JP4408665B2 - Speech recognition apparatus for speech recognition, speech data collection method for speech recognition, and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a device and a method for collecting utterance data for speech recognition, and a computer program that can efficiently collect the utterance data while keeping the precision of speech recognition high. <P>SOLUTION: Provided are a voice interaction device including a means of storing interaction scenario information in which a progress procedure of interaction is described, a means of accepting inputted utterance, a means of recognizing the speech of the inputted utterance, a means of advancing the interaction according to the speech recognition result and interaction scenario information, and a means of outputting an answer to the utterance, a means of storing a state transition history of the interaction based upon the interaction scenario information, a means of judging whether or not the inputted utterance is correctly recognized according to the speech recognition result and state transition history, and a means of storing the speech recognition result and inputted utterance so that they correspond to each other when it is judged that the inputted utterance is correctly recognized. <P>COPYRIGHT: (C)2005,JPO&amp;NCIPI

Description

  The present invention relates to a speech recognition speech data collection apparatus, speech recognition speech data collection method, and computer program for collecting and storing speech data used for speech recognition.

  In recent years, an interactive voice response (IVR) such as a voice portal using a speech recognition system (ASR: Auto Speech Recognition) has begun to spread. The recognition performance of the speech recognition system that recognizes utterances greatly affects the ease of use of the speech dialogue system. The accuracy of an acoustic model modeled by an HMM (Hidden Markov model) or the like greatly affects the performance of the speech recognition system.

  In general, in order to develop an accurate acoustic model, a large amount of utterance data whose utterance content is clear is required when learning the acoustic model. When a statistical model such as HMM is used, the amount of utterance data is directly related to the performance of the acoustic model. Therefore, it is one of the important issues in the development of a speech recognition system to easily collect a large amount of utterances to which data indicating the utterance content is added. Note that “utterance data” means the entire data used for speech recognition including information about the utterance by the speaker and the associated utterance content.

  In addition, it is desirable that the environment for inputting the utterance used for learning the acoustic model is the same as the environment in which the utterer uses the speech recognition system. For example, when the speech data used when learning the acoustic model is acquired by preparing a text to be read out and recording the speech read by the speaker, the speech data has a tone of reading. On the other hand, in an environment where the speech recognition system is actually used, utterances with spoken tones close to free utterances are greater than when reading a document. Therefore, the divergence of the speech feature amount or the like between the utterance data used when learning the acoustic model and the actual utterance increases, and the recognition accuracy deteriorates.

  In addition, when collecting utterance data by the above method, a full-time person is required for operation of the equipment that records the utterance that reads out the manuscript, and collects utterance data in large quantities from the viewpoint of utterance data collection cost or collection efficiency It was difficult to do. Furthermore, the utterance data read out by the utterer of the manuscript includes misread data, and it is necessary to exclude data inappropriate as utterance data used when learning such an acoustic model.

  In order to solve such a problem, a method of generating speech data by associating the recognition result of the speech recognition system with the speech is used. For example, by associating the recognition result of the voice recognition system as the utterance identification information, the content of the utterance becomes clear. However, it does not solve the problem that it is necessary to exclude misread data.

  Therefore, as disclosed in Patent Document 1, the recognition result of the voice recognition system is associated with the utterance to generate and store the utterance data, and as disclosed in Patent Document 2, the utterer himself notices. Provide a means to exclude misread data. FIG. 10 is a functional block configuration diagram of a speech recognition utterance data collection device that generates and stores speech data by associating the recognition result of the speech recognition device with an utterance.

  As shown in FIG. 10, the speech of the speaker is input from the speech input unit 1 and sent to the speech recognition unit 2. Each time the speaker speaks, the speech recognition unit 2 outputs a speech recognition result corresponding to the input speech, and stores it in the recognition result storage unit 3 and the speech storage unit 4, respectively.

When the speaker determines that the speech recognition result is not correct and a recognition result cancellation request signal is input from the speech cancellation input unit 7, the utterance data generation unit 5 stores the corresponding input utterance and speech recognition stored therein. The results are deleted from the recognition result storage unit 3 and the utterance storage unit 4, respectively. When the recognition result cancellation request signal is not input from the utterance cancellation input unit 7, the utterance data generation unit 5 associates the utterance and the voice recognition result stored in the recognition result storage unit 3 and the utterance storage unit 4 as a pair of data. And stored in the utterance data storage unit 6. By doing in this way, the speech data stored in the speech data storage unit 6 does not include those erroneously recognized by the speech recognition unit 2. Therefore, it is possible to construct an acoustic model with high recognition performance.
JP 2003-150185 A JP 2002-189596 A

  However, in the speech recognition utterance data collection device shown in FIG. 10, the input utterance is an utterance that is read out from the manuscript, and learning is performed using the environment for inputting the utterance used when learning the acoustic model and the utterance data. It is practically difficult to match the environment in which the user uses the speech recognition apparatus using the acoustic model.

  In addition, it is necessary to collect many speakers in order to collect the utterances used when learning the acoustic model. When using a method such as paying a speaker who cooperated in utterance collection, it is possible to collect utterers relatively easily, but the financial burden of learning the acoustic model becomes heavy.

