JP2003044086A - Method and device for removing noise - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and a device for removing noise capable of improving the accuracy in noise removal. SOLUTION: The device has a configuration provided with a spectrum generating means 110 for generating an input spectrum which is the frequency spectrum of an acoustic signal composed of an audio signal desired to extract and a noise, a noise estimating means 130j for dividing the frequency band of the input spectrum into a plurality of frequency bands and generating the minimum value of the input spectrum as an estimated value of a noise spectrum, an SNR(signal to noise ratio) estimating means 140j for generating a ratio between the input spectrum and a value obtained by subtracting the estimated value of the noise spectrum from the input spectrum as an amplitude ratio of the audio signal and the noise, a removal spectrum estimating means 150j for generating the expected value of the frequency spectrum of the noise component to be removed concerning the phase difference of the audio signal and the noise signal, and a noise removing means 160j for subtracting the expected value of the frequency spectrum of the noise component from the input spectrum and generating the result as a frequency spectrum of the audio signal desired to extract.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise removal method in a system for analyzing an acoustic signal in the frequency domain, and more particularly to a noise removal method for removing noise from an audio signal in which voice signals to be extracted and noise are mixed. TECHNICAL FIELD The present invention relates to a denoising method and apparatus for estimating a signal.

[0002]

2. Description of the Related Art Conventionally, as a noise removing method, there is a spectral subtraction method in which a frequency spectrum of a noise component is subtracted from a frequency spectrum of an input acoustic signal in a frequency domain. This method estimates the frequency spectrum of noise in all frequency bands based on the average value of the frequency spectrum in the frames over the several frames immediately before the frame to be processed, the average value of the frequency spectrum in the time zone considered to be the unvoiced section, etc. , The power spectrum of the estimated noise is subtracted from the power spectrum of the input signal.

In this spectral subtraction method,
Since it is difficult to estimate the frequency spectrum of noise that should be subtracted based on an acoustic signal with mixed noise,
The average value of the frequency spectrum calculated as above,
Substitution as a noise spectrum to be subtracted is commonly done.

[0004]

However, the S / N ratio (hereinafter referred to as SNR (Si
gnal to Noise Ratio). ) Is not constant in general and changes with time, on the other hand, in the conventional noise removal method, noise removal is performed assuming a constant SNR without considering the temporal changes in SNR and phase. However, there is a problem that it is not possible to appropriately remove noise.

Further, in the conventional noise removal method, noise removal processing for subtracting a constant value over all frequencies is performed without dividing the frequency band, and it is not considered that the SNR differs depending on the frequency band, so the SNR is Even in the high frequency band, noise removal processing is performed using the same value as that in the low SNR frequency band, and there is a problem that the signal after the noise removal processing may rather deteriorate.

The present invention has been made to solve such a problem, and an object thereof is to reflect information on a phase difference between a voice signal and noise in noise removal, and at each point on a frequency spectrum. It is an object of the present invention to provide a noise removal method and device capable of improving the accuracy of noise removal by changing the amount of noise removal according to SNR.

[0007]

In view of the above points, the invention according to claim 1 uses a frequency spectrum of an acoustic signal composed of a voice signal to be extracted and noise which is a signal other than the voice signal to be extracted. Spectrum generating means for generating a certain input spectrum, noise estimating means for generating a minimum value of the input spectrum as an estimated value of the frequency spectrum of the noise, and SNR estimating means for generating as an amplitude ratio of the voice signal and the noise. A removal spectrum estimating means for generating an estimation value of a frequency spectrum of a noise component to be removed based on the amplitude ratio and an estimation value of the frequency spectrum of the noise, and an estimation value of the frequency spectrum of the noise component to be removed. And a noise removing means for subtracting from the input spectrum to generate the frequency spectrum of the audio signal to be extracted. We have the example was constructed.

With this configuration, noise removal processing is performed by using the amplitude ratio of the voice signal and the noise signal in consideration of the input spectrum which is the frequency spectrum of the acoustic signal.
The accuracy of noise removal can be improved.

Further, the invention according to claim 2 is based on claim 1, wherein the removal spectrum estimating means takes an expected value for a phase difference between the voice signal and the noise signal, and a frequency spectrum of the noise component to be removed. And the noise removing means has a configuration in which the estimated value of the frequency spectrum of the noise component to be removed is subtracted from the input spectrum as the estimated value of the frequency spectrum of the noise component to be removed. With this configuration, the noise removal processing is performed in consideration of the amplitude ratio between the voice signal and the noise signal in consideration of the input spectrum which is the frequency spectrum of the acoustic signal, and the phase difference between the voice signal and the noise signal. ,
The accuracy of noise removal can be improved.

