JPH04367899A - Agc control system of voice recognition device - Google Patents

Abstract

PURPOSE:To make the amplitude and the variation point of suppression most suitable by placing an AGC circuit after a feature pickup section and by AGC controlling based upon the output (power spectrum) signals. CONSTITUTION:An AGC control section 24 is located in the next stage of a power spectrum generating section 23. Means 25, 26, and 27, which control gain constants against each power spectrum by the voice power signals which are obtained by low pass filtering the voice power in the total of power spectrum then integrated, a comparing means, which judges decrease and increase of the voice power signals, and a means, which switches time constants of each of the low pass filters for each case, are provided. When the voice signal is small, a certain gain is given to amplify the signal and when the signal is large, it is suppressed.

Description

[Detailed description of the invention]

0001

TECHNICAL FIELD The present invention relates to an AGC control method for a speech recognition device, and more particularly to an AGC control method for a speech recognition device that frequency-analyzes input speech and generates a speech pattern.

0002

[Prior Art] AGC (automatic gain control) of a speech recognition device is used to normalize the loudness of a speaker's voice. The system attempts to normalize the signal level by raising the signal when a person's voice is loud or when a large signal is being input, and suppressing it when a person has a loud voice or a large signal is being input. Figure 10 shows the AGC of the audio signal.
The basic configuration of (automatic gain control) is shown. The input signal is L. P. A low-frequency component (power signal) is extracted by an F (low-pass filter) 2, and the signal is applied to an variable gain control section (AGC) 1 using a variable current element, which changes the gain of the audio signal and controls the audio signal. The output of the signal is controlled. In other words, when the signal level is low, the gain is increased to raise it, and when the signal level is high, it is suppressed. The input/output characteristics of the signal change from point A as shown in FIG. In other words, signal amplification is performed at a constant gain for small signals up to point A, and the gain is reduced to 1 for signals above point A.
Suppress as below. It is desirable that the input signal be at a level appropriate for point A, but the level of human voice fluctuates widely, and the way the microphone is attached also varies greatly. Therefore, if you set the center value of the input signal level to be approximately at point B (A and B are approximately 2
(about 0 to 30 dB) improves the audio level balance.

FIG. 12 is a diagram for explaining an example of a conventional speech recognition device. In the figure, 11 is a microphone, 12 is a
is a microphone amplifier, 13 is an AGC (automatic gain control circuit),
14 is an A/D converter, 15 is a feature extraction unit (power spectrum generation unit), 16 is a pattern generation unit, 17 is a recognition unit,
18 is a recognition dictionary, 19 is a result output unit, and the feature extraction unit 15 is constructed using B.I. P. F (band pass filter) 15a, rectifier 15b, L. P. It is composed of an F (low pass filter) 15c and the like.

However, in the AGC in the conventional speech recognition device, the AGC is performed immediately after the microphone amplifier 12.
13, after that, the feature extraction section 15 extracts the feature amount, the pattern generation section 16 generates an input pattern using the feature amount, and the pre-registered recognition dictionary 18
It compared the patterns and output the recognition results. When configured in this way, the AGC is generally configured with analog elements, and it is difficult to control the changing point of amplification and suppression of the AGC. Also, more importantly, AGC
By doing this, distortion occurs in the audio waveform itself,
In some cases, the feature extraction itself was affected.

[0005] In the conventional AGC described above, a small signal is amplified with a constant gain, and a large signal is controlled by decreasing the gain. However, if some noise is added to the audio, it will be amplified as well, which will increase the unnecessary noise.As a result, the S/N of the audio signal after passing through the AGC will deteriorate, and the feature value extraction had a negative impact on.

[0006] Furthermore, as described above, the conventional AGC is located after the microphone amplifier, performs AGC control on analog audio signals, and performs feature extraction from the signals to obtain feature quantities. Therefore, changing the characteristics requires changing the analog elements, which is practically difficult.