  Furthermore, the recognition performance of the acoustic model depends on the difference in the sex ratio, age distribution, etc. of the speakers collected for the construction of the acoustic model. It goes without saying that the gender ratio, age distribution, etc. of the speakers to be collected are preferably close to the gender ratio, age distribution, etc. of the speakers who actually use the speech recognition system.

  SUMMARY OF THE INVENTION The present invention has been made in view of such circumstances, and the speech recognition speech data collection device and speech recognition speech data collection capable of efficiently collecting speech data while maintaining high speech recognition accuracy. It is an object to provide a method and a computer program.

In order to achieve the above object, the speech data collection apparatus for speech recognition according to the first aspect of the present invention is a description for executing a process to be performed next according to the content of an utterance when an utterance is input in accordance with a progress procedure of a dialog. Means for storing dialogue scenario information including: speech data accumulating means stored as speech data for speech recognition in association with speech recognition results and input utterances; means for receiving input utterances; A voice recognition means for the input utterance using speech data stored in the utterance data storage means, a means for proceeding a dialogue based on a voice recognition result and the dialogue scenario information, and execution of the processing A voice dialogue apparatus including a detection means for detecting a voice, a means for outputting a response to the input utterance, and a dialogue state transition history based on the dialogue scenario information Means that, when the detection means detects execution of the processing, it is determined that the input speech is speech that corresponds to the progression steps of the dialogue described in the dialog scenario information, said detecting means the When the execution of the process is not detected, a determination unit that determines whether or not the input utterance is correctly recognized based on the voice recognition result and the state transition history, and the determination unit is correctly recognized. The speech recognition result and the input utterance are associated with each other and stored in the utterance data storage means. Further, the speech data collection apparatus for speech recognition according to the present invention includes means for storing dialog scenario information describing a progress procedure of a dialog, means for receiving an input speech, means for recognizing the input speech, speech A voice dialogue device including means for advancing a dialogue based on a recognition result and the dialogue scenario information; and a means for outputting a response to the inputted utterance; and storing a state transition history of the dialogue based on the dialogue scenario information Means, based on the voice recognition result and the state transition history, a means for judging whether or not the inputted utterance has been correctly recognized, and if the means has been recognized correctly, the voice recognition result And means for storing the inputted utterances in association with each other, means for receiving and storing a calling number on a telephone line, and Characterized in that it comprises a means for storing in association with identification result and the input speech.

  In the speech recognition utterance data collection device according to the first aspect of the present invention, it is evaluated whether or not the utterance is correctly recognized in the dialogue progressed by the speech dialogue device, and when the utterance is correctly recognized, the speech recognition result and the utterance Are associated and stored as utterance data.

  In the speech recognition utterance data collection device according to the second aspect of the present invention, in the first aspect of the invention, the speech stored in association with the speech recognition result is speech waveform data or speech characteristics obtained by acoustic analysis of the speech. It is characterized by a quantity.

  In the speech recognition utterance data collecting apparatus according to the second aspect of the invention, speech waveform data or utterance feature values obtained as a result of acoustic analysis of the utterance are associated with the speech recognition result and stored as utterance data.

  In addition, the speech recognition utterance data collection device according to the third invention is the first invention or the second invention, wherein means for receiving and storing a calling number on a telephone line, the calling number, the voice recognition result and And means for storing in association with the inputted utterance.

  In the speech recognition utterance data collection device according to the third aspect of the invention, it is possible to store a unique call number and utterance data in association with each telephone line, depending on noise, filtering, modulation, etc. specific to each telephone line. Generate an acoustic model.

In addition, the speech recognition speech data collection method according to the fourth aspect of the invention is a description that uses a computer to execute a process to be performed next according to the content of an utterance when the utterance is input in accordance with the progress procedure of the dialog. Storing dialogue scenario information including: utterance data storage step stored as speech recognition speech data by associating the speech recognition result and the input utterance; and receiving the input utterance; said input utterance, by using the speech data accumulated in the speech data storage step, the steps of advancing a voice recognizing a conversation based on the speech recognition result and the dialog scenario information, the processing Using the detection step of detecting execution, the step of outputting a response to the input utterance, and the spoken dialogue method, Storing a state transition history of the conversation based on the information, when said detecting step detects the execution of the processing, in speech input utterance corresponding to the progression steps of the dialogue described in the dialog scenario information A determination step of determining whether or not the input utterance is correctly recognized based on the voice recognition result and the state transition history when the detection step does not detect the execution of the process. When it is determined that the speech is recognized correctly as a result of the determination in the determination step, a step is performed in which the speech recognition result and the input utterance are associated with each other and stored in the utterance data storage step.

  In the speech recognition utterance data collection method according to the fourth aspect of the present invention, it is evaluated whether or not the utterance is correctly recognized in the dialogue progressed by the speech dialogue apparatus. If the utterance is correctly recognized, the speech recognition result and the utterance Are associated and stored as utterance data.