The invention according to claim 3 provides the noise removing device according to claim 1 or 2, wherein the frequency band of the input spectrum is divided into a plurality of frequency bands, and the noise estimation is performed for each frequency band. Means produces an estimate of the frequency spectrum of the noise, the SNR estimate means produces the amplitude ratio, the remove spectrum estimate means produces an estimate of the frequency spectrum of the noise component to be removed, The removing means is configured to subtract the estimated value from the input spectrum to generate the frequency spectrum of the audio signal to be extracted. With this configuration,
Since the noise removal process is performed in consideration of the difference in SNR depending on the frequency band, it is possible to reduce the deterioration of the processed voice signal.

According to a fourth aspect of the invention, there is provided a spectrum generating step of generating an input spectrum which is a frequency spectrum of an audio signal composed of a voice signal to be extracted and noise which is a signal other than the voice signal to be extracted, A noise estimation step of generating a minimum value of an input spectrum as an estimated value of the frequency spectrum of the noise, an SNR estimation step of generating an amplitude ratio of the voice signal and the noise, and a step of estimating the amplitude ratio and the frequency spectrum of the noise. A removal spectrum estimation step of generating an estimated value of a frequency spectrum of a noise component to be removed based on the estimated value, and an audio signal to be extracted by subtracting the estimated value of the frequency spectrum of the noise component to be removed from the input spectrum And a noise removal step for generating the frequency spectrum of To have.

With this configuration, noise removal processing is performed by using the amplitude ratio of the voice signal and the noise signal in consideration of the input spectrum which is the frequency spectrum of the acoustic signal.
The accuracy of noise removal can be improved.

The invention according to claim 5 is the noise removing method according to claim 4, wherein the frequency band of the input spectrum is divided into a plurality of frequency bands, and the noise estimation step is performed for each frequency band. The estimated value of the frequency spectrum of the noise is generated, the amplitude ratio is generated in the SNR estimation step, and the frequency of the noise component to be removed with respect to the phase difference between the voice signal and the noise signal in the removal spectrum estimation step. Generate the expected value of the spectrum,
In the noise removing step, an expected value of the frequency spectrum of the noise component to be removed is subtracted from the input spectrum as an estimated value of the frequency spectrum of the noise component to be removed to generate a frequency spectrum of the voice signal to be extracted. Have

With this configuration, noise removal processing is performed in consideration of the amplitude ratio between the voice signal and the noise signal in consideration of the input spectrum, which is the frequency spectrum of the acoustic signal, and the phase difference between the voice signal and the noise signal. Therefore, the accuracy of noise removal can be improved, and since the noise removal processing is performed in consideration of the difference in SNR depending on the frequency band, deterioration of the processed voice signal can be reduced. .

[0015]

BEST MODE FOR CARRYING OUT THE INVENTION A noise removing method and apparatus according to a first embodiment of the present invention will be described below with reference to the accompanying drawings. FIG. 1 is a block diagram showing the configuration of the noise eliminator according to the first embodiment. The noise removal apparatus 100 shown in FIG. 1 includes an FFT analysis means 110, a filter bank analysis means 120, a noise estimation means 130, an SNR estimation means 140, a removed spectrum estimation means 150, and a noise subtraction means 160.

The filter bank analysis means 120 is composed of a plurality of filters for dividing an input signal into a plurality of frequency bands (hereinafter referred to as M frequency bands) and processing the same. A number for designating a frequency band (hereinafter referred to as a band) is represented by a symbol j. The noise estimation means 130, the SNR estimation means 140, the removed spectrum estimation means 150, and the noise subtraction means 160 exist for each filter that constitutes the filter bank analysis means 120, and perform the processing described later for each band.

In FIG. 1, noise estimating means 130j, S
NR estimation means 140j and removal spectrum estimation means 150
j and the noise subtraction means 160j, the symbol j added to the end of the symbols designating these means performs processing on the signal in the j-th band output from the filter bank analysis means 120. It is a symbol indicating that it is a means.

In FIG. 1, an acoustic signal 10 in which a voice signal and noise are mixed is input to a noise eliminating device 100, and an acoustic spectrum 20 after noise elimination processing is output from the noise eliminating device 100. The FFT analysis means 110 applies a fast Fourier transform (hereinafter, F
It is called FT (Fast Fourier Transformation). ) Processing is performed to generate an input spectrum which is a frequency spectrum of the acoustic signal 10.