Furthermore, in the method of performing AGC control after power spectrum generation, when the power value is small, a constant gain is applied and amplified, and when the power value is large, the gain is changed to suppress it, so that the power value remains constant. The purpose is to However, in this method, only the switching point between amplification and suppression is a parameter for AGC control, so the setting of the switching point is related to both amplification and suppression. for example,
If the amplification range is set small, the dynamic range of the suppressor cannot be maintained. In other words, with this AGC control method, only the range of the amplification section cannot be changed.

[0008]

[Purpose] The present invention has been made in view of the above-mentioned circumstances, and in particular, by placing an AGC circuit at a stage subsequent to the feature extraction section,
AGC control is performed based on the output (power spectrum) signal, and the change point of amplification and suppression is controlled to be optimal, or the gain setting when the AGC signal is small is
To perform AGC control effectively on audio signals without constantly increasing the noise level in the small signal region, or to control the power value for large power values by not operating the AGC function for large power values. This was done with the purpose of making it possible to preserve the

[0009] In order to achieve the above object, the present invention provides (
1) A preprocessing unit that amplifies and corrects the audio signal input from the microphone, a power spectrum generation unit that frequency-analyzes the obtained audio signal and converts it into a power spectrum, and a binarization pattern from these power spectra. and a speech recognition device that compares the word pattern of the binarized audio signal with a dictionary pattern registered in advance and takes the most similar word pattern as a recognition result. An AGC control unit is located at the next stage of the power spectrum generation unit, and the audio power signal obtained from the summation of the power spectra is obtained by passing it through a low-pass filter, and the audio power signal that is further integrated is used to generate each power spectrum. means for controlling the gain of the power spectrum by controlling a gain constant for the voice power signal; and comparison means for determining whether the further integrated audio power signal decreases or increases;
It has a means for switching the time constant of the low-pass filter in each case, and is characterized in that when the audio signal is small, it is amplified by giving a constant gain, and when it is large, it is suppressed. Yes, and furthermore, (2) in (1) above,
The switching point between amplification and suppression can be changed, and (3) in (2) above, the switching point between amplification and suppression can be changed by setting an external register; (4) The switching point between the amplifying section and the suppressing section is controlled by the average power of past sounds, or (5) the audio signal input from the microphone is amplified and corrected. A preprocessing section, a power spectrum generation section that frequency-analyzes the obtained audio signal and converts it into a power spectrum, a binary pattern generation section that generates a binarized pattern from these power spectra, and a binary pattern generation section that generates a binarized pattern from these power spectra. A speech recognition device that compares a word pattern of a voice signal with a pre-registered dictionary pattern and selects the most similar word pattern as a recognition result, the speech recognition device comprising:
An AGC control unit is located at the next stage of the power spectrum generation unit, and the audio power signal obtained from the sum of the power spectra is obtained through a low-pass filter,
means for controlling the gain of the power spectrum by controlling a gain constant for each power spectrum using the further integrated audio power signal; a comparing means for determining whether the further integrated audio power signal decreases or increases; It has a function of switching the time constant of the low-pass filter when the audio signal is small, and has a function of amplifying it by giving a constant gain when the audio signal is small, and suppressing it when it is large, and has a function of amplifying and suppressing the audio signal. (6) In the AGC control method of a speech recognition device that can change the switching point of the voice recognition device, it has a means for detecting a steady noise level, and suppresses the sound by keeping the gain constant in the voice power range of the steady noise level, or (6) A pre-processing section that amplifies and corrects the audio signal input from the microphone, a power spectrum generation section that analyzes the frequency of the obtained audio signal and converts it into a power spectrum, and generates a binarized pattern from these power spectra. a binary pattern generation unit;
A speech recognition device that compares the word pattern of the binarized speech signal with a pre-registered dictionary pattern and selects the most similar word pattern as a recognition result, An AGC control unit is located, and controls the gain constant for each power spectrum using the further integrated audio power signal obtained by passing the low-pass filter to the audio power signal obtained from the summation of the power spectra. means for controlling the gain of the power spectrum; comparison means for determining a decrease or increase in the further integrated audio power signal; and means for switching the time constant of the low-pass filter in each case; It has the function of amplifying by giving a constant gain when is small, and suppressing it when it is large, and when the gain of the suppressor is less than 1, setting the gain to 1 increases the power value. The feature is that the AGC control function is not activated for
Furthermore, (7) a suppressor is provided when the power value exceeds a certain value. Hereinafter, the present invention will be explained based on examples.