According to a fifth aspect of the present invention, there is provided a computer program according to the fifth aspect of the present invention, wherein the computer includes a description of dialog scenario information including a description for executing a process to be performed next in accordance with the content of the utterance when the utterance is input in accordance with the progress of the dialog The speech recognition result and the input utterance, the speech data storage means stored as the speech data for speech recognition, the means for accepting the input utterance, and the input utterance Using speech data stored in the speech data storage means, voice recognition means, means for advancing a dialogue based on the voice recognition result and the dialogue scenario information, and detection for detecting execution of the processing A voice dialog device including means, a means for outputting a response to the input utterance, and a dialog state transition history based on the dialog scenario information Means for憶, when the detection means detects execution of the processing, it is determined that the input utterance is the utterance corresponding to the progression steps of the dialogue described in the dialog scenario information, said detecting means the When the execution of the process is not detected, a determination unit that determines whether or not the input utterance is correctly recognized based on the voice recognition result and the state transition history, and a result of the determination by the determination unit, When it is determined that the speech is recognized correctly, the speech recognition result and the input utterance are associated with each other and functioned as a means for storing in the utterance data storage means.

  By introducing the computer program according to the fifth aspect of the present invention into a computer, it is evaluated whether or not the utterance is correctly recognized in the dialogue progressed by the voice dialogue device, and if the utterance is correctly recognized, the voice recognition result and Utterances are associated and stored as utterance data.

  According to the speech recognition utterance data collection device according to the first aspect of the present invention, it is possible to collect utterances associated with correct speech recognition results based on whether or not a voice dialogue is established, and generate an acoustic model with high speech recognition accuracy. It is possible to efficiently collect utterance data.

  Moreover, according to the speech recognition utterance data collection device according to the second aspect of the present invention, it is possible to collect utterances associated with correct speech recognition results based on whether or not a voice dialogue is established, and an acoustic model with high speech recognition accuracy It is possible to efficiently collect the utterance data for generating.

  Further, according to the speech recognition utterance data collection device according to the third aspect of the present invention, for a user utterance via a telephone line, an environment for inputting an utterance used for learning an acoustic model, and an acoustic model learned using the utterance data This makes it easy to match the speech recognition device using the voice and the environment used by the speaker, and it is possible to efficiently collect speech data for generating an acoustic model with high speech recognition accuracy.

  Further, according to the speech recognition utterance data collection method according to the fourth aspect of the present invention, it is possible to collect utterances associated with correct speech recognition results based on whether or not a voice dialogue is established, and an acoustic model with high speech recognition accuracy It is possible to efficiently collect the utterance data for generating.

  Moreover, according to the computer program which concerns on 5th invention, the speech matched with the correct speech recognition result can be collected based on the possibility of establishment of a speech dialogue, and it is for generating an acoustic model with high speech recognition accuracy It is possible to collect utterance data efficiently.

  Hereinafter, the present invention will be specifically described with reference to the drawings showing embodiments thereof.

(Embodiment 1)
Hereinafter, the speech recognition utterance data collection apparatus according to Embodiment 1 of the present invention will be specifically described with reference to the drawings. In the first embodiment, a case will be described in which the speech recognition utterance data collection device is implemented using a single computer. Of course, the speech data used for voice recognition may be stored in a storage device of another computer connected via the communication means, a portable recording medium such as a DVD, and the communication means is also particularly limited. is not.

  FIG. 1 is a schematic configuration diagram of a computer that embodies the speech recognition speech data collection apparatus according to Embodiment 1 of the present invention. As shown in FIG. 1, a computer embodying a speech recognition utterance data collection device includes at least a CPU (central processing unit) 11, a storage unit 12, a RAM (memory) 13, and a communication unit connected to an external communication unit. 14, an input means 15 such as a mouse and a keyboard, an output means 16 such as a monitor, and an auxiliary storage means 17.

  The auxiliary storage means 17 is a portable recording medium 18 in which a program used by a computer that embodies the speech data collection apparatus for speech recognition is recorded, and corresponds to a DVD, a CD-ROM, or the like. Moreover, the portable recording medium 18 etc. which record the data used with the speech data collection apparatus for speech recognition, such as speech data used for speech recognition, are included.

  A computer that embodies the speech recognition utterance data collection device according to Embodiment 1 of the present invention incorporates a voice interaction device 20. The voice interaction device 20 also includes a CPU (central processing unit) 11, a storage unit 12, a RAM (memory) 13, a communication unit 14 connected to an external communication unit, an input unit 15 such as a mouse and a keyboard, a monitor, etc. The output unit 16 and the auxiliary storage unit 17 function.

  First, in order to urge the speaker to speak, the computer performs voice output from the output unit 16 according to a command from the CPU 11 in accordance with the dialogue scenario information stored in the storage unit 12. For example, a question that can limit the utterance input by the speaker next, such as “Which is XX, XX,... The output from the output means 16 is not limited to audio output, but may be display output on a screen.

  Note that the dialogue scenario information is described so that an utterance in the dialogue can be received by a scenario description language such as VoiceXML. That is, the dialogue scenario information describes the contents of the output from the computer, the transition of the dialogue according to the utterance, the processing to be performed next according to the utterance content, and the like.

  When an utterance is input from the input unit 15 with respect to the output voice, the input utterance is stored as waveform data of the voice or data indicating an utterance feature value as a result of acoustic analysis of the input utterance. 12 and the RAM 13, and speech recognition is performed on the utterance stored in the RAM 13 according to a command from the CPU 11. The speech recognition engine used for speech recognition processing is not particularly limited, and any speech recognition engine that is generally used may be used. The voice recognition result is stored in the storage unit 12 and the RAM 13.

