JPH07210193A - Voice conversation device - Google Patents

Abstract

PURPOSE:To provide a voice conversation device capable of storing a new word into a machine while performing voice conversation. CONSTITUTION:By providing a phoneme/syllable recognition part 3 performing voice recognition and outputting a phoneme or syllable line, a phoneme/syllable lattice candidate storage part 4 storing the output, a voice synthesis part 5 synthesizing a sound signal from a phoneme/syllable lattice, a voice output part 6, a word dictionary 8 storing a phonetic symbol and a meaning of a word and a conversation control part 7, this device compares the last recognition result in the phoneme/syllable lattice candidate storage part with the latest recognition result being the output of the phoneme/syllable recognition part 3, and replaces the extremely different phoneme/syllable or phoneme group/ syllable group to the candidates different from the last time to output to the voice synthesis part 5, and judges a synthesis voice outputting immediately before as the right answer, and stores the phoneme/syllable lattice in the word dictionary 8, and the acquisition of language are realized by the voice conversation.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice dialogue device for controlling voice equipment while interacting with the equipment.

[0002]

2. Description of the Related Art The use of voice interaction devices for performing control while interacting with devices using voice is gradually expanding from toys such as chattering dolls to bank balance inquiry using a telephone.

FIG. 4 shows an example of the configuration of a conventional voice dialog device. In the figure, a voice input unit 1 converts a voice into an electric signal, and a voice analysis unit 2 converts the voice signal into an FFT or LPC.
A part that is converted into speech feature information such as a frequency spectrum or a cepstrum by analysis, a phoneme / syllable recognition unit 3 performs speech recognition based on the speech feature information of the speech analysis unit, and outputs a phoneme or a syllable string, a dialogue control unit. Reference numeral 12 is a portion for managing voice dialogue, recognizing a control word from the output of the phoneme / syllable recognition portion to judge the next operation, keyboard portion 13 is a portion for inputting information that cannot be input by voice, and voice synthesizing portion 7 Is the part that synthesizes the speech signal from the phoneme / syllable lattice,
The voice output unit 8 is a unit that converts the output of the voice synthesis unit into an acoustic signal, and the word dictionary 10 is a unit that stores phonetic symbols and meanings of words.

As an example, consider a schedule management system using the above-described voice interaction device. First, the dialogue management unit 12 sends the text "Please schedule" to the voice synthesis unit 7, and the voice synthesis unit 7 and the voice output unit 8 inquire as "Please schedule" with a synthetic voice. In response to this, the user
Speaking of "time meeting", this voice is converted into voice feature information by the voice input unit 1 and the voice analysis unit 2, converted into phoneme / syllable lattices by the phoneme / syllable recognition unit 3, and the dialogue control unit 1 is used.
2 recognizes the words "13:00" and "meeting" with reference to the word dictionary 10, and as the next operation,
The text “Please input the meeting partner with the keyboard” is sent to the voice synthesizer 7. The voice synthesizer 7 and the voice output unit 8 inquire about this text with a synthetic voice. When the user inputs "Matsushita" on the keyboard unit 13, the dialogue control unit 12 generates the text "A meeting with Matsushita-san at 13:00?", And when the user says "Yes", the dialogue control unit Stores the content of "Meeting with Matsushita-san at 13:00" in the database.

The problem with this conventional example is that when a proper noun such as "Matsushita" is input, it is usually not in the word dictionary 10 and voice recognition cannot be performed, so that it depends on a keyboard or the like. That is, at present, the number of words that humans can use in speech is limited, and it is difficult for machines to learn new words through dialogue like humans. Therefore, in order to be able to apply the voice interaction technology, it is only possible to control with a limited number of words, and in other cases, input means such as a keyboard is required.

[0006]

As explained in the above-mentioned conventional example, the number of words that a human can use in a voice is limited in the current voice dialog device, and the machine can use a new word through a dialogue like a human. Is difficult to learn. Therefore, an input means such as a keyboard is required for inputting proper nouns such as names.