FIG. 1 shows the AG of the speech recognition device according to the present invention.
This is a configuration diagram for explaining one embodiment of the C control method, in which 21 is a microphone, 22 is a microphone amplifier, 23 is a feature extraction section, 24 is an AGC control section, 25 is a voice pattern generation section, and 26 is a 27 is a gain setting section, 28 is a pattern generation section, 29 is a recognition section, 30 is a recognition dictionary, and 31 is a result output section. In FIG. 1, an audio signal input from a microphone 21 is amplified by a microphone amplifier 22, and an n-channel B.I. P. Frequency analysis is performed by F (band pass filter) 23a, and this is applied to rectifier 2.
3b and then L.3b. P. It is converted into a power spectrum through an F (low pass filter) 23c. The power spectrum obtained here is optimally level-converted in the AGC control unit 24, an input pattern is generated in the pattern generation unit 28, and the pattern is compared with a recognition dictionary 30 registered in advance in the recognition unit 29, and the result is output. Ru. A
In contrast to the GC control unit 24, the audio power generation unit 25 generates the L. P. Based on the power spectrum from the F23c, the audio power generation unit 25 calculates the audio power by summing the power spectra of n channels, and the cumulative average calculation unit 26
The past average powers are accumulated. The register in the gain setting unit 27 is updated based on this average power, and the AGC control unit 24 is updated based on the value set in this register.
The gain of is determined.

FIG. 2 is a detailed diagram of the AGC control section 24. As shown in FIG. and in case of increase, L. P. Through F1 (attack time constant filter) 42, and in case of decrease, L. P. F2 (decay time constant filter) 43
The gain for the input signal is determined by the gain calculation unit 44 and the gain control unit 45 based on the power of the voice obtained here. The gain control unit 45 changes the change point between amplification and suppression using the gain setting register, and determines the optimum gain.

For example, in AGC control with the input and output relationship shown in FIG. 11, if the audio power of the input signal as shown in FIG. 3(a) is obtained, the change in gain as shown in FIG. 3(b) and Become. In other words, when the audio power is small, a constant gain is obtained, and when the signal becomes large, the gain decreases, changing from amplification to suppression. Further, the gain setting register sets the changes in I and II in FIG. 3(c). In other words,
When the center point of the audio level is near A, the gain is set to I, and when the average audio power becomes small, it is moved to II using the gain setting register. By doing so, it becomes possible to generate the AGC output signal in the same way.

FIG. 4 is a diagram for explaining the operating principle of the invention set forth in claim 5.
Two gain change points, a and A, are provided. FIG. 5 is a diagram for comparing and explaining the present invention and the prior art. In the prior art, the change point was set only at point A as described above. In other words, small signals were amplified with a constant gain. In this case, if the noise shown in the shaded area (for example, microphone line noise or ambient steady noise) is mixed into the audio power signal shown in FIG. 5(a), a certain gain is applied. , as shown in FIG. 5(b), increases the noise level. In contrast, in the present invention, as shown in FIG. 4, a change point is also provided at point a, and the gain is kept constant for signals below point a. By doing this, as shown in FIG. 5(c), the noise level is not amplified, but only the audio signal is amplified.