  Note that the storage unit 12 is not limited to a built-in hard disk, and stores a large amount of data such as a hard disk built in another computer connected via the communication unit 14. Any recording medium can be used.

  The CPU 11 determines whether or not the utterance is correctly recognized based on the voice recognition result stored in the RAM 13. Various methods can be used for determining whether or not the utterance is correctly recognized. Hereinafter, a specific example will be described.

  For example, the state transition history of the dialogue based on the dialogue scenario information is stored in the storage means 12 or the RAM 13, and the input utterance is correctly recognized based on the stored voice recognition result and the state transition history. There is a method for determining whether or not. FIG. 2 shows a state transition diagram in a dialogue scenario for confirming the name. As shown in FIG. 2, the conversation scenario starts in state 1, a system utterance “Please name” is output, and the state transitions to state 2.

  In state 2, the input utterance is recognized as speech, and the speech recognition result is stored in RAM 13. When the stored speech recognition result is “XX”, a system utterance “You are Mr. XX” is output in the dialogue scenario, and the state transitions to state 3.

  In state 3, the input utterance is recognized as speech, and the speech recognition result is stored in RAM 13. Since it can be determined that the voice recognition result is “Yes” or “No” in State 3, the reliability of the voice recognition result in State 3 is high. When the stored speech recognition result is “Yes”, it is possible to determine that the speech recognition result in State 2 is correct while transitioning to State 4 to end the dialogue scenario.

  As the determination method described above, a method of determining whether or not there is feedback in the state transition can be used. FIG. 3 shows a state transition diagram in a dialogue scenario for purchasing a ticket. As shown in FIG. 3, the dialogue scenario starts in state 1, and a system utterance “Please name the destination station” is output, and the state transitions to state 2.

  In state 2, the input utterance is recognized as speech, the speech recognition result is stored in RAM 13, and the state transitions to state 1a. When the stored speech recognition result is “XX station”, the system utterance “Is it an XX station” and the system utterance “Is it an adult or a child” are output in the dialogue scenario, and the state 2a Transition to.

  In the state 2a, the input utterance is recognized by speech and the speech recognition result is stored in the RAM 13. When the speech recognition result is “ΔΔ” which is neither “adult” nor “child”, the state transitions (feeds back) to state 1. As described above, when the state transition includes a state transition that goes against the dialogue scenario information, it can be determined that the speech recognition result in the state 2 or the state 2a is not correct. In addition, it is possible to change the determination criteria such as determining that the speech recognition result is not correct only when there are continuous state transitions in the same location that reverse to the dialogue scenario information.

  Further, it is possible to use a method of accumulating the number of times the speech recognition result is corrected based on the state transition history and determining whether the speech recognition result is correct according to the accumulated number. In FIG. 3, when the speech recognition result in the state 2a is “adult” or “child”, the state transitions to the state 1b, and a system utterance “I am an adult” or “I am a child” is output. After the system utterance “please number” is output, the state transitions to state 2b.

  In the state 2b, the input utterance is recognized by speech and the speech recognition result is stored in the RAM 13. When the speech recognition result is “◎ sheets”, a system utterance “◎ is this” is output and the state transitions to state 3.

  In state 3, the input utterance is recognized as speech, and the speech recognition result is stored in RAM 13. Since it can be determined that the voice recognition result is “Yes” or “No” in State 3, the reliability of the voice recognition result in State 3 is high. When the stored speech recognition result is “No”, the state is changed to the state 1b, and the speech recognition result is corrected by performing the utterance to input the number of tickets again.

  In this way, the number of times the speech recognition result is corrected is accumulated, and when the accumulated number is equal to or less than a predetermined number, it is determined that the speech recognition result is correct. That is, if the number of times the speaker has corrected the error in the speech recognition result is small, it can be determined that the speech recognition engine is outputting the correct recognition result.

  FIG. 4 is another state transition diagram in a dialogue scenario for purchasing a ticket. As shown in FIG. 4, it is finally determined whether or not the speech recognition result is correct only once. When it is determined that the speech recognition result is correct, all the speech recognition results in the respective states passed so far are correct. It can also be judged. In FIG. 4, the dialogue scenario starts in state 1, a system utterance “please name the destination station” is output, and the state transitions to state 2.

  In state 2, the input utterance is recognized as speech, the speech recognition result is stored in RAM 13, and the state transitions to state 1a. When the stored speech recognition result is “XX station”, the system utterance “Is it an XX station” and the system utterance “Is it an adult or a child” are output in the dialogue scenario, and the state 2a Transition to.

  In the state 2a, the input utterance is recognized by speech and the speech recognition result is stored in the RAM 13. Regardless of whether the speech recognition result is “adult” or “children”, if the speech recognition result is “ΔΔ”, the state transitions to state 1b. In the state 1b, after the system utterance “please give the number of tickets” is output, the state transitions to the state 2b.

  In the state 2b, the input utterance is recognized by speech and the speech recognition result is stored in the RAM 13. When the speech recognition result is “◎”, a system utterance that summarizes the speech recognition results in the previous state is output. For example, the system utterance “To the station XX, △△ ticket ◎ sheets. If there is an error, please specify the station name, type, number of sheets, and what to correct.” Is output to state 5 Transition.