An object of the present invention is to provide a voice dialogue system in which a machine can learn a new word through dialogue like a human being in view of the above problems of the conventional voice dialogue system.

[0008]

A voice interactive apparatus according to the present invention comprises a voice input unit for converting voice into an electric signal, and a voice signal from the voice input unit such as a frequency spectrum or a cepstrum by FFT or LPC analysis. A voice analysis unit for converting into characteristic information, a phoneme / syllabic recognition unit for performing voice recognition based on the output of the voice analysis unit and outputting a phoneme or a syllable string, and a phoneme for storing the output of the phoneme / syllable recognition unit /
Syllable lattice candidate storage unit, voice synthesizing unit for synthesizing voice signal from phoneme / syllable lattice, voice output unit for converting output of the voice synthesizing unit into acoustic signal, and word dictionary for storing phonetic symbols and meanings of words And manage the voice dialogue, "Yes"
・ When a control word such as “No” should be input, the control word is recognized from the output of the phoneme / syllable recognition unit, the next operation is judged, and when a new word that is not a control word is input for the first time, The phoneme / syllable lattice with the highest likelihood is selected from the contents of the phoneme / syllable lattice candidate storage unit and output to the speech synthesis unit, and when a new word should be reworded and input, the phoneme / syllable lattice is selected. The previous recognition result in the candidate storage unit is compared with the latest recognition result output from the phoneme / syllable recognition unit, and the most different phoneme / syllable or phoneme group / syllable group is replaced with a candidate different from the previous one. When an affirmative word such as “Yes” is input as a control word to the synthesizing section, the synthesized speech output immediately before is determined as the correct answer, and the phoneme / syllable lattice is stored in the word dictionary. Is shall.

The present invention also relates to a power analysis unit connected to the voice input unit for analyzing the power of the input voice per unit time, a power analysis result storage unit for storing the analysis result, and the power analysis result storage. Output of the power analysis unit and a power comparison unit that compares the previous analysis result of the power analysis unit with the latest analysis result that is output from the power analysis unit To a speech element / syllable or a phoneme group / syllable group having the most different power value from the above to a candidate different from the previous one, and a dialogue control section for outputting to the speech synthesis section is further provided.

The present invention also includes a pitch analysis unit connected to the voice input unit for analyzing the pitch frequency of the input voice,
A pitch analysis result storage unit that stores the analysis result, a pitch comparison unit that compares the previous analysis result in the pitch analysis result storage unit and the latest analysis result that is the output of the pitch analysis unit, and a new one. When a word is to be rephrased and input, a dialogue control unit which replaces a phoneme / syllable or a phoneme group / syllable group having the most different pitch frequency with a candidate different from the previous one from the output of the pitch comparison unit and outputs to the speech synthesis unit Is further provided.

[0011]

In the present invention, the voice input unit, the voice analysis unit and the phoneme /
By including a syllable recognition unit, a phoneme / syllable lattice candidate storage unit, a voice synthesis unit, a voice output unit, a word dictionary, and a dialogue control unit, even when a new word is input, it is converted into a phoneme or syllable lattice candidate sequence and the user Re-synthesize and request confirmation, and if wrong, utter again and present a recognition candidate different from the previous candidate again, and write the phoneme or syllable string newly in the word dictionary with user confirmation. New words can be acquired.

Further, by providing a power analysis section for analyzing the power per unit time of the input voice, a power analysis result storage section, and a power comparison section for comparing the previous analysis result and the latest analysis result, When a word should be reworded and input, it is possible to correct the candidate more accurately by replacing the phoneme / syllable or phoneme group / syllable group with the most different power value from the output of the power comparison unit with a candidate different from the previous time. become.

By further comprising a pitch analysis section for analyzing the pitch frequency of the input voice, a pitch analysis result storage section for storing the analysis result, and a pitch comparison section for comparing the previous analysis result and the latest analysis result. , When a new word should be reworded and input, the more accurate correction of the candidate by replacing the phoneme / syllable or the phoneme group / syllable group having the most different pitch frequency from the output of the pitch comparison unit with a candidate different from the previous one. Will be possible.

