JPH01200294A - Sound recognizing device - Google Patents

Abstract

PURPOSE:To prevent erroneous recognition due to an environmental noise by sound- gathering the environmental noise with using a main microphone and an auxiliary microphone and after that, obtaining sound data for recognition processing with using a noise correcting coefficient, which is calculated by frequency-dividing the environmental noise. CONSTITUTION:The environmental noise is sound-gathered by a main microphone 1 and an auxiliary microphone 11 and divided in each frequency band by BPFs 3 and 13. Then, the correcting coefficient for noise reduction is calculated by a correcting coefficient calculating means 24 and stored in a memory 25. Next, when a sound signal is sound-gathered by the main microphone 1, the sound signal to include the environmental noise is frequency-divided and supplied to a sound data calculating means 26. Then, the environmental noise at such a time is sound-gathered by the auxiliary microphone 11, frequency divided and supplied to the sound data calculating means 26. The correcting coefficient just before a word is generated is read from the memory 25 and the sound data as data for recognition processing are calculated by the sound data calculating means 26. Thus, the erroneous recognition can be prevented from being executed by the environmental noise in a sound recognizing part 27.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a speech recognition device that uses, in addition to a live microphone, an auxiliary microphone that mainly collects environmental noise.

[Summary of the invention]

The present invention provides a voice recognition device that performs voice recognition by comparing acoustic parameters obtained by frequency-dividing and normalizing an audio signal collected by a main microphone with a preset standard pattern. A correction coefficient for noise removal is calculated by comparing the frequency-divided data of the environmental noise collected by the auxiliary microphone with the frequency-divided data of the environmental noise collected by the auxiliary microphone. By calculating the noise-removed data from the audio signal collected by the auxiliary microphone and the auxiliary microphone, and performing acoustic recognition using this audio data, it is possible to eliminate misrecognition due to environmental noise. It is designed to prevent this.

[Conventional technology]

Conventionally, various types of speech recognition devices have been proposed. For example, a directional microphone is placed near the speaker's mouth, an omnidirectional microphone is placed away from the speaker's mouth and near the speaker, and both The outputs of the microphones are each amplified and the difference is determined by a differential intensifier, and the difference output is inputted to a voice identification device, where the specific word type is pre-stored. There is a method of identification by comparing the
Publication No. 2604).

In addition, there is a method of extracting the frequency spectrum of the human voice itself by calculating the time average value of the frequency spectrum of human noise before inputting the voice and subtracting this average value from the frequency spectrum of the voice input signal. (Japanese Unexamined Patent Publication No. 55-33126.) [Problem to be Solved by the Invention] However, in the case of the conventional device as described in Unexamined Japanese Patent Publication No. 51-62604, noise entering the directional microphone It is difficult to detect the phase difference between the phase of the noise and the noise coming from the omnidirectional microphone, and therefore it is difficult to stably cancel out the noise, resulting in erroneous recognition due to the noise.

Furthermore, in the case of a conventional device such as that described in Japanese Patent Application Laid-open No. 55-33126, although it is effective against deterministic noise,
It is not possible to deal with intermittent noises or noises that change easily, such as people's conversations or music, and tj4
The present invention has been made in view of this problem, and provides an acoustic recognition device that can prevent erroneous recognition due to environmental noise.

[Means to solve the problem]

This invention performs frequency division (3) and normalization (27a) of an audio signal collected by a live microphone (1).

A speech recognition device that performs speech recognition by comparing (27b) the acoustic parameters obtained with a preset standard pattern (27c) includes an auxiliary microphone (11) that collects at least environmental noise, and a main microphone (1). ) and the data obtained by frequency-dividing the environmental noise collected by the auxiliary microphone (11) to calculate a correction coefficient for noise removal. , 25) and an audio signal including environmental noise collected by the main microphone (1) and at least the environmental noise collected by the auxiliary microphone (11) and a correction coefficient to calculate audio data from which noise has been removed. and a data calculation means (26),
It is configured to perform voice recognition (27) using voice data.