  In state 5, the input utterance is recognized as speech, and the speech recognition result is stored in RAM 13. In state 5, when the speech recognition result is "station name", the state transitions to state 1, to "type", to state 1a, and to "number of sheets", to state 1b. If the voice recognition result is “Yes”, the process goes to the state 6 to end the process. That is, when the reliability of the speech recognition result in state 5 is determined to be high and the speech recognition result is “Yes”, all the speeches in state 1, state 1a, state 1b, and state 5 that have transitioned so far Judge that the recognition result is correct.

  Furthermore, the recognition evaluation value output from the speech recognition engine can be used in combination. In this case, when the evaluation value of the speech recognition result in units of sentences or words is higher than a predetermined threshold value, it is determined that the speech recognition result is correct. That is, it is determined only by the evaluation value of speech recognition, not the content of dialogue. Therefore, it goes without saying that the accuracy of determining whether or not the voice recognition result is correct is further improved by using it together with the method described above.

  When it is determined that the speech recognition result is correct, the speech recognition result is stored in the utterance data storage unit 121 of the storage unit 12 as utterance data associated with the stored utterance. When it is determined that the voice recognition result is not correct, the voice recognition result and the speech are deleted from the storage unit 12. Note that “utterance data” means the entire data used for speech recognition including information on utterances and associated utterance contents.

  As described above, the speech recognition result is correctly stored in the speech data storage unit 121 as speech data by determining whether or not the speech recognition result image is correct based on the dialogue between the speaker and the computer using the speech dialogue apparatus. You can select only to collect as utterance data.

  Also, in a voice interaction device, since the speaker performs a natural conversation, it is possible to eliminate unnaturalness like when reading a manuscript and unconsciously collect speech data for speech recognition that is close to natural conversation. Can do. Therefore, it is possible to easily construct an acoustic model having a high speech recognition rate that matches a normal dialogue environment. In addition, it is possible to avoid the difference between the collected speech data and the usage environment of the speech recognition device from any viewpoint, such as the speech form, the characteristics of the input system, the distribution of the speaker's age group, etc. It becomes.

  Note that since the state transition history of the dialogue is stored in the storage means 12 or the RAM 13, the collection of the utterance data by the above-described processing does not necessarily have to be in real time, and after completion of the voice dialogue according to the dialogue scenario, Utterance data may be collected.

  FIG. 5 is an exemplary diagram of a data configuration stored in the utterance data storage unit 121. In FIG. 5, a case where “computer” is spoken will be described. Hereinafter, the utterance will be described as an example up to FIG. In the example of FIG. 5, utterances are stored as time-series data, and phoneme symbols as recognition results are given together with start points and end points. Note that the start point and the end point are indicated by accumulated values of the number of samples depending on the sampling frequency.

  FIG. 6 is an exemplary diagram showing the relationship between phonemes and time points in the data structure stored as utterance data. As shown in FIG. 6, a start point and an end point are stored for each phoneme, indicating which part of the utterance shown in FIG. 5 corresponds to each phoneme.

  Further, as shown in FIG. 7A, the start point and the end point may be stored for each syllable instead of the phoneme, or as shown in FIG. 7B, the start point and the end point for each phrase. May be stored. When the user utterance is one utterance unit, the start point and the end point are not specified for the user utterance including the silent period before and after the utterance cut out by the speech recognition apparatus, and the phoneme and syllable as the recognition result are specified. Etc. may be associated with each other.

  In the first embodiment, the start point and the end point are shown based on the number of samples depending on the sampling frequency. However, the present invention is not particularly limited to this. It may be used.

  In addition, waveform data is used as an utterance constituting the utterance data. However, the present invention is not particularly limited to this, and for example, voice feature quantities used for voice recognition such as voice spectrum, MFCC (Mel-Frequency Cepstral Co-efficients), etc. It is also possible to use.

  Furthermore, the utterance data generated and stored for speech recognition need not be in utterance units, and the result of speech recognition of only a part of the utterance is associated with the speech data of the corresponding section as in the word spotting process. May be stored.

  Next, processing of a computer program that embodies the speech recognition speech data collection apparatus according to Embodiment 1 of the present invention will be described. FIG. 8 is a flowchart of a computer program that embodies the speech recognition speech data collection apparatus according to Embodiment 1 of the present invention.

  In FIG. 8, first, a message that prompts the speaker to speak is output along the dialogue scenario information (step S801). Then, an utterance input for the message is accepted (step S802).

  Next, speech recognition processing is performed on the input utterance (step S803), and whether or not the input utterance has been correctly recognized based on the speech state transition history based on the speech recognition result and the dialog scenario information. Judgment is made (step S804).

  If it is determined that the speech recognition result is correctly recognized (step S804: YES), the input utterance and the speech recognition result are associated with each other and stored as one utterance data (step S805).

  The above-described processing is continued until the dialogue scenario information is finished (step S806), and when the dialogue scenario information is finished (step S806: YES), the collection of the utterance data is finished.