[0014]

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

FIG. 1 shows the configuration of a voice dialog device according to an embodiment of the present invention. In the figure, a voice input unit 1 is a unit that converts voice into an electric signal, and a voice analysis unit 2
Is a portion that converts the audio signal from the audio input unit 1 into audio feature information such as frequency spectrum and cepstrum by FFT or LPC analysis. The phoneme / syllable recognition unit 3 is a unit that performs voice recognition based on the output of the voice analysis unit 2 and outputs a phoneme or a syllable string. The phoneme / syllable lattice candidate storage unit 4 is a unit that stores the output of the phoneme / syllable recognition unit 3. The voice synthesis unit 5 is a unit that synthesizes a voice signal from the phoneme / syllable lattice.
The voice output unit 6 is a unit that converts the output of the voice synthesis unit 5 into an acoustic signal. The word dictionary 8 is a part that stores the phonetic symbols and meanings of words. The dialogue control unit 7 is a part that manages a voice dialogue.

The operation of the speech dialogue system of the present embodiment having the above-mentioned configuration will be described below.

Consider a schedule management system using the above-described voice interaction device as in the conventional example. First, the dialogue control unit 7 according to the present invention sends the text "Please schedule" to the voice synthesizing unit 5, and the voice synthesizing unit 5 and the voice output unit 6 inquire as "Please schedule" with a synthetic sound. On the other hand, when the user says "13:00 meeting", the voice input unit 1
The voice analysis unit 2 converts this voice into voice feature information, and the phoneme / syllable recognition unit 3 converts it into a phoneme / syllable lattice, and the dialogue control unit 4 converts the words “13:00” and “meeting” into the word dictionary 8. By referring to and recognizing, as the next operation, the text “Please have a meeting partner.” Is sent to the voice synthesizer 5. The voice synthesis unit 5 and the voice output unit 6 inquire with a synthesized voice, "Please have a meeting partner." When the user utters "Matsushita", the voice input unit 1 and the voice analysis unit 2 convert the voice "Matsushita" into voice feature information, and the phoneme / syllable recognition unit 3 converts it into phoneme / syllable lattice. At this time, the dialogue control unit 7 recognizes from the progress of the dialogue that the word "Matsushita" is a person's name and is not in the word dictionary 8. The dialogue control unit 7 confirms that the phoneme / syllable lattice candidate storage unit 4 is empty and stores the result of the phoneme / syllable recognition unit 3 in the phoneme / syllable lattice candidate storage unit 4 and at the same time in the recognition candidate lattice. Using the candidate with the highest likelihood, for example, "13
Sometimes you have a meeting with Matsushika-san? Produces the text ". Here, it means that the candidate was "Matsushika". When the user says "Matsushita" again, as in the previous time, the voice "Matsushita" is converted into voice feature information by the voice input unit 1 and the voice analysis unit 2, and the phoneme / syllable recognition unit 3 makes a phoneme / syllable lattice. Is converted to. From the result of the phoneme / syllable recognition unit 3, the dialogue control unit 7 searches for a candidate with the highest likelihood other than “Matsushika”. Here it is assumed that it was "Matsushita". The dialogue control unit 7 again generates the text "Meeting with Matsushita-san at 13:00?" When the user says "Yes", the dialogue control unit 7 stores the content "Meeting with Matsushita-san at 13:00" in the database, and at the same time stores the word "Matsushita" in the word dictionary 8. At the same time, the system understands that the word "Matsushita" is a person's name, and also writes it in the word dictionary 8.

In this way, the new word "Matsushita" was acquired by the voice dialogue. Next, a second embodiment according to the present invention will be described with reference to FIG. In FIG. 2, a power analysis unit 9 is a part that analyzes the power of the input voice per unit time. The power analysis result storage unit 10 is a part that stores the analysis result of the power analysis unit 9. The power comparison unit 11 is a unit that compares the previous analysis result with the latest analysis result.

