation of the speech recognition result of the second utterance speech or the threshold used for the reliability calculation of the re-speech recognition result of the first utterance speech. The voice dialogue apparatus according to claim 4 or 5, wherein a threshold value higher than a threshold value used for the reliability calculation of the voice recognition result of the uttered voice is used. A speech recognition step of recognizing the input speech and generating a system response according to the speech recognition result;
An output step of outputting the system response generated in the voice recognition step,
In the speech recognition step, if the first collation result obtained by collating the first uttered speech using the first language model including the vocabulary to be speech-recognized is not adopted, the attribute of the vocabulary is set. The second utterance speech is collated using the second language model including the other vocabulary to represent, and the first utterance speech is recognized again using the language model limited by the obtained second collation result. And generating a system response according to the speech recognition result.

Images