  In the first embodiment, utterance data is collected from one voice interactive device. However, the number of voice interactive devices is not limited to one, and utterance data accumulated by a plurality of voice interactive devices are aggregated. It is also possible to do. By collecting utterance data with a plurality of voice interactive devices, a larger amount of utterance data can be collected, and an acoustic model with high voice recognition accuracy can be constructed.

  In the first embodiment, whether or not the speech recognition result is correct is determined by focusing on the state transition. However, the present embodiment is not limited to the method focusing on the state transition. Any method can be used as long as it is secured. For example, there is a method in which a storage slot is provided in the storage unit 12 for each input item by an utterance used in a dialogue scenario, and the slot value, the slot value, and information for identifying the speech-recognized speech are stored in association with each other. Is possible.

  FIG. 9 is an explanatory diagram of slot management in the storage unit 12. In FIG. 9, a slot is assigned to each input item such as “station name”, “type”, “number of sheets”, etc. in the dialog scenario for purchasing a ticket, and the speech recognition result is voice-recognized as a slot value. The utterance ID is stored as information for identifying the utterance. By doing so, for example, when it is determined that the voice recognition of the station name is not correct and the voice is recognized again, the slot value is corrected to “Kochi” which is the correct recognition result, and the corresponding utterance ID is also corrected. Therefore, only the result of finally correctly recognized is stored in the storage unit 12.

(Embodiment 2)
A schematic configuration diagram of a computer embodying the speech recognition utterance data collecting apparatus according to the second embodiment of the present invention is the same as that of the first embodiment. The second embodiment is different in that a speaker recognition unit using voice is provided and the speaker can be confirmed based on the utterance.

  Information about the speaker is stored in advance in the speaker information storage unit 122 of the storage unit 12. The stored information regarding the speaker is personal information including at least information for identifying the speaker and other information such as gender, age, residential area, nationality, and the like. The information about the speaker is not limited to storing in advance in the speaker information storage unit 122, and information found in the conversation scenario information of the voice interaction device may be added as needed. . The speaker information storage unit 122 is a recording medium capable of storing a large amount of data such as a hard disk built in not only the storage unit 12 but also another computer connected via the communication unit 14. If it exists, it may be provided for anything.

  When an utterance by a speaker is input, the utterance is recognized by voice, and information for identifying the corresponding speaker, for example, a user ID, in order to confirm who the speaker is registered is, for example, a user ID To get. Then, when the information about the speaker associated with the user ID is extracted from the speaker information storage unit 122 of the storage unit 12 and the speech data is generated, the speech and voice recognition result and the extracted speaker are related. Information is stored in the speech data storage unit 121 as a pair of speech data.

  By doing this, speaker data stored for speech recognition can be collected by conditions such as gender, age, and residential area, and an acoustic model corresponding to the conditions of the speaker can be created Can do. Therefore, it is possible to efficiently collect utterance data that can be expected to improve speech recognition accuracy under specific conditions.

(Embodiment 3)
A schematic configuration diagram of a computer embodying the speech recognition utterance data collecting apparatus according to the third embodiment of the present invention is the same as that of the first embodiment. The third embodiment is different in that user utterances are input using various telephone lines.

  That is, the input means 15 is connected to a telephone line such as a fixed telephone, a cellular phone, a PHS, and an IP telephone, and the input utterance is input in a state in which each line is subjected to a specific voice processing. Therefore, when the speech data is collected, the information for identifying which speech line is used for the speech data is added, so that the speech recognition accuracy can be improved.

  In order to connect to the voice interactive apparatus, the speaker calls a predetermined telephone number connected to the voice interactive apparatus from a telephone such as a fixed telephone at home or a mobile phone owned by the speaker. When the computer receives a call, it connects to the line, and at the same time, the caller's caller ID information (whether or not the caller ID is notified and, if it is notified, the notified caller ID) is stored in the storage device 12 or the memory 13 is stored. A voice message is output along with the dialogue scenario information, and the utterance data of the utterer is accumulated as in the first embodiment.

  When the calling number is notified at the time of calling by the speaker, the line CODEC type or carrier (telephone carrier) type can be estimated based on the calling number. For example, when the first three digits of the transmission number are “090”, it can be estimated that the mobile phone is a mobile phone. Therefore, the utterance data is generated by associating the calling number information with the utterance and voice recognition result.

  In this way, since the line CODEC type or carrier type can be identified based on the transmission number, speech data can be classified for each line CODEC type or carrier type. Therefore, it is possible to generate an acoustic model in consideration of CODEC distortion generated for each line, and it can be expected that voice recognition accuracy is improved when the telephone line is used.

  On the other hand, when the calling number is not notified at the time of the call by the speaker, an acoustic model for line determination is stored in the storage unit 12 as an acoustic model, and which acoustic model matches the input utterance It is possible to determine whether the used line is a fixed phone, a mobile phone (PDC, W-CDMA, etc.), a PHS, an IP phone, or the like.

  In this way, it is possible to specify an acoustic model corresponding to each CODEC type or carrier type of the line, and by using an acoustic model that takes into account the CODEC distortion that occurs for each line, The improvement of voice recognition accuracy can be expected.

  As described above, according to the third embodiment, it is possible to collect speech data classified by line information that can be acquired for each conversation, and by using an acoustic model corresponding to the line to be used, It is possible to improve the accuracy of voice recognition.