The second aspect of the present invention configured as described above
The operation of the voice dialogue apparatus of the embodiment will be described below.

Consider a schedule management system using the above-described voice interactive apparatus as in the first embodiment. First, the dialogue management unit 7 according to the present invention sends the text “Please schedule” to the voice synthesis unit 5, and the voice synthesis unit 5 and the voice output unit 6 are provided.
Asks "Please have a plan" with a synthetic voice. On the other hand, when the user says “13:00 meeting”, the voice input unit 1 and the voice analysis unit 2 convert this voice into voice feature information, the phoneme / syllable recognition unit 3 converts it into phoneme / syllable lattice, and the dialogue control unit 7 Recognizes the words "meeting" and "meeting" with reference to the word dictionary 8 and sends the text "Please have a meeting partner." To the speech synthesizer 5 as the next operation. The voice synthesis unit 5 and the voice output unit 6 inquire with a synthesized voice, "Please have a meeting partner." When the user utters "Matsushita", the voice input unit 1 and the voice analysis unit 2 convert the voice "Matsushita" into voice feature information, and the phoneme / syllable recognition unit 3 converts it into phoneme / syllable lattice. At the same time, the power analysis unit 9 analyzes the power value and stores it in the power analysis result storage unit 10. At this time, the dialogue control unit 7 recognizes from the progress of the dialogue that the word "Matsushita" is a person's name and is not in the word dictionary 8. The dialogue control unit 7 confirms that the phoneme / syllable lattice candidate storage unit 4 is empty and stores the result of the phoneme / syllable recognition unit 3 in the phoneme / syllable lattice candidate storage unit 4 and at the same time in the recognition candidate lattice. Using the candidate with the highest likelihood, for example, the text “13:00, you have a meeting with Matsuka-san?” Is generated. Here, it means that the candidate was "Matsushika". When the user says "Matsushita" again, as in the previous time, the voice "Matsushita" is converted into the voice feature information by the voice input unit 1 and the voice analysis unit 2, and the phoneme / phoneme / syllabic unit recognition unit 3
Converted to syllable lattice. Further, the power analysis unit 9 analyzes the power value again, the power comparison unit 11 compares the output of the power analysis unit with the previous value of the power analysis result storage unit 10, and sends the result to the dialogue control unit 7. The dialogue control unit 7 replaces the syllable having the largest power value difference from the result of the phoneme / syllable recognition unit 3 and the comparison result of the power comparison unit 11 to search for a candidate with the highest likelihood other than “Matsushika”. FIG. 4 is an example showing this difference in power. Here, since the difference in "ta" is the clearest, the dialogue control unit 7 replaces the "ka" in "Matsushika" with the next candidate. Here it is assumed that it was "Matsushita". The dialogue control unit 7 again generates the text "Meeting with Matsushita-san at 13:00?" If the user says yes,
The dialogue control unit 7 stores the content “Meeting with Matsushita-san at 13:00” in the database, and at the same time stores the word “Matsushita” in the word dictionary 8. At the same time, the system understands that the word "Matsushita" is a person's name, and also writes it in the word dictionary 8.

By using the power information in this way, the new word "Matsushita" can be more accurately determined.

This can be done by using pitch information instead of power, and the third embodiment according to the present invention shown in FIG. 3 is possible. In FIG. 3, the pitch analysis unit 12
Is the part that analyzes the pitch of the input voice. The pitch analysis result storage unit 13 is a unit that stores the analysis result of the pitch analysis unit 12. The pitch comparison unit 14 is a unit that compares the previous analysis result with the latest analysis result.

The third aspect of the present invention configured as described above
The voice interaction device of the second embodiment performs the same operation as that of the second embodiment.

Further, by combining both pitch and power, more accurate estimation becomes possible.

[0025]

As described above, according to the present invention, the machine is provided with a new word by providing the dialogue control means for continuously performing the speech synthesis of the phoneme / syllabic lattice presumed to be the phoneme / syllable recognition means and the question-and-answer operation. Can be acquired by voice dialogue. In addition, by comparing the power value of the word that is reworded and input with the word of the previous utterance, and replacing an appropriate syllable candidate, more accurate estimation of the candidate becomes possible. Further, if pitch information or pitch and power information is used, more accurate estimation can be performed and unknown words can be acquired more smoothly.

