JP2975772B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for controlling various devices by voice.

[0002]

2. Description of the Related Art In recent years, research and development of a voice recognition device capable of recognizing voice have been actively carried out, and practical use of this type of device is desired.

[0003] This type of apparatus generally performs data processing on a voice pattern such as that shown in FIG. 7, which comprises parameters representing voice characteristics obtained by analyzing voice, and is stored in advance for a plurality of voices. Each of the voice patterns (voice standard patterns) is compared with an unknown voice pattern by a method of pattern matching, and a voice standard pattern with the smallest error (that is, a high similarity) is found, so that a signal corresponding to this standard pattern is obtained. Is output as a recognition result.

In such a voice recognition device, a user first performs a registration operation of storing voice information to be recognized in a memory, and performs an original recognition process after completion of the registration. In this case, if the registered voice is not correctly registered, that is, if the information of the voice stored in the memory is wrong, it has caused a cause of erroneous recognition. Therefore, in order to improve the recognition performance, how to correctly store voice information (voice standard pattern) in the memory,
It is a big problem. Conventionally, in order to make the voice standard pattern correct, at the time of registration, the same voice must be spoken at least three times,
An audio standard pattern is created by a method of creating an audio standard pattern from the two patterns that are the most similar (detailed in Japanese Patent Publication No. 1-36639) or a method of checking an audio standard pattern once registered in a test mode or the like. The pattern was more accurate. In this case, it is necessary to utter a voice at least twice in order to register a voice pattern, and the registration is complicated.

As another method, an attempt has been made to correct a speech standard pattern using a recognition result. In this method, a recognition result is output, and a user inputs the fact that the result is correct to a voice recognition device using a switch or the like, and generates an average pattern obtained by averaging a voice standard pattern and an input voice pattern using the information. Then, processing for changing the voice standard pattern to this average pattern was performed. However, when this method is used, it is necessary for the user to input information on whether or not the recognition result is correct to the speech recognition device, which is not practical.

As another method of correcting a voice standard pattern using a recognition result, there is a method of outputting a recognition result and correcting the voice standard pattern without determining whether the result is correct. In this case, since the voice standard pattern is corrected even in the case of erroneous recognition, there is a possibility that the voice standard pattern is corrected instead to an incorrect voice standard pattern.

[0007]

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and an object of the present invention is to provide a voice recognition apparatus which simplifies a registration operation and efficiently corrects a voice standard pattern.

[0008]

A voice analysis unit for analyzing a voice input from a microphone, a pattern generation unit for generating a voice pattern based on a result analyzed by the voice analysis unit, and a plurality of voice standard patterns in advance. Is calculated, and the similarity difference between the voice pattern created by the pattern creation unit and each voice standard pattern stored in the standard pattern memory is calculated, and the most similar voice is calculated. A similarity calculation unit that outputs a standard pattern and its similarity, and a signal corresponding to the voice standard pattern when the similarity of the audio standard pattern obtained from the similarity calculation unit is greater than a preset threshold. Output as a recognition result, and when the similarity is smaller than the threshold value, store a judgment unit that judges rejection of recognition and a voice pattern judged to be rejection of recognition by the judgment unit. The first input pattern buffer memory and the voice standard pattern having the most similar voice pattern obtained by the second voice input when the determination unit determines that the recognition is rejected. Is the same as the most similar voice standard pattern, the second input pattern buffer memory for storing the voice pattern obtained by this re-input and the first or second input pattern buffer memory And a pattern correction unit that corrects the audio standard pattern in the audio standard pattern memory to which the audio pattern of the first or second input pattern buffer memory is most similar based on the audio pattern.

[0009]

According to the speech recognition apparatus of the present invention, even if there is an error in the speech standard pattern at the time of speech registration processing, the error speech standard pattern is corrected at the time of speech recognition processing. Can be.

[0010]

FIG. 1 shows the configuration of a speech recognition apparatus according to the present invention, and FIG. 2 shows the configuration of an embodiment of a main part of a pattern correction unit.