  In addition to generating the utterance data in association with the utterance and voice recognition result in association with the transmission number information, it is desirable to generate the utterance data in association with the time when the utterance was recorded. For example, if the CODEC type changes with the times, such as mobile phones, it should be avoided to use old utterances for learning and improving acoustic models, and to select utterances that should be excluded from acoustic model learning data. As the information, the data regarding the time when the utterance was recorded is valid.

  Note that if learning of an acoustic model is performed based on a small amount of utterance data, it becomes difficult to generate an acoustic model effective for an unspecified speaker. Therefore, means for detecting the amount of data in which the speech data is stored in the speech data storage unit 121 is provided, and learning of the acoustic model is started when the amount of stored speech data exceeds a predetermined threshold.

  The means for detecting the accumulated amount of utterance data is not limited to a means for detecting the total amount of stored data, and for example, an acoustic model may be regenerated at regular time intervals.

  In this way, the acoustic model can be updated quantitatively or periodically, and an acoustic model based on the latest utterance data can be generated. An acoustic model can be generated in consideration of fluctuations and the like. Therefore, it is possible to collect utterance data that further contributes to improving the accuracy of voice recognition.

(Appendix 1)
Means for storing dialogue scenario information describing the progress of the dialogue;
Means for accepting the input utterance,
Means for voice recognition of the input utterance;
Means for proceeding a dialogue based on the speech recognition result and the dialogue scenario information;
And a voice interaction device including means for outputting a response to the input utterance,
Means for storing a dialog state transition history based on the dialog scenario information;
Means for determining whether or not the inputted utterance is correctly recognized based on the voice recognition result and the state transition history;
A speech recognition utterance data collection device comprising: means for storing the speech recognition result and the input utterance in association with each other when the means determines that the speech recognition has been correctly performed.

(Appendix 2)
The utterance data collection device for speech recognition according to appendix 1, wherein the utterance stored in association with the speech recognition result is speech waveform data or an utterance feature amount obtained by acoustic analysis of the utterance.

(Appendix 3)
Means for correcting speech recognition results;
Means for accumulating the number of times the speech recognition result has been corrected;
Means for determining that the input utterance is an utterance corresponding to the progress procedure of the dialog described in the dialog scenario information when the cumulative number of corrected times is equal to or less than a predetermined number of times, The speech data collection device for speech recognition according to appendix 1 or 2.

(Appendix 4)
The dialog scenario information includes a description for executing a predetermined task when an utterance is input in accordance with the described procedure of the dialog,
Means for detecting execution of the task,
The means for determining whether or not the inputted utterance has been correctly recognized is the dialogue in which the inputted utterance is described in the dialogue scenario information when the means for detecting the execution of the task detects the execution of the task. The speech recognition utterance data collection device according to supplementary note 1 or 2, wherein the utterance data collection device is determined to be an utterance corresponding to the proceeding procedure.

(Appendix 5)
Means for storing information about the speaker including information for identifying the speaker;
Means for identifying a speaker based on the input utterance and information about the speaker;
The speech recognition according to any one of appendices 1 to 4, further comprising: means for storing information associated with the identified speaker in association with the speech recognition result and the input speech. Utterance data collection device.

(Appendix 6)
Means for receiving a calling number on a telephone line;
The speech recognition utterance data collection device according to any one of appendices 1 to 5, further comprising: a unit that stores the transmission number in association with the speech recognition result and the input utterance.

(Appendix 7)
Means for determining a line type or a carrier based on the calling number;
The speech recognition utterance data collection device according to appendix 6, further comprising means for storing a determination result of the line type or carrier in association with the speech recognition result and the input utterance.

(Appendix 8)
The speech for speech recognition according to any one of appendices 1 to 7, further comprising means for storing information related to a time when the speech is input in association with the speech recognition result and the input speech. Data collection device.

(Appendix 9)
Memorize dialogue scenario information describing the progress of the dialogue,
Accepts the input utterance,
Speech recognition of the input utterance,
Based on the speech recognition result and the dialogue scenario information, the dialogue proceeds,
Using a voice interaction method for outputting a response to the input utterance,
Storing a dialogue state transition history based on the dialogue scenario information;
Based on the speech recognition result and the state transition history, it is determined whether or not the input utterance is correctly recognized,
A speech recognition utterance data collection method characterized by storing the speech recognition result and the input utterance in association with each other when the means determines that the speech recognition has been correctly performed.

(Appendix 10)
Computer
Means for storing dialogue scenario information describing the progress of the dialogue;
Means for accepting the input utterance;
Means for voice recognition of the input utterance;
Means for proceeding a dialogue based on the speech recognition result and the dialogue scenario information;
And a voice interaction device including means for outputting a response to the utterance;
Means for storing a dialog state transition history based on the dialog scenario information;
Means for determining whether or not the inputted utterance is correctly recognized based on the voice recognition result and the state transition history;
A computer program that causes the voice recognition result and the inputted utterance to be stored in association with each other when it is determined that the means has been correctly recognized.

