JPH03274099A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To remove a noise properly even if the playback sound of a radio set, a stereophonic set, etc., is superposed as the noise on a voice signal by providing a 2nd feature quantity extraction part which extracts the feature quantity of the noise from a playback signal reproduced by a speaker placed at the periphery of a microphone outside a 1st feature quantity extraction part which extracts the feature quantity of a noise-containing voice inputted from the microphone. CONSTITUTION:The 1st noise removal part 30 subtracts the time spectrum pattern of the noise of a speaker playback signal extracted by a 2nd feature quantity extraction part 20 from the time spectrum pattern of the voice containing the noise extracted by the 1st feature quantity extraction part 10 to generate a time spectrum pattern. A 2nd noise removal part 60 estimates the time spectrum pattern of a noise other than the noise of the speaker playback signal by a noise estimation part 50 and subtracts it from the time spectrum pattern, in a voice section detected by a voice section part 40, generated by the 1st noise removal part 30. An input pattern generation part 70 generates the input pattern of a voice inputted according to a known BTSP voice recognizing means from the feature quantity of the voice generated by the 2nd noise removal part 60.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition device, and more particularly, to a noise removal technique in speech recognition in a high-noise environment, such as dialing in a car, audio equipment, etc. The present invention is suitable for application to a voice recognition device for controlling a car, a near conditioner, a navigation system, etc., and can also be applied at home, office, etc.

In recent years, voice-based information input methods have been attracting attention.
You can also make calls from your car phone while in your car.

Audio equipment control, near conditioner control,
It is being considered to apply voice recognition technology to control navigation systems, etc.

However, voice recognition in a car is difficult due to the fact that the sound of the engine, running tires, or sound played from a radio or stereo mixes into the voice signal as noise, and the distance between the mouth and the mouth of a close-talking microphone while driving. Since it is not possible to attach a microphone close to the device, the ratio of the audio signal to the noise, that is, the S/N ratio is poor, which poses a problem, and noise removal technology is indispensable.

As a noise removal technology in conventional speech recognition.

There are spectral subtraction methods such as S, F, and Boll, and adaptive noise canceling methods such as B and Iidrow.

However, the spectral subtraction method is sensitive to time-unsteady noise. For example, when there is a large amount of time-unsteady noise in the audio band, such as radio or stereo playback (speech or musical tones), the spectral subtraction method first requires speech section detection. However, even if speech sections can be detected, the estimated noise component amount is different from the actual noise component. In addition, adaptive noise canceling has the disadvantage that it cannot respond well to noise synthesized from multiple noise sources when using two people, that is, two microphones.
However, even if the number of microphones is increased unnecessarily, the amount of signal processing will be enormous, and the cost will be high, making it difficult to put it into practical use.

In addition, regarding the sound played from speakers such as radios and stereos (which is noise for voice recognition devices), it is possible to subtract the signal sent to the speakers by applying an appropriate gain from the input from the microphone. Since there is a delay from the speaker to the microphone and reflected sound from the surroundings, simply taking the difference between the two on the time waveform cannot be expected to be effective.

The present invention was made in view of the drawbacks of the prior art as described above, and is particularly applicable to radios in the voice band that occurs near the microphone for inputting voice and which has strong time non-stationarity. , even if playback sound from stereo etc. is superimposed on the audio signal as noise, the purpose was to appropriately remove this noise and achieve good speech recognition in such a noisy environment. .

In order to achieve the above object, the present invention includes: (1) a first feature extracting section that extracts features of a voice including noise input from a microphone; a second feature extraction unit that extracts a feature amount of noise caused by the speaker reproduction signal using the speaker reproduction signal to be reproduced from the speaker, and a feature extracted by the first feature extraction unit; a first noise removal unit that generates a feature quantity by removing the feature quantity extracted by the second feature quantity extraction unit from the quantity;
a speech section detection section that detects a speech section from the feature quantity generated by the first noise removal section; and a feature generated by the first noise removal section in the non-speech section detected by the speech section detection section. An estimated value of noise other than the noise caused by the speaker reproduction signal is calculated from the amount, and further, an estimated value of the noise other than the noise caused by the speaker reproduction signal is calculated from the feature amount generated by the first noise removal unit in the voice interval detected by the voice interval detection unit. a second noise removal unit that removes estimated values of noise other than the noise caused by the noise and generates a voice feature amount; and an input pattern of the voice input from the voice feature amount generated by the second noise removal unit. an input pattern generation unit that generates an input pattern, a standard pattern memory that stores a standard pattern of speech, and a recognition process that performs recognition processing using the input pattern generated by the input pattern generation unit and the standard pattern stored in the standard pattern memory. (2)
) In the speech recognition device of (1) above, the second feature extracting section extracts the speaker reproduction signal to be reproduced from speakers placed around the microphone.