The configuration of the speech recognition apparatus shown in FIG. 1 is as follows. A speech analysis unit 2 for analyzing speech input from a microphone 1, a speech section is detected from feature parameters analyzed by the speech analysis unit, and speech patterns are formed. A pattern creation unit 3, a voice standard pattern memory 4 in which voice standard patterns are stored in advance, and matching between the unknown voice analyzed by the analysis unit and a voice standard pattern stored in the voice standard pattern memory, and for each voice standard pattern The similarity calculator 5 for calculating the similarity stores the similarity calculated for each voice standard pattern by the similarity calculator, selects the most similar voice standard pattern, and determines a criterion (hereinafter referred to as a threshold). ), A determination unit 6 for determining whether the recognition result is valid, and a pattern correction unit 7 for correcting the voice standard pattern based on the result of the determination unit.

In such a speech recognition device of FIG.
The feature of the present invention resides in a pattern correction unit 7, which has a first voice buffer 71 for storing a voice standard pattern to be corrected and a second voice for storing a voice pattern of an input voice, as shown in FIG. A buffer 72, a third audio buffer 73 for storing an audio pattern of the audio input again,
A pattern similarity calculator 74 for calculating the similarity between the audio patterns of each audio buffer; a rejection number storage 76 for storing the number of the audio standard pattern transmitted to the first audio buffer 71;
A pattern correction averaging unit 77 that performs a correction averaging process on the audio pattern based on the audio pattern in the first audio buffer 71, the second audio buffer 72, or the third audio buffer 73 is provided.

The operation of the apparatus according to the present invention when the configuration shown in FIG. 2 is employed in the pattern correction unit 7 shown in FIG. 1 will be described below.

[First Embodiment] A user first registers a voice. That is, the apparatus is set to a registration mode by a registration switch (not shown), and sounds to be sequentially registered are uttered toward the microphone 1. For example, say "zero". The voice input from the microphone 1 is converted into an electric signal, and the voice analyzer 2 extracts the voice as a feature parameter. For example, a general frequency analysis as shown in FIG. 7 is performed by a band pass filter or the like. The analyzed characteristic parameters are transmitted to the pattern creating unit 3, and the pattern creating unit 3 detects a voice section and performs voice patterning. The voice pattern patterned by the pattern creation unit is stored in a predetermined area of the voice standard pattern memory 4 as a voice standard pattern. Subsequently, "one", "d",... "Kyu" are sequentially stored in the voice standard pattern memory 4, and all registrations are performed.

Next, actual speech recognition will be described.
The case where the operator utters “zero” toward the microphone 1 will be described. The voice input from the microphone 1 is subjected to the same processing as in the registration mode, and the pattern generation unit 3 generates a voice pattern. In the recognition mode, this voice pattern is transmitted to the similarity calculation unit 5. The similarity calculation unit 5 calculates the similarity between the voice pattern created by the pattern creation unit 3 and each of the voice standard patterns stored in the voice standard pattern memory 4,
The similarity is transmitted to the determination unit 6. For example, the similarity is transmitted to the input voice “zero” as shown in FIG. Subsequently, the determination unit 6 determines the voice standard pattern that gives the maximum similarity and the similarity. For the input “zero”, the voice standard pattern that gives the maximum similarity is “zero”, and the similarity is 70, as shown in FIG. The determination unit 6 determines the threshold value and the degree of similarity that are set in advance, and determines the validity of the input voice. Here, the threshold is 80, and it is determined that recognition is rejected. The determination unit 6 notifies the pattern correction unit 7 to that effect. [This state is referred to as <state 1>] Here, if the maximum similarity is 90, it is determined that the recognition has been performed.
The determination unit 6 outputs a signal corresponding to the audio standard pattern. At this time, the pattern correction unit 7 does not correct the audio standard pattern.

Next, the processing of the pattern correction unit 7 will be described. When receiving the recognition rejection signal from the determination unit, the pattern correction unit 7 reads the maximum similarity, stores the number of the voice standard pattern that gives the maximum similarity in the rejection number storage unit 76, and stores the maximum number from the voice standard pattern memory 4. An audio standard pattern giving similarity is read into the first audio buffer 71, and an audio pattern of the input audio is read from the pattern creation unit 3 into the second audio buffer.

When it is determined that the input speech is rejected, the user usually utters the same word again. Here, it is assumed that “zero” is uttered again. Similarly, the similarity of this input is calculated as shown in FIG. As shown in FIG. 6, the voice standard pattern that gives the maximum similarity is "
The determination unit 6 determines that the value is “zero”. The pattern correction unit 7
Then, the number of the voice standard pattern giving the maximum similarity is transmitted to the rejection number storage unit 76. The rejection number storage unit 76
If the number already stored and the transmitted number match, the audio pattern created by the pattern creation unit is transmitted to the third audio buffer.

