JP3499113B2 - Noise removal device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise eliminating device for eliminating noise contained in an input signal.

[0002]

2. Description of the Related Art As one of general methods for noise reduction,
Spectral difference method (eg SFBoll, "Suppression o
f Acoustic Noise in Speech Using Spectral Subtract
ion ", IEEE Trans. ASSP-27 No.2, 1979/4, pp.113-12
There is 0).

An outline of the processing of the above spectral difference method will be described below. A noise section is detected from the input signal in which voice and noise are mixed, and the noise power spectrum is extracted. The voice power spectrum is estimated by subtracting the previously extracted noise power spectrum from the input power spectrum of the input signal. It is composed of a process of restoring a voice waveform from the voice power spectrum and the information of the phase spectrum obtained from the input signal.

In the spectral difference method, frequency conversion means such as FFT is generally used, but a method for realizing frequency analysis / synthesis by a filter bank for band division is disclosed in Japanese Patent Laid-Open No. 59-67732. ing. Similar to the spectral difference method, it consists of the process of estimating the energy of noise for each band and subtracting it from the signal in which voice and noise are mixed.

[0005]

However, in the above-mentioned prior art, it is necessary to determine the noise section, and in the method of determining the noise section from information such as a change in power and spectrum, especially for non-stationary noise. Has a problem that it is difficult to determine the noise section and it is difficult to extract the noise power spectrum.

[0006]

In order to solve the above-mentioned problems, a noise elimination device of the present invention estimates a noise component contained in the input signal and a frequency analyzer for calculating frequency spectrum information of the input signal. Noise removal provided with a noise power spectrum estimator, a subtractor for subtracting the noise power spectrum from the input power spectrum, and a frequency synthesizer for restoring the phase spectrum of the input signal and the power spectrum obtained by the subtractor into a time signal In the apparatus, the noise power spectrum estimator is provided with means for extracting, for each frequency, a small small level power spectrum within a certain fixed period as a noise power spectrum.

That is, the frequency analyzer calculates the power spectrum and phase spectrum of the input signal in which voice and noise are mixed by a method such as FFT. The noise power spectrum estimator inputs the power spectrum output from the frequency analyzer and extracts a small small level power spectrum within a certain fixed period for each frequency as a noise power spectrum. The subtractor subtracts the noise power spectrum from the power spectrum of the input signal in which voice and noise are mixed, and estimates the voice power spectrum. The frequency synthesizer calculates the inverse FF from the speech power spectrum estimated by the subtractor and the phase spectrum of the input signal in which speech and noise are mixed.
The audio signal is restored by a method such as T. The noise noise
The power spectrum estimator is
A non-linear converter for weighting
Converted power spectrum from long-term power spectrum
Long-time power spectrum calculator to calculate
To convert the power spectrum between
And a linear inverse converter. In another aspect, above
The noise power spectrum estimator is the input power spectrum
A non-linear converter that weights small values of
Converted converted power spectrum to long-term power spectrum
A long time power spectrum calculator to calculate the torque
Estimate the variance of the non-linearly transformed power spectrum
Variance estimator,
Calculate the noise reduction amount from the variance of the converted power spectrum
Noise rejection amount calculator and the estimated noise removal amount
And an exponential calculator for returning the power spectrum.
In one embodiment, the denoising device comprises an input power spectrum.
From the power spectrum and noise power spectrum,
Voice detector that detects voice by comparing information
Control the amount of sound signal attenuation according to the judgment result.
Attenuation controller and attenuator that attenuates the amplitude of audio signals
It is desirable to have

[0008]

DETAILED DESCRIPTION OF THE INVENTION

[Embodiment 1] FIG. 1 is a block diagram showing Embodiment 1 of the noise elimination apparatus of the present invention. The noise eliminator shown in FIG. 1 has an input terminal 101 for inputting an input signal in which voice and noise are mixed, a window multiplier 102 for applying a humming window or the like to analyze the input signal by FFT, power by a method such as FFT. Frequency analyzer 103 for calculating the spectrum and phase spectrum, noise power spectrum estimator 104 for estimating the noise power spectrum included in the input signal, subtraction of the noise power spectrum from the power spectrum of the input signal, and subtraction for estimating the voice power spectrum From the estimated speech power spectrum and the phase spectrum obtained from the frequency analyzer 101, the frequency synthesizer 106 that restores the time signal and the signal obtained by the frequency synthesizer 106 are appropriately superposed and added. An overlap adder 107 and an output terminal 108 that obtain continuous time signals by Et al constructed.