[Effect]

First, the main microphone +11 and the auxiliary microphone (11) collect environmental noise in a state where no words are being generated, and the bandpass filter bank (3, 13) divides the sound into each frequency band, and uses each data to perform compensation. Stop coefficient calculation means (
24, 25), the noise removal correction coefficient An is calculated and stored for each frequency band. This correction coefficient An is updated and corrected so as to always maintain the optimum value. Next, when words are generated and the audio signal is collected by the main microphone (1), this audio signal containing environmental noise is frequency-divided as described above and supplied to the audio data calculator [fi (26)]. Auxiliary microphone (11) at least eliminates environmental noise at this time.
Collects sound, divides it into frequencies, and calculates audio data (26)
The correction coefficient immediately before the word was generated is read out from the memory (25), and the speech data calculation means (26)
), voice data as recognition processing data is calculated, and this voice data is compared with a standard pattern to perform voice recognition. As a result, environmental noise is substantially canceled out on the frequency axis, so that erroneous recognition due to environmental noise can be reliably prevented without being affected by phase shifts unlike the conventional method performed on the time axis.

〔Example〕

Hereinafter, one embodiment of the present invention will be described in detail based on the accompanying drawings.

The figure shows the circuit configuration of this embodiment, and in the figure,
[1) is a main microphone mainly for collecting audio signals, (2) is an amplifier that amplifies the audio signal from the main microphone (1), and (3) is a band-pass filter bank, which has, for example, 16 channels. Bandpass filter (
131) to (13ts), the total frequency band of which is, for example, from 200 Hz to 6 kHz, and is allocated to each bandpass filter so that the frequency bands are equally spaced. (4) represents each bandpass filter (31) to (
This is a multiplexer that switches and extracts the output of 3ts) in a time-division manner. (5) is a low pass filter, (6) is a sample hold circuit, and (7) is an A/D converter.

Further, (11) is an auxiliary microphone that mainly collects environmental noise, (12) is an amplifier that amplifies the environmental noise from the auxiliary microphone (11), and (13) is a band pass filter bank, which is the same as described above. Bandpass filter (3
) Similar 16 channel band pass filter (131)
~(131ε), have similar frequency bands, and are distributed to each bandpass filter at equal intervals. (14) is a multiplexer, (15) is a low pass filter, (16) is a sample hold circuit, (17)
is an A/D converter.

(20) is a switch circuit, which is an interlocking switch (
20a) and (20b), and has a switch (20a) and (20b).
The switch (20b) is supplied with digital data from the A/D converter (17), and the switch (20b) is supplied with digital data from the A/D* converter (17). (21
) is a switch circuit in which an interlocking switch (21a
).

(21b), the switch (21a) is supplied with digital data from the A/D converter (7), and the switch (21b) is supplied with digital data from the A/D converter (17). Ru.

These switch circuits (20) and (21) are controlled by the output signal from the level detection circuit (22). That is, the level detection circuit (22) has a predetermined threshold level i'h, and when the level of the signal from the main microphone (11 colors) is greater than this threshold level '1'h, it generates 0N(i) on its output side. Then, this ON signal turns on the switch (21a) of the switch circuit (21).
, (21b) are closed and become ON, and the ON signal generated by the level detection circuit (22) is inverted by the inverter (23) and becomes an OFF signal, and this OF
The switch (20a) of the switch circuit (20) is activated by the F signal.
) and (20b) are opened and become OFF state. Furthermore, when the level of the audio signal from the main microphone (1) is lower than the threshold level T)I, an OFF signal is generated on its output side, and this OFF signal causes the switches (21a) and (21b) of the switch circuit (21) to be activated.
) is released and becomes OFF, and the OFF signal generated from the level detection circuit (22) is sent to the inverter (23).
) is inverted and becomes an ON signal, and this OFF signal causes the switch (20a) of the switch circuit (20),
(20b) is closed and becomes ON state. In other words, when the level of the audio signal from the main microphone (1) is higher than the threshold level 'I'h of the level detection circuit (22), the switch (21a) of the switch circuit (21) is activated.
, (21b) is in ON state, switch circuit (20)
switches (20a) and (20b) are OFF
When the level of the audio signal from the main microphone +11 is lower than the threshold level Tk of the level detection circuit (22), the switches (21a) and (21b) of the switch circuit (21) are in the F state. A switch (20a) of a switch circuit (20).

(20b) is turned on.

