JP4757775B2 - Noise suppressor - Google Patents

Description

  The present invention relates to noise suppression for the purpose of high-quality sound collection of speech, and more particularly to a noise suppression device that suppresses noise other than speech signals from an acoustic signal in which stationary noise and speech are mixed.

  As a technique related to this type of noise suppression, a spectral subtraction method (hereinafter referred to as SS method) is well known. In the SS method, an acoustic signal input to a microphone is subjected to Fourier transform to convert it into frequency domain amplitude information (hereinafter referred to as spectrum) and phase information. Then, by using this spectrum information, noise is suppressed by subtracting the spectrum of only the estimated noise signal from the spectrum of the acoustic signal in which noise and voice are mixed.

  As a method for estimating a noise signal by the SS method, several methods have been proposed so far. For example, the technique described in Patent Document 1 obtains an S / N ratio at a certain time called a frame and obtains a speech interval and a noise interval. The noise interval is estimated and the spectrum in that interval is used as the noise spectrum.

  In the technique described in Patent Document 2, in the noise section, the spectrum in the previous frame is set as the noise spectrum. In the voice section, only the speech component is extracted from the input signal and the output signal and the estimated noise component, and the difference from the input signal is calculated. In this way, only the noise signal is obtained, and the accuracy of noise estimation is improved by updating with the component.

  The SS method is a method of subtracting the noise component from the input signal, but this can also be regarded as reducing noise by creating and adding noise from the input signal. New noise called musical noise caused by the added noise also becomes a problem. Further, the SS method uses Fourier transform in order to convert a time domain acoustic signal that can be picked up by a microphone into a frequency domain. As the Fourier transform processing method, Fast Fourier Transform (FFT) is well known, and although the amount of computation of FFT itself is not so large, since the signal in the frequency domain is a complex number, computation is performed in comparison with real number computation. In many cases, complex arithmetic operations are a problem when speech is processed in real time.

  In order to solve the increase in the scale of the system and the increase in the amount of calculation, band division is performed using a filter bank using a well-known general single sideband (hereinafter referred to as SSB) modulation, and the theory of Wiener filter is used. There is a known technique for performing noise suppression (see, for example, Patent Document 3). In this method, it is assumed that exponential smoothing of noise estimation and filter coefficient update is effective, thereby eliminating the reduction of musical noise.

  Further, since the SS method performs non-linear processing in the frequency domain, the sound is inevitably distorted. Therefore, the attenuation coefficient in the frequency domain is smoothed by weighted averaging of the attenuation coefficients of an arbitrary number of bands around the band for correcting the attenuation coefficient (see, for example, Patent Document 4). ).

Japanese Patent Laid-Open No. 10-097288 (page 3 to page 4, FIG. 1) Japanese Patent No. 3270480 (pages 4-5, FIG. 2) Japanese translation of PCT publication No. 2004-502977 (page 7-10, FIG. 1) Japanese Patent Laying-Open No. 2005-348173 (pages 6-7, FIG. 1)

  However, in the technique described in Patent Document 1, noise is not estimated in a speech section, and if there is a change in noise during that time, the speech is distorted. In addition, when an error occurs in the estimation of the noise interval, an error occurs in the estimated noise spectrum due to speech, which also causes speech distortion. As a result, there is a problem that the sound collection signal is accompanied by noise and sound distortion, and high-quality sound collection is not realized.

  Further, in the technique described in Patent Document 2, the noise component is estimated using the output signal of the previous frame, and if the audio signal is a signal having a band of about 3.4 kHz (the audio band used for a telephone), Although it does not become a big problem, there is a problem that a large error occurs in the estimation when dealing with a voice of 7 kHz or higher which is a wide band (voice conference, web conference, etc.).

  Further, the technique described in Patent Document 3 has a problem in that noise is estimated by using an average value to the extent that the sound to be collected is distorted.

  Further, in the technique described in Patent Document 4, although the continuity between the bands is improved by leveling the attenuation coefficient of each band, it interferes with noise suppression and causes a new distortion in the voice.

  Therefore, an object of the present invention is to propose a noise estimation method and a filter coefficient correction method in a noise suppression method using a band-division filter bank, and does not include complex number calculation in the scale-up or calculation of the device, Another object of the present invention is to provide a noise suppression device that realizes noise suppression with little distortion.