Shinsuke's special yll amount, which has been subjected to processing corresponding to the transfer function between the speaker reproduction signal measured in advance in a predetermined positional relationship between the speaker and the microphone, and the signal input from the microphone, is transmitted to the speaker. It is characterized in that it is a feature amount of noise due to a reproduction signal, and furthermore, (
3) In the speech recognition device of (1) or (2) above, when there are multiple speakers placed around the microphone, the speaker reproduction signal for reproduction from each speaker is A signal that is the sum of all the signals that have been processed in accordance with the transfer function between each of the above speaker reproduction signals and the signal input from the above microphone, which are measured in advance in a predetermined positional relationship between the above microphone and the above microphone. This feature is characterized in that the feature quantity is used as the feature quantity of the noise due to the speaker reproduction signal. An explanation will be given below based on one embodiment of the present invention.

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention. In the figure, 1 is a microphone, 2 is a speaker placed around the microphone 1, and 10 includes noise input from the microphone 1. 20 is a second feature extraction unit that extracts the feature amount of noise due to the speaker reproduction signal; 30 is the first noise removal unit; and 40 is a voice section. 50 is a noise estimator, and 60 is a second noise remover.

70 is an input pattern generation section, 80 is a standard pattern memory, and 90 is a recognition section. A second feature amount that extracts a feature amount of noise due to a speaker reproduction signal using a first feature amount extraction unit 10 and a speaker reproduction signal to be reproduced from the speakers 2 placed around the microphone. Extraction part 20 and the above ↓
a first noise removal section 30 that generates a feature amount by removing the feature amount extracted by the second characteristic amount extraction section 20 from the feature amount extracted by the feature amount extraction section 10; a speech section detection section 40 that detects a speech section from the feature amount generated by the noise removal section; and a feature amount generated by the first noise removal section 30 in the non-speech section detected by the speech section detection section 40. The noise estimator 50 estimates and obtains an estimated value of noise other than the noise caused by the speaker reproduction signal from A second noise removing unit 60 that removes estimated values of noise other than the noise caused by the speaker reproduction signal from the generated feature amount to generate a voice feature amount; an input pattern generation unit 70 that generates an input pattern of the input voice from the feature amount of the input voice; a standard pattern memory 80 that stores a standard pattern of voice; and the input pattern 70 generated by the input pattern generation unit and the It is separated from the recognition section 90 that performs recognition processing using the standard pattern stored in the standard pattern memory 80.

To explain in more detail, the first feature extraction unit 10 extracts the feature amount of the voice including noise input from the microphone 1 installed in the car for inputting voice. , perform amplification, pre-emphasis 1
2 emphasizes the high frequency range, and the bandpass filter bank 13 includes a group of bandpass filters whose center frequencies are 15 frequencies arranged at equal intervals on the logarithmic axis from 25011z to 6.35KIlz, and each band. It consists of a rectifier and a low-pass filter for each channel, and the spectrum of the input voice is determined using these. The multiplexer 14 switches the data of each band, and the A/D converter 15 digitizes the data of each band at a sampling period of IOMS.

Therefore, the signal input to the first feature extraction unit 10 is
After passing through the microphone amplifier 11, pre-emphasis 12, bandpass filter bank 13, multiplexer 14, and A/D converter 15, the time spectrum pattern of the audio including noise X(t, f) (where t is the time axis and f is (on the frequency axis). Further, the second feature amount extraction unit 20 extracts the feature amount of noise due to the speaker reproduction signal using the speaker reproduction signal to be reproduced from the speaker 2 placed around the microphone 1, The pre-emphasis 22 emphasizes high frequencies in the same manner as the pre-emphasis 12, the band-pass filter bank 23 obtains the spectrum of the speaker reproduction signal in the same way as the band-pass filter bank 13, and the multiplexer 24
The data of each band is switched in the same manner as in 4, and the A/D converter 25 digitizes the data of each band in the same manner as the A/D converter 15. The signal input to the second feature extraction unit 20 is processed by pre-emphasis 22. Bandpass filter bank 23. Multiplexer 24, A/
After passing through the D converter 25, the time spectrum pattern N(t, f) of the signal for speaker reproduction is obtained. Further, for example, the time spectrum pattern N(t, f) of this speaker reproduction signal and the transfer function H(f) between the speaker 2 and the microphone 1 measured in advance and stored in the transfer function memory 28.
The product N(t,,f)・H(f) is calculated by the multiplier 27, and the time spectrum pattern of noise due to the speaker reproduction signal N1(t,f)=N(t,f)・H(f ). Note that the transfer function H(f) between the speaker 2 and the microphone 1 is the second
This can be obtained in advance by giving an impulse signal to the speaker reproduction signal input to the feature extracting unit 20 of the feature extraction unit 20, and performing Fourier transform on the impulse response obtained by collecting the sound reproduced through the speaker 2 from the microphone 1. It can also be obtained by reproducing white noise or frequency sweep signals.