The threshold level Th in this level detection circuit (22) is set to a level that is lower than the level of the voice signal of normal speaking and higher than the environmental noise.

switch (20a) of the switch circuit (20),
(20b), each digital data (that is, the switch (20a) of the switch circuit (20), (
20b) is in the ON state, no words are being generated and there is only environmental noise, so each digital data at this time consists only of noise components) is sent to the correction coefficient calculation circuit (24). The correction coefficient for noise removal is calculated here. That is, when the noise removal correction coefficient is An, calculation is performed according to the following equation.

Above (in formula 11, n is A/D converter + 7),
The number of signal samplings in (17), a 11 is the data from the main microphone (1) (in this case, the noise component), b
n is the data from the auxiliary microphone (■1) (in this case, the noise component), N and M are constants whose values are selected appropriately depending on the nature of the environmental noise and sampling frequency, and An-1 is the correction coefficient before one sampling. .

The correction coefficient calculation circuit (24) is used when there is no audio signal from the main ficrophone (11, that is, the switch circuit (20)
When the switches (20a) and (20b) are closed and in the ON state, this correction coefficient An is sequentially calculated and updated for each frequency band. Correction coefficient calculation circuit (24)
The correction coefficient An calculated in is stored in the memory (
25).

In addition, this correction coefficient An is used as much as possible during normal times, that is, when no words are generated, for recognition processing data Cn, which will be described later.
is set to a value close to zero.

Also, a switch (21a) of the switch circuit (21).

Each digital data passed through (21b) (that is, 2 of the switch circuit (21), isona (21a),
(When 21b> is turned on, words are generated and at least the main microphone (1) is given an audio signal, so the digital data from the A/D converter (7) at this time is the capital signal. is the noise component, and A/
The digital data from the D converter (17) is almost only a noise component) and is supplied to a difference detection circuit 1f8 (26) as an audio data calculation means. Further, the correction coefficient An stored in the memory (25) for each frequency +2h range is read out and input to the difference detection circuit (26). Therefore, the difference detection circuit (26) calculates the confirmation processing data Cn using the following equation based on each supplied digital data and the correction coefficient An.

Cn=an An-bn...121
In the above equation (2), al is data consisting of an audio signal and noise components in this case, and bn is data consisting mostly of noise components. As mentioned above, the correction coefficient An is set so that the recognition processing data Cn becomes zero as much as possible when no words are generated, so it is essentially the noise component contained in al and the noise contained in b11. The components are canceled out, and in the end, only the audio signal included in an is extracted as the recognition processing data Cn.

Recognition processing data Cn from this difference detection circuit (26)
is supplied to the sound source information normalizer (27a) of the speech recognition unit (27), normalized, and extracted as an acoustic parameter. This acoustic parameter is determined by the pattern matching circuit (
27b). Before recognition, the analysis result of each recognition target word of the speaker is registered in advance as a standard pattern in the standard pattern memory (27c), and when recognizing, the standard pattern of each recognition target word is stored in the memory (27c). A pattern matching circuit (27b) reads out acoustic parameters corresponding to the input audio pattern.
The 81 words to be recognized that are closest to each other, that is, those with the smallest VIi distance, are selected and outputted to the output terminal (28) as a recognition result indicating the human voice.

Next, the operation of the circuit shown in the figure will be explained. During normal times when no words are being generated, the main microphone ill and the auxiliary microphone (
■1) Since only environmental noise is collected, an OFF signal is obtained on the output side of the level detection circuit (22), which causes the switch (21a) of the switch circuit (21) to
(21b) opens and turns OFF, and turns OFF.
The ON signal obtained by inverting the signal with the inverter (23) causes the switches (20a), (
20b) is closed and becomes ON state.

Then, the noise collected by the microphones (11, (11)) is separated into frequency bands by the bandpass filter banks (31, (13), respectively), and then sent to the multiplexer (4).

(14), each is taken out in a time-divisional manner and sent to the A/D converter (
7) and (17), respectively. and A/1)
Each digital data (noise component) obtained on the output side of circuits (7) and (17) is sent to a correction coefficient calculation circuit (24).
The correction coefficient An is sequentially calculated for each frequency band according to the above equation (L) and stored in the memo 'J (25). This correction coefficient An is continued until a word is generated, that is, at least until the main microphone (1) starts collecting the audio signal, and a new correction coefficient An is successively applied for each frequency band so that the recognition processing data Cn becomes zero as much as possible. It is corrected to maintain the optimal value that is close to the value and stored in the memory (25).