The noise suppressor of the present invention is a noise suppressor that employs a spectral subtraction method. A signal in a frequency band in which a time domain input signal in which a speech signal and a stationary noise signal are mixed is limited by using SSB modulation. In each of the processing units, the frequency band dividing means (20 in FIG. 1), the frequency band processing unit (30 in FIG. 1) that performs processing to suppress noise in the divided frequency band input signal, and 1 is composed of a band synthesizer (40 in FIG. 1) that outputs one signal in which noise is suppressed by synthesizing the processed signals, and each processing unit (30 in FIG. 1) receives the input by leak integration. Amplitude estimation means (1 and 2 in FIG. 1) for smoothing on the time axis by suppressing a difference in attenuation coefficient between signal frames and estimating the current input signal amplitude value and noise amplitude value, and noise amplitude estimation Value input signal amplitude An attenuation coefficient determining means (5, 6 in FIG. 1 ) that obtains a value γ divided by the estimated value and uses (1-γ) as an attenuation coefficient, and the attenuation coefficient determined by the attenuation coefficient determining means on the time axis and in the band There is provided attenuation coefficient correction means (7 and 8 in FIG. 1) for performing correction by smoothing between them, and multiplication means (9 in FIG. 1) for multiplying the input signal by the attenuation coefficient obtained from the attenuation coefficient correction means. Then, the amplitude estimating means (2 in FIG. 1) is characterized in that noise estimation is not stopped even in the speech section by multiplying the term of the input signal amplitude estimated value in the leakage integral expression by the value γ .

  Further, the noise suppression apparatus of the present invention compares the input signal amplitude estimated value obtained by the amplitude estimating means (1, 2 in FIG. 1) with the noise amplitude estimated value to determine the noise interval, and at this time, the noise amplitude The estimated value is multiplied by a coefficient of about 2, and the noise amplitude obtained by the noise estimation unit (3 in FIG. 1) that outputs a noise bias value of about 0.5 in the noise interval and about 1.0 in the speech interval, and the amplitude estimation unit The estimated value and the input signal amplitude estimated value are multiplied by a noise bias value and compared. If the input signal amplitude estimated value is small, a maximum attenuation coefficient flag that maximizes the attenuation coefficient of the band is set as the attenuation coefficient determining means (FIG. 1). 5 and 6) may be provided with noise amplitude comparison means (4 in FIG. 1).

More specifically, the attenuation coefficient determining means (5, 6 in FIG. 1) calculates the value γ, but if the input signal amplitude estimation value is smaller than the noise amplitude estimation value , the signal / noise ratio calculation is made with the value γ being 1.0. Means (5 in FIG. 1), and when the maximum attenuation coefficient flag has been passed, attenuation coefficient calculation means (6 in FIG. 1) outputs the attenuation coefficient to the attenuation coefficient correction means as 0 . It is comprised by these.

Attenuation coefficient correction means (7, 8 in FIG. 1) includes attenuation coefficient leveling means (7 in FIG. 1) that smoothes the attenuation coefficient from the attenuation coefficient determination means on the time axis by leak integration, and in the frequency band. With respect to the attenuation coefficient, an inter-band attenuation coefficient leveling means (8 in FIG. 1) that examines the attenuation coefficient of adjacent bands and corrects only in the direction of decreasing the attenuation coefficient so that the ratio with the attenuation coefficient does not exceed a certain value. Composed.

  In the present invention, the increase in the amount of calculation due to complex number calculation is first solved by band division using SSB modulation. With regard to noise estimation errors, noise estimation is continued and suppressed even during the speech period by using noise estimation using the ratio of noise to the current amplitude value (hereinafter referred to as SNR). By using the noise amplitude spectrum and the input signal spectrum obtained as described above, the noise is reduced by using the Wiener filter theory. At this time, the nonlinearity in the time domain is reduced by smoothing the attenuation coefficient obtained by the Wiener filter theory in the time domain, and the attenuation coefficient in each band is corrected based on the attenuation coefficient between adjacent bands. It reduces nonlinearity in the frequency domain and suppresses audio distortion. Furthermore, with respect to musical noise, the noise interval is estimated, and the noise is apparently increased in the noise interval, and the attenuation coefficient is increased.