From the input terminal 101, a voice signal s [t] which is a time signal (where t represents time) and a noise signal n
The input signal x [t] (= s [t] + n [t]) added by [t] is input and output to the window multiplier 102.

The window multiplier 102 blocks the input start position of the input signal obtained from the input terminal 101 into a certain fixed sample length (for example, 256 samples) while shifting it by a certain fixed time shift amount (for example, 128 samples). ,
The time signal x ′ [t] multiplied by a Hamming window or the like is output to the frequency analyzer 103. In an analysis method such as FFT, since a blocked signal is analyzed as being repeated, the influence of signal discontinuity appearing at a block boundary causes
The spectrum analysis result may not be obtained correctly. Windowing is performed to reduce this effect. However, when the window is a rectangular window (an operation of simply blocking a signal into a certain length), the window multiplier 102 and the overlap adder 107 are unnecessary.

The frequency analyzer 103 inputs the time signal x '[t] obtained from the window multiplier 102 and outputs a power spectrum X [ω] (where ω represents a frequency) using a technique such as FFT. The phase spectrum Θ [ω] is calculated. The power spectrum X [ω] is calculated by the noise power spectrum estimator 104,
The phase spectrum Θ [ω] is output to the subtractor 105, and is output to the frequency synthesizer 106.

The noise power spectrum estimator 104 calculates the power spectrum X [ω] obtained from the frequency analyzer 103 as
It comprises a memory 111 that stores a certain period of time and a small level power spectrum extractor 112 that extracts a small level power spectrum of a small level for each frequency from the past power spectrum stored in the memory. In general, a voice signal is a non-stationary signal whose power spectrum shape changes, and includes a silent or weak energy section within a certain period (for example, 1 second). It is considered that spectral information can be obtained.

FIG. 5a is a noise-free speech waveform (speaking content is “far” about 0.5 seconds), and FIG. 5b is a waveform in which white noise is added to the signal of FIG. 5a. FIGS. 6a and 6b show the power spectra of FIGS. 5a and 5b, respectively.
Comparing FIGS. 6a and 6b, it can be seen that the noise spectrum is observed in the interval where the power of the voice signal is weak.

As an embodiment of the noise power spectrum estimator 104, the power spectrum of the past certain period is stored in the memory 111, and the small level power spectrum extractor 112 stores the past power spectrum stored in the memory 111. The noise power spectrum N [ω] is estimated by extracting a small-level power spectrum by a method such as extracting the minimum value for each frequency. The subtractor 105 subtracts the noise power spectrum N [ω] obtained from the noise power spectrum estimator 104 from the power spectrum X [ω] of the signal in which voice and noise are mixed, obtained from the frequency analyzer 103. , Voice power spectrum S [ω] (= X [ω] −
N [ω]) is calculated and output to the frequency synthesizer 106.

The frequency synthesizer 106 uses the audio power spectrum S [ω] obtained from the subtractor 105 and the frequency analyzer 10.
The phase spectrum Θ [ω] of the signal obtained by mixing the voice and noise obtained from 3 is input, the time signal waveform y ′ [t] is restored by a method such as inverse FFT, and the time signal waveform y ′ [t] is output to the overlap adder 107. The overlap adder 107 shifts the signal waveform y ′ [t] obtained from the frequency synthesizer 106 by an appropriate time shift amount (the same as the time shift amount of the window multiplier 102), and superimposes the added signal y.
[t] is output to the output terminal 108. Noise removal is executed by the above processing.

[Second Embodiment] FIG. 2 is a view for explaining a second embodiment of the noise removing apparatus of the present invention. 2, the same components as those in FIG. 1 are given the same numbers as in FIG. As different components, instead of the noise power spectrum estimator 104, a non-linear converter 201 for weighting a small value of the power spectrum, and a long-time power spectrum for estimating a long-time power spectrum from the non-linearly converted power spectrum. Calculator 202
And a non-linear inverse converter 203 that returns the long-term power spectrum to an exact value.

In the first embodiment, in estimating the noise power spectrum, a memory for storing the past power spectrum is required. The second embodiment is characterized in that the value of the small power spectrum is weighted and then estimated by the smoothing process, thereby significantly reducing the memory amount. The difference from the first embodiment is only the estimation portion of the noise power spectrum, and therefore the processing of this portion will be described.

The non-linear converter 201 is a frequency analyzer 10.
The power spectrum X [ω] obtained from No. 3 is input, and a signal L [ω] in which a small value of the power spectrum is weighted is output to the long-time power spectrum calculator 202.
Here, as an example of the non-linear processing, there is logarithmic conversion or the like. By weighting a small value, the power spectrum of the background noise, which has a weaker power than the voice and is constantly present, can be extracted.

