JPH0416900A - Speech recognition device - Google Patents

Abstract

PURPOSE:To eliminate the influence of a noise upon speech recognizing ability by inputting the outputs of two microphones which gather a voice and a noise mainly to two filter banks which are equipped with band-pass filters by plural channels, and subtracting the outputs of corresponding channels by respective subtracting circuits. CONSTITUTION:The voice input and noise component which are gathered by the two microphones 11 and 12 are outputted to the filter banks 15 and 16 through analog processing circuits 13 and 14 respectively. Those filter banks 15 and 16 branch the voice input and noise by plural frequency bands and subtracting circuits 17-1 to 17-n perform subtraction as to outputs of corresponding frequency bands. Thus, the signals which are gathered by the two microphones and outputted to the filter banks are processed by subtraction, channel by channel, to cancel the noise components which contribute to the smoothing outputs of the filters. Consequently, even when the device is used in noisy environment, deterioration in speech recognizing performance due to the noise is prevented.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition device equipped with a filter bank that branches speech collected by a microphone into a plurality of frequency bands.

[Summary of the Invention] The present invention inputs the output from two microphones that mainly collect voices and noise into two voice analysis filter banks each equipped with a plurality of channels of band-pass filters, and The feature is that the output of the corresponding channel of the bank is subtracted by each subtraction circuit.

[Prior Art] For example, the one shown in FIG. 2 is known as a speech recognition device that performs speech analysis using a plurality of channels of band-pass filters.

In FIG. 2, an analog processing circuit 2 performs processing such as amplification and filtering on an audio signal collected by a microphone 1, and the results are input to a filter bank 3. In this filter bank 3, frequency analysis is performed using a plurality of band pass filters 31 each having a different center frequency Fl, . . . , find the average energy level of each filter.

The output of this filter bank 3 is passed through a multiplexer 4, and taken in as a voice feature quantity by an A/D converter 5.
It is stored in the pattern memory 6.

A series of operations in the A/D converter 5 and pattern memory 6 are as follows:
A microcomputer 7 consisting of a CPU 71, ROM 72, and RAM 73 supervises and controls the system.

FIG. 3 shows a waveform diagram of the frequency band for one channel of the filter bank 3 in FIG.
(e) is the output of the smoothing circuit 33.

As can be seen from these figures, the signal input to the A/D converter 5 is smoothed by the filter bank 3 as an analog circuit part, and its output is a signal with a frequency component lower than that of the original signal. It has become. -This is because the energy of voice does not change rapidly.

[Problems to be Solved by the Invention] When the above-mentioned speech recognition device is used in a noisy environment, the signal input from the microphone 1 becomes a sound with noise superimposed on it, and this is caused by the difference between each of the smoothed outputs of the filter bank. This affects the spectral pattern and causes deterioration of speech recognition performance.

Another reason why speech recognition performance deteriorates when noise is mixed into input speech is that the energy level of the filter becomes higher than when there is no noise.

At this time, if a certain threshold value is set for the energy level and the start and end of the voice is determined in order to detect the voice section, the accuracy of the section detection decreases.

[Object of the Invention] An object of the present invention is to provide a speech recognition device that prevents noise from deteriorating speech recognition performance even when used in a noisy environment.

[Means for Solving the Problem] In order to achieve such an object, the present invention provides first and second microphones that mainly collect voice and noise, respectively, and a plurality of outputs of each of these microphones. The present invention is characterized by comprising first and second filter banks branching into frequency bands, and subtracting means having a plurality of subtraction circuits for subtracting outputs of frequency bands corresponding to each of the filter banks.

[Function] In the present invention, two types of sound are mainly used to collect voices and noises.
The outputs of the microphones are input to two filter banks each equipped with multiple channels of band-pass filters, and the outputs of the corresponding channels of each filter bank are subtracted by the respective subtraction circuits. There is no impact on the

[Example] Hereinafter, one embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of a speech recognition device according to the present invention.

As shown in the figure, in the present invention, two microphones 11, 12, analog processing circuits 13, 1, 4, and
filter banks 15 and 16, one 16 is used for noise detection and the other 15 is used for voice input.

Also, a plurality of subtraction circuits 17- for subtracting the outputs of the corresponding frequency bands of the filter banks 15 and 16.
1 to 17-n are provided, and their outputs are input to the multiplexer 4. Furthermore, a level setting section 18 for setting the output level of one filter bank 16 is provided, and this level setting section 18 is controlled by the microcomputer 7.

Of the two microphones 11 and 12, the microphone 11 for voice input is placed in a place where the voice of the speaker is likely to be input, for example, at a position corresponding to the mouth of the forehead of the first speaker. In order to detect ambient noise, it is installed at a place where only noise is input, for example, near the back of the head of one speaker. Since the microphones 11 and 12 are installed relatively close to each other, the characteristics of the input noise (spectral distribution, etc.) are similar.