  According to the present invention, it is possible to perform noise suppression with a smaller amount of calculation than that of the conventional method and less distortion in the output speech. This is because the SNR is used for noise estimation, the attenuation coefficient of each band is smoothed on the time axis, and the attenuation coefficient of each band is smoothed based on the attenuation coefficient of the adjacent band.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[Description of configuration]
FIG. 1 shows the configuration and processing flow of the noise suppression apparatus of the present invention. This noise suppression apparatus includes a band dividing unit 20 that divides an input signal input from a microphone 10 into a signal limited to a frequency band, a processing unit 30 that has a one-to-one correspondence with the divided signal, and each processing unit. And a band synthesizing unit 40 that synthesizes the signal processed in 30.

  An audio signal and a noise signal are mixed in the input signal picked up by the microphone 10. Therefore, the band dividing unit 20 divides the input signal into bands, each processing unit 30 performs a process of suppressing noise in the input signal in the frequency band, and the band synthesizing unit 40 synthesizes the signals of each band. , Output a noise-suppressed signal.

  The band dividing unit 20 uses a filter bank using general SSB (Single Side Band) modulation. If a Fourier transform method such as FFT is used for the band dividing unit 20, the signal in the band division becomes a complex number, which increases the amount of calculation. Therefore, band division using SSB modulation is performed in order to suppress such a wasteful increase in calculation amount.

  This is a specific flow of the band dividing unit 20. First, the input signal is cut out with a frame length depending on the number of band divisions. This depends on the division method performed in the band dividing unit 20, but in the band division using the SSB modulation used in the present invention, for example, when it is desired to divide into 16 bands, the frame is about 10 samples. Process. The number of divisions should be changed according to the sampling frequency. For reference, 32 divisions are sufficient when dealing with 16 kHz sampled audio. In this case, the frame length is about 20 samples.

  The signal cut out by the frame length is input to a filter bank using SSB modulation and becomes a set of signals of one sample limited to each frequency band. Since the signals in the respective frequency bands are processed almost independently, the processing unit 30 is represented by a layer structure in FIG.

  Each processing unit 30 includes an input signal amplitude estimating unit 1, a noise signal amplitude estimating unit 2, a noise interval estimating unit 3, a noise amplitude comparing unit 4, a signal / noise ratio calculating unit 5, an attenuation coefficient calculating unit 6, an attenuation coefficient smoothing unit. 7 and an inter-band attenuation coefficient smoothing unit 8 and a multiplier 9. Each part 1-8 calculates | requires the attenuation coefficient with respect to an input signal rather than analyzing the input signal from the band division part 20. FIG. The multiplier 9 multiplies the input signal from the band dividing unit 20 by the attenuation coefficient obtained by each unit 1 to 8 and outputs the result to the band synthesizing unit 40.

  The input signal amplitude estimator 1 estimates the amplitude of the input signal by a process called leak integration. Since the SS method performs nonlinear processing in the frequency domain, inevitably distortion occurs in the signal. As a countermeasure against the distortion of the signal, smoothing is performed on the time axis by suppressing the difference in attenuation coefficient between frames of the input signal, and the amplitude value (average value) of the current signal is estimated.

  Similar to the input signal amplitude estimation unit 1, the noise signal amplitude estimation unit 2 estimates the noise amplitude value using leak integration. For the estimation here, the output of the signal / noise ratio calculation unit 5 in the subsequent stage is used, and the equation of the leakage integral is modified by the reciprocal γ (= An / As) of SNR (SignalNoiseRatio: Here As / An). Therefore, it is possible to accurately estimate noise even in a voice section without requiring discrimination between a voice section and a noise section.

  The noise interval estimation unit 3 inputs the estimated amplitude values from the input signal amplitude estimation unit 1 and the noise signal amplitude estimation unit 2 and compares them to estimate the noise interval. In this case, since the estimated noise amplitude value is an average value, multiplying a coefficient of about 2 halves the false estimation of a noise interval as a speech interval. For this reason, the noise section estimation unit 3 passes a noise bias value of about 0.5 in the noise section and about 1.0 in the voice section to the noise amplitude comparison unit 4 in the subsequent stage.