The long-time power spectrum calculator 202 is
The signal L [ω] obtained by weighting the small value of the power spectrum obtained from the non-linear converter 201 is input, and the long-term smoothing process is performed to obtain the long-term power spectrum M.
[ω] is calculated and output to the non-linear inverse converter 203. Here, as an example of the smoothing process for a long time, it is possible to calculate for each frequency component by the integral process with leak as shown in the equation (1). In Expression (1), r1 is a coefficient that controls the degree of smoothing (for example, 0.9). In this formula, the time index T is attached and displayed. The unit of this time is a frame.

M _T [ω] = r1 × M _T-1 [ω] + (1-r1) × L _T [ω] ··· (1) The power spectrum of a voice signal is relatively large. The value of the power spectrum M [ω] for a long time does not increase for a signal that changes rapidly, but the power spectrum of a signal that constantly exists such as background noise can be extracted. The nonlinear inverse transformer 203 receives the long-time power spectrum M [ω] obtained from the long-term spectrum calculator 202, calculates the signal N [ω] that has been subjected to the inverse transform processing of the nonlinear transformer 201, and subtracts it. Output to 105. For example, when logarithmic transformation is performed as an example of non-linear processing, the non-linear inverse transformation processing is exponential transformation. With the above processing, the estimated value of the noise power spectrum N [ω] is obtained.

FIG. 7 shows the noise power spectrum N [ω] obtained by this processing for the signal shown in FIG. 5b in which white noise is added. When compared with FIG. 6b, it can be seen that the fluctuation of the power spectrum due to the voice is smoothed and a value close to the noise power spectrum is extracted. In order to display the result in an easy-to-see manner, FIG. 8 shows the spectrum of the processing result for the signal in the vicinity of 5.78 seconds in FIG.

Plot 1 in FIG. 8 is a power spectrum of noise, plot 2 is a power spectrum of a signal in which voice and background noise are added, and plot 3 is a long-time power spectrum calculated by the above processing. It can be seen that the noise spectrum is being estimated. With this configuration, estimation of the noise power spectrum can be realized with a small amount of memory.

[Third Embodiment] FIG. 3 is a diagram for explaining a third embodiment of the noise removing apparatus of the present invention. 3, the same components as those in FIGS. 1 and 2 are designated by the same reference numerals as those in FIGS. A different component is a variance estimator 30 that estimates the variance of the power spectrum that has been nonlinearly transformed.
1, and the noise removal amount calculator 302 that calculates the noise removal amount from the long-term power spectrum and the variance of the power spectrum.
Equipped with.

In the first and second embodiments, the small value of the power spectrum is weighted to estimate the noise power spectrum. Noise having a stationary power spectrum such as power supply noise can be removed by the above processing, but the amount of removal is not sufficient for noise with varying power spectrum. The third embodiment is characterized in that the noise removal amount is controlled by investigating the stationarity of the noise power spectrum.

Portions different from the second embodiment will be described. The signal L [ω] obtained by the nonlinear converter 201 is output to the long-time power spectrum calculator 202 and the variance estimator 301.
The output of the long-term power spectrum calculator 202 is output to the variance estimator 301 and the noise removal amount calculator 302. The dispersion estimator 301 uses the long-time power spectrum calculator 20.
The long-time power spectrum M [ω] obtained from 2 and the non-linearly transformed power spectrum L [ω] obtained from the non-linear converter 201 are input, and the variance V [ω] of the power spectrum is estimated.

Here, as an example of the variance estimation process, it is possible to calculate each frequency component by the leaky integration process as shown in equation (2). In Expression (2), r2 is a coefficient that controls the degree of smoothing (for example, 0.9). This formula (2)
In the figure, the time index T is attached and displayed. The unit of this time is a frame.

V _T [ω] = r2 × V _T-1 [ω] + (1-r2) × (L _T [ω] −M _T [ω]) × (L _T [ω] −M _T [ω ]) (2) The noise removal amount calculator 302 uses the long-time power spectrum M [ω] obtained from the long-time power spectrum calculator 202,
Variance V of power spectrum obtained from variance estimator 301
With [ω] as an input, the noise removal amount T [ω] is calculated and output to the non-linear inverse converter 203.

Here, as an example of the noise removal amount estimation processing, as shown in equation (3) for each frequency component, the noise power spectrum is estimated larger as the variance of the power spectrum is larger, and conversely the variance is smaller. A process of setting the power spectrum value closer to the value for a long time can be considered. In Expression (3), K is a constant (for example, 2). If K is set to a negative value here, stationary power spectrum components are extracted as noise power spectra and noise is removed. Conversely, non-stationary (large variance) power spectrum components are calculated as small values as noise power spectra. Therefore, the removal amount can be controlled to be small.