In the present invention, the audio input and noise components collected by the two microphones 11 and 12 are respectively input to the analog processing circuit 13.
14 to filter banks 15 and 16. In this filter bank 15 and 16, microphone 1
A circuit 1 divides the audio input and noise from 1 and 12 into multiple frequency bands, and subtracts the outputs of the corresponding frequency bands.
Subtract from 7-1 to 17-n.

In this way, by collecting sound with two microphones and subtracting the signal output to the filter bank for each channel, it is possible to cancel out the contribution of the noise component to the smoothed output of the filter. At this time, since the microphones are installed at different positions, the level of the signal collected by the microphones differs depending on the direction of arrival of the noise, so the level setting unit 18 adjusts the levels. That is, the level setting unit 18 sets the level of the smoothed output so that the influence of noise on the smoothed output of the filter is reduced and the output of the subtraction circuit is reduced while no voice is input. This operation is hereinafter referred to as calibration.

Specifically, the subtraction results for each channel of the filter bank 15 and 16 are sequentially switched by the multiplexer 4, and the A/
The D-converter 5 digitizes the output, and the microcomputer 7 monitors the output. When performing a calibration operation, the level of the level setting section 18 is set so that the subtraction result of the subtraction circuit 17 is below a predetermined level. do.

FIG. 4 shows the block configuration of the level setting section 18, in which an AGC amplifier 19, a D/A converter 20, and a latch 21 are provided for each channel, and setting data sent from the microcomputer 7 is The data is taken into the latch 21 corresponding to the designated channel under the control of the microcomputer 7. The setting data taken into the latch 21 is converted into an analog voltage by the D/A converter 20, and this is used as the gain control voltage of the AGC amplifier 19.

Since the smoothed output from the filter bank 3 is input to the AGC amplifier 19, it is amplified or attenuated to a predetermined magnitude and output.

FIG. 5 is a flowchart showing the overall operation of the system of the present invention. First, the above-mentioned calibration operation is performed (step 22). Next, it is determined whether or not to perform speech recognition (step 23). is executed (step 24), and if control is not transferred to another device, the system operation is ended (steps 25 and 26).

FIG. 6 is a flowchart showing the specific operation of calibration. As mentioned above, the channels are switched by the multiplexer 4 (step 40), and the latch 21 of the level setting section 18 is switched based on the judgment result of the microcomputer 7. Set a setting level (Step 41), read the subtraction result of the subtraction circuit 17 at the setting level into the microcomputer 7 (Step 42), and determine whether the calculation result is below the threshold (Step 43). Repeat steps 41 and 42 until the threshold is below.

Such operations are performed for all channels (step 44).

According to the embodiment of the present invention described above, it is possible to cancel out the contribution of noise to the smoothed output.

Therefore, it is possible to reduce the deformation of the spectral pattern and the deterioration of speech segment detection accuracy due to an increase in the noise level, and as a result, it is possible to prevent the deterioration of the speech recognition performance.

Furthermore, since the smoothed output signal, which has a lower frequency than the input original signal, is subjected to calculation processing, the influence of noise can be reduced without performing high-speed calculation or complicated processing.

Furthermore, the system of the present invention can be easily realized by adding a circuit to a conventional speech recognition system.

[Effects of the invention] As stated above. According to the present invention, even when used in a noisy environment, noise does not degrade speech recognition performance.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an embodiment of a speech recognition device according to the present invention, FIG. 2 is a block diagram showing the configuration of a conventional speech recognition device, and FIG. 3 is a block diagram showing the configuration of a conventional speech recognition device.
FIG. 4 is a block diagram of a specific configuration of the level setting section of FIG. 1, and FIGS. 5 and 6 are flowcharts for explaining the present invention in detail. 11.12・・・・・・Microphone, 13.1
4... Analog processing circuit, 15.16.
......Filter bank, 17-1 to 17-n
・・・・・・・・・Subtraction circuit, 18・・・・・・・・・
Level setting section. Patent Applicant Tararion Co., Ltd. Agent Patent Attorney Nagai 1) Takeshi Part 3 Figure 3 (C) 1-i〉-1 Figure 4 Figure 5

Claims (2)

[Claims] (1) A first microphone that mainly collects sound, a second microphone placed near the first microphone that mainly collects noise, and a plurality of microphones that collect the output of the first microphone. A first filter bank that branches into frequency bands, a second filter bank that branches the output of the second microphone into a plurality of frequency bands, and corresponds to each of the first filter bank and the second filter bank. a subtractor having a plurality of subtraction circuits for subtracting outputs of frequency bands. (2) Furthermore, detect the average energy level of the noise collected from the first microphone in the absence of the sound and the average energy level of the noise collected from the second microphone, and detect the average energy level of the noise collected from the second microphone. 2. The speech recognition apparatus according to claim 1, further comprising a determining means for determining the magnitude of the subtracting means, and a setting means for setting the weight of the subtracting means in accordance with the determination result of the determining means.