JPH07219596A - Voice recognizer - Google Patents

Abstract

(57) [Abstract] [Purpose] To provide a speech recognition device that can be used in various environments with little reduction in recognition rate even in a noisy environment. A means 102 for dividing a voice into a plurality of frequency bands having different center frequencies, and a means 103 for obtaining the magnitude of the correlation included in each signal waveform obtained by the division.
And means 104 for detecting the power component of each signal waveform obtained by the means 102 for dividing into a plurality of frequency bands.
And a means 105 for normalizing a synchronizing frequency component included in each divided signal waveform by using the power component, and a means 106 for adding the normalized synchronizing signal components. apparatus.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus, and more particularly to a speech recognition apparatus which can be used even in a noisy environment without impairing the recognition performance.

[0002]

2. Description of the Related Art As a conventional voice recognition device, a structure as shown in FIG. 4 has been generally used. In FIG. 4, 401 is a microphone for inputting voice, and 402
Is a feature extractor for extracting the feature amount of the voice input by the microphone 401, and 403 is an identification unit for recognizing the input voice using the feature amount of the voice extracted by the feature extractor 402. . The operation of the conventional speech recognition apparatus configured as described above will be described below. In FIG. 4, the voice input from the microphone 401 for inputting voice is input to the feature extractor 402 and the feature amount of the voice is extracted. Conventionally known filter analysis, DFT (Discrete Fourier Transform) analysis, LPC (Linear Predictive Coding) analysis and the like have been used to extract the feature amount of the voice. The voice feature amount obtained by the voice feature extractor 402 is input to the identification unit 403 and a known DTW (Dinamic
Time Warping), HMM (Hidden Markoff Model), NN
It recognizes the input voice by various voice pattern identification methods such as (Neural Network).

[0003]

In the conventional speech recognition apparatus having the above-described configuration, known filter analysis, DFT (Discrete Fourier Transform) analysis, L
In the feature extractor 402 using PC (Linear Predictive Coding) analysis or the like, there is a problem that the feature analysis of voice under a large noise is not accurately performed and the recognition performance is deteriorated.

In view of the above points, the present invention has an object to provide a voice recognition apparatus which can obtain a high recognition rate even under a large noise.

[0005]

SUMMARY OF THE INVENTION The present invention comprises means for dividing speech into a plurality of frequency bands having different center frequencies,
A means for obtaining the magnitude of the correlation between the signal waveforms of the divided audio signal, a means for detecting the power component of each signal waveform of the divided audio signal, and a signal waveform of each of the signal waveforms of the divided audio signal. A unit for normalizing the magnitude of the correlation using the power component, a unit for adding the normalized magnitude of the correlation, and a discriminating unit for discriminating the input voice using the added signal. It is a voice recognition device characterized by the above.

[0006]

With the above-described structure, the present invention can be used in an environment with a large amount of noise without lowering the recognition rate of the voice recognition device, and the environment in which the voice recognition device is used can be greatly expanded.

[0007]

1 is a block diagram of a voice recognition apparatus according to a first embodiment of the present invention, in which 101 is a microphone for inputting voice, and 102 is voice in a plurality of frequency bands having different center frequencies. Filter bank for partitioning, 10
3 is an autocorrelator for obtaining an autocorrelation of each signal waveform of the voice signal frequency-analyzed by the filter, 104 is a power detection circuit for obtaining a power component of the voice signal waveform frequency-analyzed by the filter, and 105 is an autocorrelator 1
A divider for normalizing the output of No. 03 using the power of each filter output obtained by the power detector 104, 10
6 is an adder for adding the autocorrelation values of the respective filters normalized by the divider 105, and 106 is an adder 10
This is an identification unit for performing voice recognition using the voice feature amount obtained in 5.

The operation of the conventional speech recognition apparatus configured as described above will be described below. In FIG.
The voice input by the microphone 101 for inputting voice is input to the filter bank 102 and divided into voice frequency bands having a plurality of frequency bands having different center frequencies. The center frequency of this band filter bank is set by dividing the audio signal band logarithmically at equal intervals.The low-frequency components responsible for important components of audio information are fine, and the high-frequency components are roughly. The frequency of the voice is efficiently analyzed by analyzing. The audio signal analyzed by the filter bank 102 is input to the autocorrelator 103 and the power detector 104, and the periodicity of each band component included in each filter output and the power included in the filter are detected. After that, the signal is input to the divider 105 for normalization with the power component, and the output of the autocorrelator 103 is normalized according to the power of the filter output. The normalized autocorrelator outputs are input to the adder 106, and the normalized autocorrelation values of the filters are added. The identifying unit 107 uses the added value of the autocorrelation to recognize the input voice by DTW (Dinamic Time Warping) which is a known voice pattern identifying method. HMM (Hidden Markoff Model) and NN (Neural Neural Neural
Various voice pattern identification methods such as twork) can be used instead of DTW.

FIG. 2 shows the result of comparing the speech recognition rate in the noise environment of the speech recognition apparatus using the present invention described above with the recognition rate when the conventional recognition apparatus is used. The vertical axis of FIG. 2 represents the voice recognition rate, and the horizontal axis represents the S / N ratio indicating the noise level of the environment in which the voice is uttered. As a conventional recognition device, the result of a speech recognition device using DFT (Discrete Fourier Transform) which is a widely used speech analysis method is shown. As is clear from this figure, in the conventional method, when the noise becomes high (the S / N ratio becomes low), the recognition rate greatly decreases, whereas in the voice recognition device of the present invention, the noise becomes large. It can be seen that the recognition rate does not decrease even if it becomes worse and good recognition performance can be maintained.