The filter bank analysis means 120 performs the conversion process shown in the following equation (1) on the above input spectrum.

[Equation 1] Here, the argument t is a number that specifies the frame to be processed,
The subscript i is a number that specifies the numerical value of the frequency.

Further, X _i (t), S _i (t) and N _i (t) are the input spectrum and the speech signal at the frame designated by the argument t and the numerical frequency designated by the subscript i, respectively. And the frequency spectrum of noise and the frequency spectrum of noise. Further, θ _i (t) is a phase difference between the frequency spectrum S _i (t) and N _i (t), and Mel _ij is a filter coefficient.

In the following, the filter bank analysis means 120
The filter that processes the signal in the j-th band is referred to as the j-th filter. Also,
A mel filter bank can be used as the filter bank analysis unit 120 for performing band division and filtering. In that case, Mel _ij is a filter bank coefficient that correlates the i-th value in the input spectrum with the output BX _j (t) from the j-th filter that constitutes the filter bank analysis means 120.

The conversion result when a process similar to the process for the input spectrum X _i (t) is performed on the frequency spectrum S _i (t) of the speech signal is BS _j (t), and the following equation ( It is defined by 2).

[Equation 2]

Hereinafter, the acoustic spectrum 20 after the removal processing output from the noise removal apparatus 100 will be derived. The component Z _j (t) corresponding to the noise after being processed by the filter bank analysis means 120 can be expressed by the following equation (3) using the equations (1) and (2).

[Equation 3] Where r _i (t) is | S _i (t) | / | N _i (t) | (hereinafter, r _i (t)
Is referred to as an amplitude ratio), and an amount corresponding to SNR.

The component Z _j (t) corresponding to the noise after processing by the filter bank analysis means 120 has a term represented by the following equation (4) in the absolute value | N _i (t) | of the noise spectrum. It is shown that the result of multiplication is equivalent to being converted into the mel filter bank and converted.

[Equation 4]

Here, the expected value γ (r) of the above equation (4) for the phase difference θ is defined as the following equation (5).

[Equation 5] In the above formula (5), the subscript i is omitted for simplification. As is clear from the equation (5), the expected value γ (r) is a function of the amplitude ratio r.

The estimated value j (t)> of the noise spectrum of each band is defined by the following equation (6).

[Equation 6] This expression (6) is used for processing the target frame t and this frame t
In each frame from the frame to the frame before the T frame, the value at which the output from the filter bank analysis unit 120 becomes the minimum for the j-th band is the estimated value j (t)> (of the noise spectrum of each band. Hereinafter, it is defined as M _minj (t).).

Here, the amplitude ratio, which is an amount corresponding to the SNR,
r _i a (t), as a function of the number j that specifies the band rather than a function of the number i specifying the frequency, substitute approximated by r _j ^'(t) represented by the following equation (7).

[Equation 7]

From the above, the estimated value of the frequency spectrum of the noise component to be removed is expressed by equation (8).

[Equation 8] By subtracting Z _j (t) derived in this way from the input signal of the j-th band, it is possible to obtain the frequency spectrum of the audio signal to be extracted.

The components of the noise eliminator 100 will be described below.
Returning to the explanation of the processing for opening. Noise estimation means 130
Is the noise spectrum of each band based on equation (6).
Estimated value M_minj(t) is the output of the filter bank analysis means 120
Generate from. The SNR estimation means 140 is based on the equation (7).
Then, of the amplitude ratio, which is the amount corresponding to the SNR in each band,
Estimate r_j ^'(t) is the output of the filter bank analysis means 120
And noise spectrum estimate M_minjgenerate using (t)
It The removal spectrum estimation means 150 uses the equations (5) and
Based on equation (8), the frequency spectrum of the noise component to be removed is
Estimated value of Koutor Z_j(t) is the noise spectrum estimate M _minj
(t) and the estimated value r of the above amplitude ratio_j ^'generate using (t)
It

The noise subtraction means 160 performs noise removal processing on the output from the filter bank analysis means 120 based on the following equation (9).

[Equation 9] Here, η is a flooring parameter, which defines a lower limit of the estimated value BS _j (t) of the frequency spectrum of the voice signal.