It is a schematic block diagram of the computer which embodies the speech data collection apparatus for speech recognition which concerns on Embodiment 1 of this invention. It is a state transition diagram in the dialogue scenario for confirming the name. It is a state transition diagram in the dialogue scenario which purchases a ticket. It is another state transition diagram in a dialogue scenario for purchasing a ticket. It is an illustration figure of the data structure memorize | stored in an utterance data storage part. It is an illustration showing the relationship between phonemes and time points. It is an illustration figure which shows the relationship between a syllable or a word, and a time point. It is a flowchart of the program used with the speech data collection apparatus for speech recognition which concerns on Embodiment 1 of this invention. An explanatory diagram of slot management in the storage means is shown. It is a functional block diagram which shows schematic structure of the conventional speech data collection apparatus for speech recognition.

Explanation of symbols

11 CPU
12 storage means 13 RAM
DESCRIPTION OF SYMBOLS 14 Communication means 15 Input means 16 Output means 17 Auxiliary storage means 18 Portable recording medium 121 Utterance data storage part 122 Speaker information storage part

Claims (6)

Means for storing dialogue scenario information including a description for executing a process to be performed next in accordance with the content of the utterance when an utterance in accordance with the dialogue progress procedure is input;
An utterance data storage unit that associates the speech recognition result and the input utterance and is stored as utterance data for speech recognition;
Means for accepting the input utterance;
Means for recognizing the input speech using speech data stored in the speech data storage means;
Means for proceeding a dialogue based on a speech recognition result and the dialogue scenario information;
Detecting means for detecting execution of the processing ;
A voice interactive apparatus including means for outputting a response to the input utterance;
Means for storing a dialog state transition history based on the dialog scenario information;
When the detection unit detects the execution of the process, the input unit determines that the input utterance is an utterance corresponding to the progress procedure of the dialog described in the dialog scenario information, and the detection unit executes the process . If not detected, based on the voice recognition result and the state transition history, a determination means for determining whether or not the input utterance has been correctly recognized;
Speech recognition data collection for speech recognition, comprising: means for storing the speech recognition result and the input utterance in association with each other when the judgment means determines that the speech recognition result is correctly recognized. apparatus. 2. The speech recognition utterance data collection device according to claim 1, wherein the utterance stored in association with the speech recognition result is speech waveform data or an utterance feature amount obtained by acoustic analysis of the utterance. Means for receiving and storing a calling number on a telephone line;
The speech recognition utterance data collection device according to claim 1, further comprising: a unit that stores the transmission number in association with the speech recognition result and the input utterance. Using a computer
Storing dialogue scenario information including a description for executing a process to be performed next in accordance with the content of the utterance when an utterance in accordance with the dialogue progress procedure is input;
Associating the speech recognition result and the input utterance, and storing the speech data storing step stored as speech data for speech recognition;
Accepting the input utterance,
Recognizing the input utterance using the utterance data stored in the utterance data storage step; and
Advancing a dialogue based on the speech recognition result and the dialogue scenario information;
A detection step of detecting execution of the process ;
Using a step of outputting a response to the input utterance and a voice interaction method;
Storing a dialogue state transition history based on the dialogue scenario information;
When the detection step detects the execution of the process , it is determined that the input utterance is an utterance corresponding to the progress procedure of the dialog described in the dialog scenario information, and the detection step detects the execution of the process. If not, a determination step of determining whether or not the input utterance is correctly recognized based on the voice recognition result and the state transition history;
A speech recognition utterance data collection method for executing a step of storing the speech recognition result and the inputted utterance in association with each other in the speech data storage step when it is determined that the speech is recognized correctly as a result of the determination in the determination step. Computer
Means for storing the described dialogue scenario information including a description for executing a process to be performed next in accordance with the content of the utterance when an utterance in accordance with the progress procedure of the dialogue is input;
An utterance data storage unit that associates the speech recognition result and the input utterance and is stored as utterance data for speech recognition;
Means for accepting the input utterance;
Means for recognizing the input speech using speech data stored in the speech data storage means;
Means for advancing a dialogue based on the speech recognition result and the dialogue scenario information;
Detecting means for detecting execution of the processing ;
A voice interactive apparatus including means for outputting a response to the input utterance;
Means for storing a dialog state transition history based on the dialog scenario information;
When the detection unit detects the execution of the process , the input unit determines that the input utterance is an utterance corresponding to the progress procedure of the dialog described in the dialog scenario information, and the detection unit detects the execution of the process. If not, based on the speech recognition result and the state transition history,
Determining means for determining whether or not the input utterance is correctly recognized;
A computer program that functions as means for storing the speech recognition result and the inputted utterance in association with each other when the judgment means judges that the speech is recognized correctly. Means for storing dialogue scenario information describing the progress of the dialogue;
Means for accepting the input utterance,
Means for voice recognition of the input utterance;
Means for proceeding a dialogue based on a speech recognition result and the dialogue scenario information;
And a voice interaction device including means for outputting a response to the input utterance,
Means for storing a dialog state transition history based on the dialog scenario information;
Means for determining whether or not the inputted utterance is correctly recognized based on the voice recognition result and the state transition history;
Means for storing the voice recognition result and the inputted utterance in association with each other when it is determined that the means is correctly recognized;
Means for receiving and storing a calling number on a telephone line;
A speech recognition utterance data collection device comprising: means for storing the transmission number in association with the speech recognition result and the input utterance.

Images