As described above, according to this embodiment, it is possible to obtain an excellent speech recognition apparatus with a small reduction in recognition rate even in a large noise environment, and to use the speech recognition apparatus even in an environment with large noise. Its practical value is extremely large, such as enabling.

FIG. 3 is a block diagram of a speech recognition apparatus according to the second embodiment of the present invention. Reference numeral 301 is a microphone for inputting voice, 302 is a filter bank for dividing voice into a plurality of frequency bands having different center frequencies, and 303 is the center of each filter for each signal waveform of the voice signal frequency-analyzed by the filter. A delay circuit for delaying the filter output signal by a delay amount determined by the reciprocal of the frequency, 304 is a multiplier for multiplying the signal delayed by the delay circuit 303 and a filter output signal which is not delayed, 305 is An integrating circuit for integrating the output of the multiplier 304 for the integration time set for each filter,
306 is a square circuit for calculating the square of the filter output signal, 307 is an integrating circuit for integrating the output of the squarer 306 for the integration time set for each filter, and 308 is an integrating circuit of the output of the integrating circuit 306. A divider for normalizing the output of the multiplier 304 by using the power of each filter by dividing by 307, and an adder 309 for adding the output of the divider 308 connected to each filter output,
Reference numeral 10 is an identification unit for performing voice recognition using the voice feature amount output from 309.

The operation of the conventional speech recognition apparatus configured as above will be described below. In FIG.
The voice input by the microphone 301 for inputting voice is a plurality of filters 302 having different center frequencies.
Is divided into a plurality of audio frequency bands having different frequency bands. The center frequency of the bandpass filter 302 is set by logarithmically dividing the audio signal band into equal intervals as in the first embodiment, and is responsible for important components such as low-order formants of audio information. The low-frequency components are finely divided, and the high-frequency components are roughly analyzed to efficiently analyze the frequency of the voice. The audio signal analyzed by the filter bank 302 has a delay amount determined by the reciprocal of the center frequency of each filter of each signal waveform of the audio signal frequency-analyzed by the filter (for example, when the center frequency of the filter is 200 Hz). The delay amount is set to 5 ms) and is input to the delay circuit 303 for delaying the filter output signal and delayed, and then input to the multiplier 304 to obtain the product with the filter output signal that is not delayed, and then each filter It is input to the integrator 305 having the integration time set for each. The output of the integrator 305 is
The higher the correlation between the delayed signal (output of the delay circuit 303) and the non-delayed signal, the larger the output value, and conversely, the lower the correlation, the lower the output. The output of the integrator 305 is the output of the squaring circuit 306 (power of the filter output),
The output of the integrated signal by 7 is used for division by the divider 308. By this calculation, the center frequency component included in the filter output is obtained as a normalized amount using the power included in the filter. Divider 308-1-
The output of 308-N is added by the adder 309, and the identification unit 31
Input to 0. The identification unit 310 uses the input voice feature amount to perform known DTW (Dinamic Time Warping),
HMM (Hidden Markoff Model), NN (Neural Network)
The input voice is recognized by various voice pattern identification methods such as.

As described above, according to this embodiment, it is not necessary to obtain the autocorrelation of each signal waveform of the audio signal frequency-analyzed by a plurality of filters, and the correlation of the filter output can be reduced with a simpler configuration. It is possible to obtain an excellent speech recognition device with a small reduction in recognition rate even in a noisy environment, and it is possible to use the speech recognition device even in an environment with large noise. Is very large.

[0014]

As described above, according to the present invention,
It is possible to obtain a speech recognition apparatus with a small reduction in recognition rate even in a noisy environment, and to enable the speech recognition apparatus to be used in various environments, which is of great practical value.

[Brief description of drawings]

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

FIG. 2 is an explanatory diagram of performance of a voice recognition device using the first embodiment of the present invention.

FIG. 3 is a configuration diagram of a voice recognition device according to a second embodiment of the present invention.

FIG. 4 is a block diagram of a conventional voice recognition device.

[Explanation of symbols]

 101 Input Microphone 102 Filter Bank 103 Auto Correlator 104 Power Detection Circuit 105 Divider 106 Adder 107 Discriminator 301 Microphone 302 Filter Bank 303 Delay Circuit 304 Multiplier 305 Integrator 306 Square Circuit 307 Integrator 308 Adder 309 Divider 310 Identification unit

Claims (4)

[Claims] 1. A means for dividing a voice into a plurality of frequency bands by filters having different center frequencies, and a means for obtaining the magnitude of correlation in the time direction contained in each filter output waveform obtained by the division. A means for detecting the power component of each signal waveform obtained by the means for dividing into the plurality of frequency bands; a means for normalizing the magnitude of the correlation between the divided signal waveforms using the power component; A voice recognition device comprising: means for adding the normalized signals of the plurality of frequency bands; and an identification section for identifying an input voice using the added signal components. 2. A means for obtaining the magnitude of correlation contained in each signal waveform divided by a filter, and a means for detecting a power component of each signal waveform of the divided audio signal,
The speech recognition apparatus according to claim 1, wherein the speech waveform is obtained by using autocorrelation of a signal waveform divided by the means for dividing into a plurality of frequency bands. 3. A means for determining the magnitude of correlation contained in each signal waveform divided by a filter, is a voice signal divided by the plurality of filters and a delay amount determined by a center frequency of each filter. 2. The voice recognition apparatus according to claim 1, wherein the voice signal is obtained by using a means for multiplying a signal obtained by delaying a voice signal divided by the plurality of filters and a means for integrating an output of the multiplying means. 4. The voice according to claim 1, wherein the center frequency of the filter for dividing the voice into a plurality of frequency bands is set by dividing the voice signal band logarithmically at equal intervals. Recognition device.