In the above expression (9), the estimated value BS _j (t) of the frequency spectrum of the voice signal obtained by removing the estimated value Z _j (t) of the frequency spectrum of the noise component is defined by η. Is larger than the estimated value η BX _j (t), noise is removed by subtracting the estimated value Z _j (t) of the frequency spectrum of the noise component from the estimated value BS _j (t). The estimated value BS _j (t) of the frequency spectrum is the value η BX defined by η.
When it is smaller than _j (t), it is set to a value η BX _j (t) defined by η in order to prevent deterioration of the extracted audio signal.

In the above, the input acoustic signal 1 is calculated using the filter bank analysis means 120 composed of a plurality of filters.
Although the apparatus for dividing the frequency spectrum of 0 into a plurality of frequency bands and performing the processing for noise removal for each band has been described, the noise removal apparatus according to the present invention does not necessarily perform filter bank analysis for performing band division. The means 120 is not required and may be configured by one filter. The configuration of the noise eliminator at this time is shown in FIG. The processing in each component shown in FIG. 2 is the same as that in each component designated by the same number.

FIG. 3 is a flow chart showing the flow of processing in the noise elimination method performed by the noise elimination device according to the first embodiment of the present invention. Step S301
Then, the FFT analysis means 110 receives the input acoustic signal 10
To the fast Fourier transform (hereinafter, FFT (Fast Fourier
er Transformation). ) Processing is performed to generate an input spectrum which is a frequency spectrum of the acoustic signal 10.

In step S302, the filter bank analysis means 120 adds the following equation (1) to the above input spectrum.
The conversion process shown in 0) is performed.

[Equation 10] Here, the argument t is a number that specifies the frame to be processed,
The subscript i is a number that specifies the numerical value of the frequency.

In step S303j, the noise estimation means 13
0 produces _{| generates} the estimated value _Mminj (t) of the noise spectrum in each band from the output of the filter bank analysis means 120 based on Formula (11).

[Equation 11]

In step S304j, the SNR estimating means 1
The reference numeral 40 generates an estimated value r _j ^' (t) of an approximate value of the amplitude ratio, which is an amount corresponding to the SNR in the j-th filter constituting the filter bank analysis means 120, based on the following formula (12). .

[Equation 12]

In step S305j, the removed spectrum estimation means 150 calculates the estimated value Z _j (t) of the frequency spectrum of the noise component to be removed based on the following equation (13) and the estimated value M _minj (of the noise spectrum). t) and the estimated value r _j ^' (t) of the above amplitude ratio.

[Equation 13]

Here, γ (r) is a quantity defined by the following equation (14).

[Equation 14]

In step S306j, the noise subtraction means 16
0 applies noise removal processing to the output from the filter bank analysis means 120 based on the following equation (15).

[Equation 15] Here, η is a flooring parameter, which defines a lower limit of the estimated value BS _j (t) of the frequency spectrum of the voice signal.

In the above formula (15), the estimated value BS _j (t) of the frequency spectrum of the voice signal obtained by removing the estimated value Z _j (t) of the frequency spectrum of the noise component is defined by η. Is larger than the estimated value η BX _j (t), noise is removed by subtracting the estimated value Z _j (t) of the frequency spectrum of the noise component from the estimated value BS _j (t). The estimated value BS _j (t) of the frequency spectrum is the value η defined by η.
When BX _j (t) less than are those values defined by η in order to prevent degradation of the audio signal ηBX _j (t) to be extracted.

In the above, the method of dividing the frequency spectrum of the input acoustic signal 10 into a plurality of frequency bands and performing processing for noise removal for each band has been described, but the noise removal method according to the present invention is not necessarily required. It is not necessary to perform processing for band division, and a configuration in which all frequency bands are subjected to filter processing at one time may be used. FIG. 4 is a flowchart showing the flow of processing in the noise removing method at this time. The processing in each step shown in FIG. 4 is the same as the processing in each of the above steps designated by the same number when M is 1.

As described above, since the noise removing method and apparatus according to the first embodiment perform the noise removing process in consideration of the SNR and the phase, the noise removing accuracy can be improved. it can. Also, depending on the frequency band, SNR
Since the noise removal processing is performed in consideration of the difference between, the deterioration of the processed voice signal can be reduced.
Furthermore, since the band is divided using the filter bank and the noise removal processing is performed, the number of parameters to be processed can be reduced and the processing amount can be reduced.

[0043]

As described above, according to the present invention, the information on the phase difference between the voice signal and the noise is reflected in the noise removal, and the noise removal amount is adjusted according to the SNR at each point on the frequency spectrum. Since it can be changed, it is possible to provide a noise removal method and device capable of improving the accuracy of noise removal.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a noise elimination device according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing a device configuration in the case of performing processing without dividing a frequency band in the noise removal device according to the first embodiment of the present invention.