Then, the words are generated at least from the main microphone +
1) collects an audio signal, its level exceeds the threshold level ゛t'h of the level detection circuit (22), so the output side of the level detection circuit (22) is turned on.
A signal is obtained, which closes the switches (21a) and (21b) of the switch circuit (21) and turns them on.
At the same time, the switches (20a) and (20b) of the switch circuit (20) are released by the OFF signal obtained by inverting the ON signal by the inverter (23), and the switches (20a) and (20b) are turned OFF. Therefore, in the memory (25), 16 correction coefficients An are stored in each frequency band, which are optimal for canceling the noise component immediately before a word is generated.

On the other hand, the audio signal including the noise component collected by the raw microphone (1) and the mainly environmental noise collected by the auxiliary microphone (2) undergo the same signal processing as described above and are supplied to the difference detection circuit (2b). be done. And the difference detection circuit (26
), the correction coefficient An of the corresponding frequency band is stored in memory (
25) and calculate the recognition processing data Cn according to the above equation (1). Since the correction coefficient An is set so that the recognition processing data Cn has a value as close to zero as possible during normal times, that is, when no words are generated, it is essentially a noise component contained in the audio signal from the main microphone ill. is canceled out by the noise component from the auxiliary microphone (11), and in the end, only the audio signal becomes the true recognition processing data C.
It is taken out as n.

This recognition processing data Cn is supplied to the music information normalizer (27a) of the Ondo recognition section (27), where it is normalized and extracted as acoustic parameters. These acoustic parameters are stored in memory (
27c) and is output to the output terminal (28) as a recognition result indicating human-powered sound.

In this way, in this embodiment, when subtracting data from the auxiliary microphone (11) from data from the main microphone (1), separate correction coefficients for noise removal are prepared for each frequency band, and the correction coefficient is always optimal. Since the correction is made to maintain the same value, it is possible to reliably cancel out environmental noise.Especially in the conventional method, it was difficult to predict the phase shift on the time axis, so it was difficult to cancel out environmental noise, but in this example, the frequency The data is divided so that phase information is not included in the data, and environmental noise is essentially canceled out on the frequency axis, so it is not affected by phase shifts (environmental noise can be definitely canceled out). .

In addition, since environmental noise is canceled for each frequency band on the frequency axis, it is not affected by differences in the characteristics of the main microphone + 1) and auxiliary microphone (11), or differences in frequency distribution or phase depending on the location. Since the noise removal correction coefficient is corrected until just before the signal is input, the size of the noise source,
Amplifier (2), which is not easily affected by position, sound quality, etc.
The noise generated by (12) can also be canceled out.

〔Effect of the invention〕

As described above, according to the present invention, environmental noise is collected using the main microphone and the auxiliary microphone, and then frequency-divided to calculate a correction coefficient for noise removal, and the sound including this correction coefficient and the environmental noise from the main microphone is collected. Since the audio data for recognition processing is obtained from the signal and at least the environmental noise from the auxiliary microphone, it is possible to cancel out the environmental noise without being affected by any phase shift on the frequency axis, thereby ensuring that the environmental noise is Misrecognition due to noise can be prevented.

[Brief explanation of the drawing]

The figure is a circuit configuration diagram showing an embodiment of the present invention. +11 is a raw microphone, (3) and (13) are band pass filter banks, (7) and (17) are A/
D converter, (11) is auxiliary microphone, (20),
(21) is a switch circuit, (22) is a level detection circuit, (24) is a correction coefficient calculation circuit, (25) is a memory,
(26) is a difference detection circuit, and (27) is a speech recognition section.

Claims (1)

[Scope of Claims] A speech recognition device that performs speech recognition by comparing acoustic parameters obtained by frequency-dividing and normalizing an audio signal collected by a main microphone with a preset standard pattern, which mainly uses environmental noise. A correction coefficient for noise removal is calculated by comparing data obtained by frequency-dividing the environmental noise collected by the auxiliary microphone with the data obtained by frequency-dividing the environmental noise collected by the auxiliary microphone. and a correction coefficient calculation means for calculating noise-free audio data from an audio signal including environmental noise collected by the main microphone, at least environmental noise collected by the auxiliary microphone, and the correction coefficient. 1. A speech recognition device, comprising: data calculation means, and performs speech recognition using the speech data.