  The noise amplitude comparison unit 4 inputs and compares the respective estimated amplitude values from the input signal amplitude estimation unit 1 and the noise signal amplitude estimation unit 2, and if the input signal amplitude estimation value is small, the attenuation coefficient of the band is maximized. Is passed to the attenuation coefficient calculation unit 6. At this time, the estimated value is compared with the noise bias value from the noise interval estimation unit 3 for comparison. This is to effectively perform noise suppression in the noise interval in response to the noise interval estimation unit 3 multiplying the noise amplitude estimation value by a coefficient.

  The signal / noise ratio calculation unit 5 receives the estimated amplitude values of the input signal amplitude estimation unit 1 and the noise signal amplitude estimation unit 2 and divides the noise amplitude estimation value by the input signal amplitude estimation value γ. = (An / As) is calculated. If the input signal amplitude estimation value is smaller than the noise amplitude estimation value, it is passed to the noise signal amplitude estimation unit 2 and the attenuation coefficient calculation unit 6 as γ = 1.0. The usage of γ in the noise signal amplitude estimation unit 2 is as described above.

  The attenuation coefficient calculation unit 6 calculates the attenuation coefficient L from the inverse SNR γ = (An / As) from the signal / noise ratio calculation unit 5 and the maximum attenuation coefficient flag from the noise amplitude comparison unit 4, and smoothes the attenuation coefficient. Pass to part 7. The attenuation coefficient smoothing unit 7 further smoothes the attenuation coefficient L from the attenuation coefficient calculation unit 6 on the time axis by leak integration, obtains a final attenuation coefficient SL, and passes it to the interband attenuation coefficient smoothing unit 8. .

  The inter-band attenuation coefficient smoothing unit 8 checks the attenuation coefficient SL of an adjacent band (hereinafter referred to as an adjacent band) for the attenuation coefficient SL in the frequency band, and compares the attenuation coefficient SL with the adjacent band (hereinafter referred to as MD). (Notation) is corrected so that it does not exceed a certain level. Correction is performed only in the direction of decreasing the attenuation coefficient. Thereby, the attenuation coefficient between adjacent bands is smoothly connected, and the distortion of the voice can be largely eliminated. For this purpose, although difficult to draw in FIG. 1, the attenuation coefficient SL is inputted to the inter-band attenuation coefficient smoothing unit 8 from the attenuation coefficient smoothing unit 7 of the processing unit 30 corresponding to the adjacent band.

[Description of operation]
Next, the operation of the present noise suppression apparatus configured as described above will be described in detail with reference to FIGS.

The signal of each frequency band divided by the band dividing unit 20 is input to the input signal amplitude estimating unit 1 of the processing unit 30. In the input signal amplitude estimation unit 1, amplitude is estimated by a process called leak integration. The leak integral is expressed by the following equation.
As (t) = δ × | S | + (1-δ) × As (t-1) (1)
Here, t represents the sample time, and (t-1) represents the time of one sample in the past. S is an input signal in which voice and noise input to the microphone are mixed. As represents the amplitude estimate of the input signal. δ is a parameter for controlling the influence of the instantaneous value on the estimated value, and is a value of 1 or less. If the parameter δ is decreased, the amplitude estimated value As is approximated to the average value of the input signal, and if it is increased, it is close to the instantaneous value of the input signal.

  In order to suppress noise by the Wiener filter theory, there is no problem if the instantaneous value of the input signal is used as the amplitude value of the input signal. If the instantaneous value of the input signal is to be used, the parameter δ may be set to 1. However, since the difference in attenuation coefficient between frames can be suppressed by setting δ to a value of about 0.5 to 0.25, smoothing is performed on the time axis and nonlinearity can be suppressed. As a result, distortion can be reduced. In the present invention, δ = 0.5 to 0.25 is recommended.

  Next, the output of the input signal amplitude estimation unit 1 is input to the noise signal amplitude estimation unit 2. Similar to the input signal amplitude estimation unit 1, the noise signal amplitude estimation unit 2 estimates the average value of noise using leak integration. In this case, since it is not necessary to follow the instantaneous value, the value of δ is made extremely small, such as 0.0001. Since the output As of the input signal amplitude estimation unit 1 that is an input to the noise signal amplitude estimation unit 2 is smoothed in advance by the input signal amplitude estimation unit 1, an improvement in noise estimation accuracy can be expected. However, the output of the input signal amplitude estimation unit 1 includes a voice component in addition to the noise to be suppressed.