T [ω] = M [ω] + K × Sqrt (V [ω]) (3) With this configuration, the noise removal amount is determined according to the stationarity of the noise power spectrum. Can be controlled.

[Fourth Embodiment] FIG. 4 is a view for explaining a fourth embodiment of the noise removing apparatus of the present invention. 4, the same components as those in FIG. 1 are designated by the same reference numerals as those in FIG. The different components are a voice detector 401 that inputs a power spectrum and a noise power spectrum to detect voice, an attenuation controller 402 that controls the attenuation of output voice from the detection result of the voice detector, and an attenuation controller. According to the attenuation amount obtained from the controller 402, the overlap adder 107
An attenuator 403 for attenuating the audio signal obtained from

In the above description, the noise removal is performed by a process such as a spectral difference method, but the entire power of the output signal can be attenuated with respect to the signal only in the noise section where no voice exists. As a method of voice detection, there has been a method of using power information in the related art, but since the input power spectrum and the noise power spectrum are calculated by the spectrum difference method, voice can be detected from these information.

The voice detector 401 inputs the input power spectrum X [ω] and the noise power spectrum N [ω], and outputs the detection result to the attenuation amount controller 402. As an example of the detection method, the input power spectrum X for each frequency component
There is a method in which the noise power spectrum N [ω] is subtracted from [ω] and if the subtraction result has a frequency component larger than a certain value, it is determined that the voice is detected. Further, as described in the third embodiment, it is possible to consider the fluctuation of the input power spectrum in the estimation of the noise power spectrum.

The attenuation controller 402 is a voice detector 401.
Input the output result of, and when the voice is detected, the attenuation ratio g
When [t] is set to 1 and no voice is detected (only noise), the attenuation ratio g [t] is set to a small value and the attenuator 403 is set.
Output to. Here, when the damping ratio is controlled smoothly over time, the discontinuity in the output waveform is reduced, so control such as that shown in equation (4) is conceivable. In the equation (4), r3 is a coefficient for controlling the degree of smoothness (for example, 0.
9).

G [t] = 1: When voice is detected g [t] = r3 × g [t−1]: When voice is not detected (4) The attenuator 403 attenuates The attenuation ratio g [t] obtained from the quantity controller 402 is input, the audio signal obtained from the overlap adder 107 is subjected to processing such as reducing the amplitude by the attenuation ratio, and output to the output terminal 108. When performing attenuation here, it is possible to use the input of the attenuator directly as an input signal instead of the signal after the spectral difference processing.

With this configuration, the power of the output signal can be reduced as a whole in the noise-free section where there is no voice. It should be noted that although all of the above processes have been described based on the spectral difference method, they can be similarly realized by band splitting process using a filter bank.

[Fifth Embodiment] FIG. 9 shows a fifth embodiment in which the noise removing device of the present invention is incorporated in a voice encoding / decoding device.
Indicates. The speech coding / decoding apparatus shown in FIG. 9 has a speech input terminal 901 for inputting speech, a noise eliminator 902 for eliminating noise of an input signal, a speech coder 903 for encoding a noise-removed signal, and a speech code. A code output terminal 904 that outputs the code from the coder 903, a medium 911 that transmits or stores the output code of the voice coder 903, a code input terminal 921 that inputs the code, and a voice signal that inputs the code. The audio decoder 922 for decoding, the noise removing device 923 for removing noise included in the decoded audio signal, and the audio output terminal 924 for outputting the audio signal are included.

In the speech coding / decoding apparatus of FIG. 9, the noise removing apparatus is incorporated in both encoding / decoding, but it is possible to incorporate it in either one or both.
The operation of the embodiment in which the present technology is applied to voice communication and voice storage will be briefly described. The noise removing device 902 on the encoding side is
Noise included in a signal input through a microphone or the like is removed, and the speech is encoded by the speech encoder 903. The coded data is output to the transmission or storage medium 911.

The decoding side inputs the coded data in the medium 911 from the code input terminal 921, and the speech decoder 9
At 22, it is decoded into an audio signal. Noise included in the decoded voice is removed by the noise remover 923, output to the voice output terminal 924, and reproduced from a speaker or the like. As an effect of the noise removing device 923 on the receiving side, for example, when used in voice communication or the like, even if there is no noise removing device on the transmitting (coding) side, the noise removing device 923 on the receiving side can remove noise. it can. In the present embodiment, it has been shown that the speech encoding / decoding device can execute the noise removal processing in each of the encoding processing and the decoding processing.