Next, the following calculation is performed in the pattern similarity calculator 74. The audio pattern of the first audio buffer is P ₁
(I, j). Set the audio pattern of the second audio buffer to P
₂ (i, j). The audio pattern of the third audio buffer is P ₃ (i, j). Let the modified pattern be P _ref (i, j). In this case, the similarity S ₁₂ between voice patterns of the first audio buffer and the second audio buffer,

[0019]

(Equation 1)

The similarity S ₁₃ between the audio patterns of the first audio buffer and the third audio buffer is

[0021]

(Equation 2)

The similarity S ₂₃ between voice patterns of the second audio buffer and the third audio buffer,

[0023]

(Equation 3)

The voice pattern having the largest value (similar one) among the calculation results S ₁₂ , S ₁₃ , and S ₂₃ is transmitted to the pattern correction averaging unit 77. The pattern correction averaging unit 77 performs an averaging process of two voice patterns as follows.

The average processing of the audio patterns in the first audio buffer and the second audio buffer is as follows.

[0026]

(Equation 4)

The average processing of the audio patterns of the first audio buffer and the third audio buffer is as follows.

[0028]

(Equation 5)

The average processing of the audio patterns in the second audio buffer and the third audio buffer is as follows.

[0030]

(Equation 6)

A correction pattern is created from such an averaging result. The created voice pattern is stored in the corresponding voice standard pattern area of the standard pattern memory 4 based on the number stored in the rejection number storage unit.

If the number already stored in the rejection number storage unit 76 and the transmitted number do not match, the contents of the first audio buffer, the second audio buffer, the third audio buffer, and the rejection number storage unit And the number giving the new maximum similarity is stored in the rejection number storage unit, the voice standard pattern giving the maximum similarity is stored in the first voice buffer, and the voice pattern created by the pattern creation unit 3 is stored in the second voice buffer. I do.

In the present embodiment, a new voice pattern is created based on two similar voice patterns, but without calculating the similarity, for example,

[0034]

(Equation 7)

As shown by the calculation formula, the first audio buffer,
A modified pattern may be created by averaging all the audio patterns of the second and third audio buffers.

In the speech recognition apparatus of the present invention, the number of input speech patterns (input pattern buffers) used is not limited to two, but may be N (for example, five). In this case, the voice input processing may be repeated N times.

[Embodiment 2] FIG. 3 shows the configuration of another embodiment of the pattern correction unit of the speech recognition apparatus of the present invention. 2 is different from that of FIG. 2 in that an input time measuring function 75 (hereinafter referred to as a timer) for starting time measurement after a recognition rejection result is obtained and measuring a time interval between re-inputs is added. On the point.

In the apparatus shown in FIG. 7, in the above-mentioned <state 1>, when receiving a signal of rejection of recognition from the judging section 6, the timer 75 starts counting time, and there is a speech input again. When the recognition rejection signal is received again, the timing is ended.
The time from the start to the end of the time measurement is a set value (for example, 10
A time of about seconds to 20 seconds is set in advance. )
If not, the pattern correction unit 7 corrects the voice standard pattern. This avoids erroneous correction when voice input after a predetermined time is not appropriate. That is,
After the first recognition rejection signal is generated, a device that allows speech input to obtain the second recognition rejection signal for an indefinite period of time may be used. When the cause is similar noise input, there is a problem that the voice standard pattern is erroneously corrected by the noise pattern. On the other hand, the device of the present invention has such a problem by providing the time limiting means as described above. Is less frequent.

[Embodiment 3] FIG. 4 shows the configuration of still another embodiment of the pattern correction unit of the speech recognition apparatus of the present invention. 2 differs from that of FIG. 2 in that the maximum similarity read after the recognition rejection result is obtained is compared with a predetermined value (hereinafter referred to as a second threshold). The point is that the section has been added.

In the state shown in FIG. 4, when the recognition rejection signal is received from the determination unit 6 in the state <State 1>, the second threshold value determination unit 78 determines whether the read maximum similarity is large or small. Make a comparison. The second threshold is set smaller than the threshold, and is set to, for example, '45'. If the maximum similarity is larger than the second threshold, the pattern correction unit 7 corrects the voice standard pattern. When the maximum similarity is equal to or less than the second threshold, the voice pattern input again has too low a similarity, so that the voice pattern is invalidated, and the same processing as the second voice input is performed for the third voice input.