The first noise removal unit 30 extracts a time spectrum pattern
The time spectrum pattern N1 (t, f) of the noise due to the speaker reproduction signal extracted by the second feature extraction unit 20 is subtracted from (t, f), and the noise due to the speaker reproduction signal is removed from the noisy audio signal. Removed time spectrum pattern x1(t,f)=x(t,f)-N1. (t,
f).

The speech section detection section 40 generates a time spectrum pattern XI (t, f
) to detect the voice section.

The voice section detection method used here detects the section in which the total of 5 Σ XI (t, f) f=1 in each frame of the time spectrum pattern This is a voice section.

The second noise removal unit 60 generates a time spectrum pattern obtained by removing noise caused by the speaker reproduction signal from the noise-containing audio signal generated by the first noise removal unit 30 in the audio interval detected by the audio interval detection unit 40. X 1s(t,
f) The noise estimator 50 estimates and subtracts the time spectrum pattern N 2 (t, f) of noise other than the noise caused by the speaker reproduction signal from (the subscript S represents the speech section) to obtain the speech time. Spectral pattern S (t, f) = X 1. s(t, f) −N 2
Generate (t, f). Here, the time spectrum pattern N2 (
tlf) is estimated by the noise estimation unit 50 and applied to the average of a plurality of frames of the output X1n(t, f) of the first noise removal unit 30 when it is not a voice section, according to the known spectral subtraction system (see appendix). The letter n represents a non-speech section).

The input pattern generation unit 70 generates a known BTSP (Bina
ry Ti+ie Spectrum Pattern
) Generate an input pattern of the input voice according to the voice pattern generation method of the voice recognition method.

The standard pattern memory 80 stores a standard pattern of speech in the standard pattern format of the known BTSP speech recognition system.

The recognition unit 90 performs recognition processing using the input pattern generated by the input pattern generation unit 70 and the standard pattern stored in the standard pattern memory 80 according to the recognition process of the well-known BTSP speech recognition method.

In addition to the means used in the embodiments described above, the speech section detection method of the speech section detection section 40, the noise removal method of the second noise removal section 60, the pattern generation method of the input pattern generation section 70, and the standard pattern Pattern format of memory 80, recognition unit 90
The present invention can also be implemented using a known method for recognition processing.

Also, the bandpass filter bank 13.23.

It may be replaced with digital signal processing such as FFT, and the A/D converters 15 and 25 may be shared by time-division processing.

FIG. 2 is a configuration diagram of an embodiment in which the embodiment shown in FIG. 1 is expanded to accommodate a case where there are multiple speakers.
As shown in the figure, if the number of speakers is two, the second
There are also two feature extraction units 20a and 20b, and the speaker 28
．． The time spectrum patterns Na(t, f) t Nb(t, f) of these speaker playback signals obtained by the bandpass filters 23a and 23b are inputted to the respective speaker playback signals to be played back from the band pass filters 23a and 23b. , a transfer function Ha(f) between the speakers 2a, 2b and the microphone 1 measured in advance.

)Ib(f) are calculated by multipliers 27a and 27b, and the sum of these products is calculated by adder 29 as a time spectrum pattern of noise due to the speaker reproduction signal N1(t,f) = Na(t , f)・Ha(f)
+ Nb(t, fl Hb(f). The subsequent processing is the same as the embodiment shown in FIG. 1. Even if the number of speakers is three or more, the same method can be used.

As is clear from the above description, according to the invention of claim 1, in the noise removal unit 30, the noise removing unit 30 detects noise in the voice band and which has strong time non-stationarity from the feature amount of the voice including noise. After removing the noise features caused by the speaker reproduction signal,
Since the voice section is detected by the voice section detecting section 40,
Noises other than speaker playback sounds 1 For example, in a car, the engine sound, the sound of running tires, and in homes and offices, it is typical of noise, all of which have a low frequency range (100 Hz to several 100 Hz). degree)
Moreover, since the speech section is detected from the main noise, which is strongly time-stationary, the detection accuracy is improved, and as a result, the recognition rate of speech under high noise is improved.