[Brief description of drawings]

FIG. 1 is a block diagram of a voice dialog device according to a first embodiment of the present invention.

FIG. 2 is a block diagram of a voice dialog device according to a second embodiment of the present invention.

FIG. 3 is a block diagram of a voice dialog device according to a third embodiment of the present invention.

FIG. 4 is a diagram showing a difference in power between two vocalizations of the word “Matsushita”.

FIG. 5 is a block diagram of a conventional voice dialog device.

[Explanation of symbols]

 1 voice input unit (means) 2 voice analysis unit (means) 3 phoneme / syllable recognition unit (means) 4 phoneme / syllable lattice candidate storage unit (means) 5 voice synthesis unit (means) 6 voice output unit (means) 7 dialogue Control unit (means) 8 Word dictionary 9 Power analysis unit (means) 10 Power analysis result storage unit (means) 11 Power comparison unit (means) 12 Pitch analysis unit (means) 13 Pitch analysis result storage unit (means) 14 Pitch comparison Department (means)

Claims (3)

[Claims] 1. A voice input unit for converting voice into an electric signal,
A voice analysis unit that converts a voice signal from the voice input unit into voice feature information such as a frequency spectrum and a cepstrum by FFT or LPC analysis, and performs voice recognition based on the output of the voice analysis unit and outputs a phoneme or a syllable string. Phoneme
A syllable recognition unit, a phoneme / syllable lattice candidate storage unit that stores the output of the phoneme / syllable recognition unit, a speech synthesis unit that synthesizes a speech signal from the phoneme / syllabic lattice, and an output of the speech synthesis unit as an acoustic signal. A voice output unit for conversion, a word dictionary for storing phonetic symbols and meanings of words, and a phoneme / syllable recognition unit for managing a voice dialogue and inputting a control word such as “Yes” or “No”. When a new word that is not a control word should be input for the first time by recognizing the control word from the output and recognizing the next operation, the phoneme / syllable lattice with the highest likelihood is selected from the contents of the phoneme / syllable lattice candidate storage unit. When it is output to the speech synthesizer and a new word should be reworded and input, the previous recognition result in the phoneme / syllable lattice candidate storage and the output of the phoneme / syllable recognizer are the latest. Recognition result of A comparison is performed and the most different phoneme / syllable or phoneme group / syllable group is replaced with a candidate different from the previous one, and the result is output to the speech synthesis unit, and when a positive word such as "Yes" is input as a control word, the synthesis output immediately before is output. A speech dialogue apparatus comprising: a dialogue control unit that judges a voice as a correct answer and stores the phoneme / syllable lattice in the word dictionary. 2. A power analysis unit connected to the voice input unit for analyzing the power of input voice per unit time, a power analysis result storage unit for storing the analysis result, and a power analysis result storage unit. The power comparison unit that compares the previous analysis result with the latest analysis result that is the output of the power analysis unit, and when a new word should be reworded and input, the most power is output from the power comparison unit. The speech dialogue system according to claim 1, further comprising: a dialogue control unit that replaces a phoneme / syllable or a phoneme group / syllable group having different values with a candidate different from the previous time and outputs the candidate to the speech synthesis unit. 3. A pitch analysis unit connected to the voice input unit for analyzing the pitch frequency of an input voice, a pitch analysis result storage unit for storing the analysis result, and a previous pitch analysis result storage unit for storing the pitch analysis result. A pitch comparison unit that compares the analysis result and the latest analysis result that is the output of the pitch analysis unit,
When a new word should be reworded and input, the phoneme / syllable or the phoneme group / syllable group having the most different pitch frequency from the output of the pitch comparison unit is replaced with a candidate different from the previous one, and the dialogue control is performed. The voice dialog device according to claim 1 or 2, further comprising a section.