FIG. 6 is a block diagram of a main part of an electric circuit for explaining an example of an AGC control circuit used to implement the present invention as described above.
52 and 53 are low pass filters (LPF), 54 is a noise level detection section, 55 is a level comparison control section, and 56 is a variable gain control section. The input power spectra are compared by the signal comparator 51 to see if the signal increases or decreases, and if the signal increases, the L. P. F1 (attack time constant filter)
52, in case of decrease L. P. The gain for the input signal is determined by the variable gain control unit 56 based on the power of the audio obtained through the F2 (decay time constant filter) 53. In the variable gain control section 56, the stationary noise level in the input signal is detected at the change point of amplification and suppression by the noise level detection section 54, and thereby the level comparison control section 55 is controlled to adjust the gain of the variable gain control section 56. In the audio power range of steady noise level, the gain is kept constant and suppressed for control. This enables more accurate AGC control of the audio signal without using the AGC control unit to remove unnecessary noise mixed into the audio signal.

After the above-mentioned power spectrum generation, AGC
The control method is such that when the power value is small, a constant gain is applied to amplify it, and when it is large, the gain is changed so as to suppress it, so that the power value remains constant. However, in this method, only the switching point between amplification and suppression is a parameter for AGC control, so the setting of the switching point is related to both amplification and suppression. For example, if the amplification range is set small, the dynamic range of the suppressor cannot be maintained. The invention set forth in claim 6 is intended to solve the above-mentioned problem, that is, the above-described AGC control method cannot change only the range of the amplifier section.

FIG. 7(a) shows the gain characteristics in the above-mentioned AGC control method, and FIG. 7(b) shows the relationship with the gain coefficient. FIG. 8(a) shows the gain characteristic in the control method according to the present invention, and FIG. 8(b) shows the relationship with the gain coefficient. As is clear from FIG. 8, according to the present invention, compared to the gain characteristics shown in FIG. It becomes possible to save power values for powers. FIG. 9 shows a characteristic for explaining the invention as claimed in claim 6, in which a suppressing effect is added to the power value above a certain level, thereby making it possible to fix the maximum power.

The main purpose of using AGC control in speech recognition is to raise consonants with low power and to maintain a constant speech level. Therefore, it is necessary to amplify the parts with low power, but it is unnecessary to suppress the parts with high power. Furthermore, the dynamic range will be reduced. The invention set forth in claim 6 makes it possible to eliminate such a suppressing portion. Furthermore, according to the seventh aspect of the invention, a section in which the gain coefficient is 1 is provided between the switching points of the amplifying section and the suppressing section, so that an excessive dynamic range can be suppressed.

[0018]

[Effects] According to the invention as set forth in claims 1 to 4, AGC control can be performed stably, and control can always be optimally performed in response to fluctuations in audio level. According to the invention set forth in claim 5, unnecessary noise mixed into the audio signal is
Since the GC control prevents the signal from being pulled up, more accurate AGC control of the audio signal becomes possible. Claim 6
According to the invention described in , when the power value is a certain value or more, that is, the gain coefficient is 1 or less, the gain coefficient is set to 1.
Since it is fixed to , the power value can be preserved for large powers. According to the seventh aspect of the invention, since a suppressing effect is added to the power value above a certain value, it is possible to fix the maximum power.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing the overall configuration when an AGC control circuit according to the present invention is applied to a speech recognition device.

FIG. 2 is a block diagram for explaining details of the AGC control circuit according to claims 1 to 4.

3 is a waveform diagram and a characteristic diagram for explaining the operation of the AGC control circuit shown in FIG. 2. FIG.

FIG. 4 is a characteristic diagram for explaining the operation of the AGC control circuit according to claim 5;

5 is a waveform diagram for explaining the operation of the AGC control circuit according to claim 5. FIG.

FIG. 6 is a block diagram for explaining details of the AGC control circuit according to claim 5.

FIG. 7 is a characteristic diagram for explaining the characteristics of a conventional AGC control circuit.

FIG. 8 is a characteristic diagram for explaining the operation of the invention recited in claim 6;

9 is a characteristic diagram for explaining the operation of the invention set forth in claim 7. FIG.

FIG. 10 is a block diagram for explaining an example of a conventional variable gain control circuit.

FIG. 11 is a characteristic diagram of the circuit of FIG. 10.

FIG. 12 is an overall configuration diagram for explaining an example of a conventional speech recognition device.