FIG. 3 is a flowchart showing the flow of processing in the noise removal method according to the first embodiment of the present invention.

FIG. 4 is a flowchart showing a processing flow in the case of processing without dividing a frequency band in the noise removing method according to the first embodiment of the present invention.

[Explanation of symbols]

10 acoustic signals 20 Output spectrum 100 noise eliminator 110 FFT analysis means 120 Filter bank analysis means 130, 130j Noise estimation means 140, 140j SNR estimation means 150, 150j Removal spectrum estimating means 160, 160j noise subtraction means

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Toru Imai             1-10-11 Kinuta, Setagaya-ku, Tokyo, Japan             Broadcasting Association Broadcast Technology Institute (72) Inventor Shoe Sato             1-10-11 Kinuta, Setagaya-ku, Tokyo, Japan             Broadcasting Association Broadcast Technology Institute (72) Inventor Ken Kobayakawa             1-10-11 Kinuta, Setagaya-ku, Tokyo, Japan             Broadcasting Association Broadcast Technology Institute (72) Inventor Hiroyuki Seki             1-10-11 Kinuta, Setagaya-ku, Tokyo, Japan             Broadcasting Association Broadcast Technology Institute F-term (reference) 5D015 EE05

Claims (5)

[Claims] 1. A spectrum generating means for generating an input spectrum which is a frequency spectrum of an acoustic signal composed of a voice signal to be extracted and noise which is a signal other than the voice signal to be extracted, and a minimum value of the input spectrum to the noise. Noise estimation means for generating as an estimated value of the frequency spectrum of SNR, SNR estimation means for generating as an amplitude ratio of the voice signal and the noise, and removal based on the estimated value of the amplitude ratio and the frequency spectrum of the noise. Removal spectrum estimating means for generating an estimated value of a frequency spectrum of a noise component to be removed, and noise removal generated as a frequency spectrum of the voice signal to be extracted by subtracting the estimated value of the frequency spectrum of the noise component to be removed from the input spectrum And a noise removing device. 2. The removing spectrum estimating means estimates an frequency spectrum of the noise component to be removed by taking an expected value with respect to a phase difference between the voice signal and the noise signal, and the noise removing means removes the removed signal. 2. The noise removing apparatus according to claim 1, wherein an estimated value of a frequency spectrum of a power noise component is subtracted from the input spectrum as an estimated value of a frequency spectrum of the noise component to be removed. 3. The noise elimination device divides the frequency band of the input spectrum into a plurality of frequency bands, and the noise estimation means generates an estimated value of the frequency spectrum of the noise for each frequency band, The SNR estimating means generates the amplitude ratio, the removing spectrum estimating means generates an estimated value of the frequency spectrum of the noise component to be removed, and the noise removing means subtracts the estimated value from the input spectrum to perform the extraction. The noise eliminator according to claim 1 or 2, wherein a frequency spectrum of a desired audio signal is generated. 4. A spectrum generating step of generating an input spectrum which is a frequency spectrum of an acoustic signal composed of a voice signal to be extracted and noise which is a signal other than the voice signal to be extracted, and a minimum value of the input spectrum is set to the noise. Based on the amplitude ratio and the estimated value of the frequency spectrum of the noise, and the noise estimation step of generating the estimated value of the frequency spectrum of the noise, the SNR estimation step of generating the amplitude ratio of the voice signal and the noise, Removal spectrum estimation step of generating an estimated value of the frequency spectrum of the power noise component, and noise removal generated as the frequency spectrum of the voice signal to be extracted by subtracting the estimated value of the frequency spectrum of the noise component to be removed from the input spectrum A method for removing noise, comprising: 5. The noise elimination method divides a frequency band of the input spectrum into a plurality of frequency bands, and generates an estimated value of the frequency spectrum of the noise in the noise estimation step for each frequency band, In the SNR estimation step, the amplitude ratio is generated, in the removal spectrum estimation step, an expected value of the frequency spectrum of the noise component to be removed is generated with respect to the phase difference between the voice signal and the noise signal, and the noise removal step is performed. The frequency spectrum of the speech signal to be extracted is generated by subtracting the expected value of the frequency spectrum of the noise component to be removed as an estimated value of the frequency spectrum of the noise component to be removed from the input spectrum. 4. The noise removal method described in 4.