Further, according to the invention of claim 2, the second feature extracting section 20 includes a signal for reproducing the speaker reproduction signal to be reproduced from the speakers placed around the microphone. A first noise removal unit sets the feature amount of the signal that has been processed corresponding to the impulse response between the speaker and the microphone measured in advance in a predetermined positional relationship as the feature amount of the noise due to the speaker reproduction signal. 30 removes the feature amount of noise due to the signal for speaker reproduction from the feature amount of the voice containing noise, so the noise component due to the signal for speaker reproduction can be accurately removed, and as a result, the recognition rate of speech under high noise can be improved. is improved.

Further, according to the third aspect of the invention, even when there are a plurality of speakers 2 placed around the microphone 1, the second feature extraction units 20a, 20b10. .

However, each speaker 2a, 2b188. and microphone 1
Each of the speakers 2a, 2b191. Multipliers 27a and 27b10 perform respective processing corresponding to the transfer function between
0. The feature amount of the signal obtained by adding all the signals processed in step 1 with the adder 29 is set as the feature amount of the noise due to the speaker reproduction signal,
Since the first noise removal unit 30 removes the characteristics of the noise caused by the speaker reproduction signal from the feature amount of the noise-containing voice,
Even when there are a plurality of speakers 2, the noise component caused by the speaker reproduction signal can be accurately removed, and as a result, the recognition rate of speech under high noise is improved.

[Brief explanation of drawings]

FIG. 1 is a configuration diagram for explaining an embodiment of the invention described in claims 1 and 2, and FIG.
FIG. 2 is a configuration diagram for explaining an embodiment of the invention described in section 1. 1...Microphone, 2.2a, 2b...Speaker, 10
...First feature extraction unit, 20, 20a, 20b...
Second feature extraction unit, 11... Microphone amplifier, 12°22.22a, 22b... Pre-emphasis circuit, 1
3.23, 23a, 23b - bandpass filter bank, 14, 24, 24 a, 24 b - multiplexer, 15, 25, 25a, 25b = A/D converter, 27, 27 a, 27 b - - multiplication vessel. 28.28a, 28b...transfer function memory, 29...
-Adder, 30...first noise removal section, 40...speech section detection section, 50...noise estimation section, 60...second
Noise removal unit, 70... Input pattern generation unit, 80.
...Standard pattern memory, 90...Recognition section. The life of the audience

Claims (1)

[Scope of Claims] 1. A first feature extraction unit that extracts the feature amount of the sound including noise input from the microphone, and a speaker reproduction unit for reproducing the sound from speakers placed around the microphone. a second feature extraction section that uses the signal to extract a feature amount of noise due to a signal for speaker reproduction; and a second feature extraction section that extracts the feature amount from the feature amount extracted by the first feature extraction section. a first noise removal section that generates a feature amount by removing the feature amount that has been removed; a speech section detection section that detects a speech section from the feature amount generated by the first noise removal section; and the speech section detection section. An estimated value of noise other than the noise caused by the speaker reproduction signal is calculated from the feature amount generated by the first noise removal section in the non-speech section detected in the non-speech section, and further a second noise removing unit that removes an estimated value of noise other than the noise caused by the speaker reproduction signal from the feature generated by the first noise removing unit to generate a voice feature; an input pattern generation unit that generates an input pattern of the input voice from the feature amount of the voice generated by the noise removal unit; a standard pattern memory that stores a standard pattern of voice; and a standard pattern memory that stores the standard pattern of voice; A speech recognition device comprising: a recognition unit that performs recognition processing using an input pattern and a standard pattern stored in the standard pattern memory. 2. The second feature extracting section includes a feature that has been measured in advance in a predetermined positional relationship between the speaker and the microphone, in the speaker reproduction signal to be reproduced from speakers placed around the microphone. Claim 1, characterized in that a feature quantity of a signal subjected to processing corresponding to a transfer function between the speaker reproduction signal and the signal input from the microphone is used as a feature quantity of noise due to the speaker reproduction signal. The speech recognition device described in . 3. When there are multiple speakers placed around the microphone, the speaker reproduction signal to be reproduced from each speaker is measured in advance in a predetermined positional relationship between each speaker and the microphone. The feature amount of the signal obtained by adding all the processed signals corresponding to the transfer function between each of the above-mentioned speaker reproduction signals and the signal input from the microphone is the feature amount of the noise due to the speaker reproduction signal. Claim 1 characterized in that
The speech recognition device according to item 1 or 2.