[Explanation of symbols]

21...Microphone, 22...Mic amplifier, 23...Feature amount extraction unit, 24...AGC control unit, 25...Audio pattern generation unit, 26...Cumulative average calculation unit, 27...Gain setting unit,
28...Pattern generation unit, 29...Recognition unit, 30...Recognition dictionary, 31...Result output unit.

Claims (7)

[Claims] [Claim 1] A preprocessing unit that amplifies and corrects an audio signal input from a microphone, a power spectrum generation unit that frequency-analyzes the obtained audio signal and converts it into a power spectrum, and a A binary pattern generation unit that generates a digitized pattern, and a speech recognition device that compares the word pattern of this binarized audio signal with pre-registered dictionary patterns and selects the most similar one as a recognition result. An AGC control unit is located at the next stage of the power spectrum generation unit, and the audio power signal obtained by passing the audio power signal obtained from the summation of the power spectra through a low-pass filter and further integrating the audio power signal, means for controlling the gain of the power spectrum by controlling a gain constant for each power spectrum; a comparison means for determining whether the audio power signal decreases or increases; and means for switching the time constant of the low-pass filter in each case. An AGC control method for a speech recognition device, characterized in that when a speech signal is small, the speech signal is amplified by giving a constant gain, and when the speech signal is large, it is suppressed. 2. The AGC control method for a speech recognition device according to claim 1, wherein a switching point between said amplification and suppression can be changed. 3. The AGC control method for a speech recognition device according to claim 2, wherein the switching point between amplification and suppression is changed by setting an external register. 4. The AGC control method for a speech recognition device according to claim 2, wherein the switching point between the amplifying section and the suppressing section is controlled based on the average power of past speech. 5. A preprocessing unit that amplifies and corrects an audio signal input from a microphone, a power spectrum generation unit that frequency-analyzes the obtained audio signal and converts it into a power spectrum, and a A binary pattern generation unit that generates a digitized pattern, and a speech recognition device that compares the word pattern of this binarized audio signal with pre-registered dictionary patterns and selects the most similar one as a recognition result. An AGC control unit is located at the next stage of the power spectrum generation unit, and the AGC control unit generates an audio power signal obtained by passing the audio power signal obtained from the summation of the power spectra through a low-pass filter and further integrating the audio power signal. means for controlling the gain of the power spectrum by controlling a gain constant for each power spectrum; a comparison means for determining whether the audio power signal is decreasing or increasing; and switching the time constant of the low-pass filter in each case. AGC of a speech recognition device that has a function of applying a constant gain and amplifying the speech signal when it is small and suppressing it when it is large, and that can change the switching point between amplification and suppression.
A voice recognition device characterized in that the control method includes means for detecting a steady noise level, and suppresses the steady noise level by keeping the gain constant in the voice power range of the steady noise level.
GC control method. [Claim 6] A preprocessing unit that amplifies and corrects an audio signal input from a microphone, a power spectrum generation unit that frequency-analyzes the obtained audio signal and converts it into a power spectrum, and generates binary values from these power spectra. a binary pattern generation unit that generates a binary pattern, and a speech recognition device that compares the word pattern of this binary audio signal with a pre-registered dictionary pattern and selects the most similar word pattern as a recognition result. An AGC control section is located at the next stage of the power spectrum generation section, and the AGC control section is located at the next stage of the power spectrum generation section, and the audio power signal obtained from the summation of the power spectra is obtained by passing it through a low-pass filter and further integrated. , means for controlling the gain of the power spectrum by controlling a gain constant for each power spectrum, a comparison means for determining whether the audio power signal decreases or increases, and means for switching the time constant of the low-pass filter in each case. It has the function of amplifying the audio signal with a constant gain when it is small, and suppressing it when it is large, and when the gain of the suppression section is 1 or less, the gain is set to 1. An AGC control method for a speech recognition device characterized in that the AGC control function is not activated for particularly large power values. 7. The AGC control method for a speech recognition device according to claim 6, further comprising a suppressor provided when the power value exceeds a certain value.