In this embodiment, if the maximum similarity is equal to or less than the second threshold, the input voice pattern is invalidated, and
It is also conceivable to stop correcting the voice standard pattern instead of waiting for the second voice input.

[0042]

According to the speech recognition apparatus of the present invention, even if there is an error in the speech standard pattern at the time of the speech registration process, only the error speech standard pattern can be operated only by a simple operation at the time of speech recognition processing. Corrections can be made. Further, since the audio standard pattern is not corrected for an incorrect input audio pattern, a highly reliable audio standard pattern can be obtained.

[Brief description of the drawings]

FIG. 1 shows a configuration diagram of a speech recognition device of the present invention.

FIG. 2 shows a configuration of an embodiment of a pattern correction unit of the speech recognition device of the present invention.

FIG. 3 shows the configuration of another embodiment of the pattern correction unit of the speech recognition device of the present invention.

FIG. 4 shows the configuration of still another embodiment of the pattern correction unit of the speech recognition device of the present invention.

FIG. 5 shows an example of a similarity to an input voice.

FIG. 6 shows an example of similarity with respect to re-inputted speech.

FIG. 7 shows an example of an audio pattern (an audio pattern subjected to frequency analysis by a band-pass filter).

[Explanation of symbols]

 Reference Signs List 1 microphone 2 voice analysis unit 3 pattern creation unit 4 voice standard pattern memory 5 similarity calculation unit 6 determination unit 7 pattern correction unit 71 first voice buffer 72 second voice buffer 73 third voice buffer 74 pattern similarity calculation unit 75 timer 76 Rejection number storage unit 77 Pattern corrected average unit 78 Second threshold value judgment unit

──────────────────────────────────────────────────続 き Continued on front page (58) Field surveyed (Int.Cl. ⁶ , DB name) G10L 3/00 521 G10L 3/00 561

Claims (5)

(57) [Claims] An audio analysis unit that analyzes audio input from a microphone, a pattern creation unit that creates an audio pattern based on a result analyzed by the audio analysis unit, and a plurality of audio standard patterns stored in advance. A standard pattern memory, and a similarity difference between the audio pattern created by the pattern creation unit and each audio standard pattern stored in the standard pattern memory, and the most similar audio standard pattern and its A similarity calculator that outputs the similarity, and a signal corresponding to the voice standard pattern is output as a recognition result when the similarity of the voice standard pattern obtained from the similarity calculator is greater than a preset threshold. When the similarity is smaller than the threshold value, the determination unit determines that the recognition is rejected, and the first input pattern that stores the voice pattern determined to be rejected by the determination unit. If the pattern buffer memory and the determination unit determine that recognition is rejected, the voice standard pattern having the most similar voice pattern obtained by the second voice input is the voice pattern of the first input pattern buffer memory which is the most similar. When the voice pattern is the same as the similar voice standard pattern, the voice pattern is obtained based on the second input pattern buffer memory for storing the voice pattern obtained by the re-input and the voice pattern of the first or second input pattern buffer memory. And a pattern correction unit for correcting a voice standard pattern in the voice standard pattern memory having the most similar voice pattern in the first or second input pattern buffer memory. 2. A speech recognition apparatus according to claim 1, wherein said pattern correction unit comprises: a speech standard pattern for providing a maximum similarity; a speech pattern of a first input pattern buffer memory; and a second input pattern buffer memory. Create an average pattern by averaging two similar patterns among the audio patterns
A speech recognition apparatus characterized in that a speech standard pattern is changed to this average pattern. 3. The speech recognition apparatus according to claim 1, wherein the pattern correction unit starts timing after a recognition rejection result is obtained, and measures an input time interval between re-input. A speech recognition apparatus having a measurement function and correcting a speech standard pattern only when a measured time is equal to or shorter than a predetermined time. 4. The speech recognition device according to claim 1, wherein when the maximum similarity of the input speech pattern is smaller than a threshold, the maximum similarity is lower than a second threshold set lower than the threshold. A speech recognition device wherein a speech standard pattern is modified only when the maximum similarity is large. 5. The speech recognition device according to claim 1, 2 or 3,
In this case, the voice standard pattern obtained at the time of voice registration is stored in a voice standard pattern memory, and at the time of voice recognition, the voice standard pattern is stored in the memory based on N input voice patterns obtained from the same input voice as the voice standard pattern of the memory. A speech recognition device characterized by correcting a speech standard pattern.