[0039]

According to the present invention, the noise power spectrum can be estimated by a simple process without detecting the noise section, and the stationary noise contained in the input signal can be removed.

[Brief description of drawings]

FIG. 1 is an explanatory diagram of a first embodiment of a noise eliminator of the present invention.

FIG. 2 is an explanatory diagram of a second embodiment of a noise removing device of the present invention.

FIG. 3 is an explanatory diagram of a third embodiment of the noise eliminator of the present invention.

FIG. 4 is an explanatory diagram of a fourth embodiment of the noise eliminator of the present invention.

FIG. 5 is a diagram showing an example of a voice signal waveform and a voice signal waveform with noise added.

FIG. 6 is a diagram showing an example of a power spectrum of a voice signal and a power spectrum of a voice signal added with noise.

FIG. 7 is a diagram showing an example of a noise power spectrum obtained by the processing of the second embodiment.

FIG. 8 is a diagram showing a state in which the noise power spectrum obtained by the processing of the second embodiment is superimposed on the input spectrum and the noise spectrum.

FIG. 9 is an explanatory diagram of a fifth embodiment of a voice encoding / decoding device including a noise removing device.

[Explanation of symbols]

101 input terminal 102 window mount 103 frequency analyzer 104 noise power spectrum estimator 105 Subtractor 106 frequency synthesizer 107 Overlap adder 108 output terminals 111 memory 112 Small level power spectrum extractor 201 Non-linear converter 202 Long-time power spectrum calculator 203 Non-linear inverse converter 301 variance estimator 302 noise removal amount calculator 401 voice detector 402 Attenuation controller 403 attenuator 901 Audio input terminal 902,923 noise elimination device 903 speech encoder 904 code output terminal 911 medium 921 code input terminal 922 speech decoder 924 audio output terminal

─────────────────────────────────────────────────── ─── Continuation of the front page (56) Reference JP-A-9-34497 (JP, A) JP-A-57-1444597 (JP, A) JP-A-3-180900 (JP, A) JP-A-8- 221092 (JP, A) JP 10-254499 (JP, A) JP 59-67732 (JP, A) JP2-3520 (JP, B2) JP146080 (JP, B2) Japanese Patent Publication No. 63-34477 (JP, B2) (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 21/02 G10L 15/20

Claims (3)

(57) [Claims] 1. Calculation of frequency spectrum information of an input signal
Frequency analyzer and the noise component contained in the input signal
Noise power spectrum to estimate the noise power spectrum of
And the noise power spectrum from the input power spectrum.
The subtractor for subtracting the cutout and the phase spectrum of the above input signal.
Tor and the power spectrum obtained by the above subtractor
A noise eliminator including a frequency synthesizer for restoring a signal
Where the noise power spectrum estimator is
A small level power spectrum of a small level within a certain period,
It came to be extracted as the above noise power spectrum
Cage, the noise power spectrum estimator, a nonlinear converter for weighting the small value of the power spectrum, the conversion power spectrum is non-linear conversion, and a long long power spectrum calculator for calculating a power spectrum, the A non-linear inverse converter that returns a long-time power spectrum to a true power spectrum, and a noise eliminator. 2. Calculation of frequency spectrum information of an input signal
Frequency analyzer and the noise component contained in the input signal
Noise power spectrum to estimate the noise power spectrum of
And the noise power spectrum from the input power spectrum.
The subtractor for subtracting the cutout and the phase spectrum of the above input signal.
Tor and the power spectrum obtained by the above subtractor
A noise eliminator including a frequency synthesizer for restoring a signal
Where the noise power spectrum estimator is
A small level power spectrum of a small level within a certain period,
It came to be extracted as the above noise power spectrum
Cage, the noise power spectrum estimator, a nonlinear converter for weighting the small value of the input power spectrum, and a long power spectrum calculator for calculating a power spectrum long from the conversion power spectrum is non-linear conversion, said nonlinear A variance estimator that estimates the variance of the transformed power spectrum, a noise removal amount calculator that calculates the noise removal amount from the long-time power spectrum and the variance of the transformed power spectrum, and the estimated noise removal amount A noise elimination device comprising: an exponential calculator for returning a power spectrum of a true number. 3. The noise eliminator according to claim 1 or 2, wherein from the input power spectrum and the noise power spectrum,
Further comprising a voice detector for the audio detected by the comparison of the power spectrum information, and attenuation controller to control the amount of attenuating the audio signal according to the determination result, an attenuator for attenuating the amplitude of the audio signal, the A noise eliminator characterized by.