In many conventional techniques, the speech component is eliminated by dividing the speech segment and the noise segment. For example, in the technique described in Patent Document 1, noise estimation is performed using
An (t) = δ × As + (1-δ) × An (t-1) (2)
δ = α when As / An ≤ TH
δ = 0 when As / An> TH
0 <α <1
(Variable names are the same as those of the present invention.) The noise is estimated only where the SNR is bad (TH is a threshold value and a constant value), that is, in the noise interval. In such an estimation method, the noise amplitude is not estimated in the speech section, the change in the noise cannot be followed, the speech is distorted, and the noise suppression effect is diminished.

In the technique described in Patent Document 2, noise estimation is performed using
An (t) = δ × As + (1-δ) × An (t-1) when ΣAs ≤ TH
An (t) = δ × (As- (η × Ao (t-1) + (1-η) × (As-An (t-1)) + (1-δ) × An (t-1) when
ΣAs> TH (3)
0.5 <η <1
(Variable names match those of the present invention.) Here, Ao represents an amplitude estimation value of the output signal. In this method, in the non-noise section, the noise component of one sample past is subtracted from the output signal (because it is the result of noise suppression processing) where only speech of one sample past exists. Considering a signal that will only be left speech, only the noise component in the input signal is extracted and noise estimation is performed, but there are many assumptions. For example, if only speech can be extracted with the expression (As-An (t-1)), noise suppression is possible only with this term. In practice, this is difficult, so an additional function is used, and it is questionable whether (As-An (t-1)) can extract only sound. In addition, the calculation is complicated.

Therefore, in the present invention, the noise is estimated as follows using the output of the signal / noise ratio calculator 5 in the subsequent stage. Specifically, the leak integral equation,
An (t) = δ × As + (1-δ) × An (t-1) (4)
Parameter γ is added to
An (t) = δ × γ × As + (1-δ) × An (t-1) (5)
And deformed. Here, An is an estimated value of noise amplitude. γ will be described in detail in the description of the signal / noise ratio calculation unit 5, but simply speaking, it is the reciprocal of SNR (SignalNoiseRatio: Here, As / An). By modifying the leakage integral equation as shown in Equation (5), learning can be performed with the noise signal itself in the noise interval, and the amplitude value of the noise signal included in As is estimated and learned in the speech interval. It becomes possible.

  Here, it is assumed that the SNR of one sample in the past is equal to the current SNR, but the time of one sample is, for example, 0.0025 seconds at 400 Hz (16 kHz is thinned out by 40 samples), and the SNR of this time interval is Since the change is very small, it can be said that the assumption that the SNR of one sample is equal to the current SNR is reasonable. As a result, it is not necessary to discriminate between speech and noise sections, and noise can be accurately estimated even in the speech sections.

  The estimated amplitude values of the input signal amplitude estimation unit 1 and the noise signal amplitude estimation unit 2 obtained from the above are input to the noise interval estimation unit 3. Here, the estimated value of the input signal amplitude and the estimated value of the noise amplitude are compared to estimate the noise interval. However, the noise amplitude estimation value merely indicates the average value of the noise amplitude, and the noise section cannot be estimated completely by simply comparing. This is because noise involves dispersion.

  As an example showing the variance of noise, the amplitude distribution of background noise in a room is shown in FIG. This histogram measures the background noise in a room, normalizes the maximum absolute value of the amplitude as 1, and sets the distribution as a histogram. The average value of the noise amplitude is indicated by a white dotted line. Comparing the average value of noise and the estimated amplitude value (substantially instantaneous value) of the input signal, nearly half of the input signals show an amplitude value higher than the average value of the noise amplitude. That is, if the noise amplitude estimation value and the input signal amplitude estimation value are simply compared to determine the noise interval, half of the noise estimation values are incorrect.

  Therefore, the noise interval estimation unit 3 reduces the estimation error by multiplying the noise amplitude estimation value by a coefficient and comparing the coefficients. If this coefficient is large, the noise interval can be reliably estimated as the noise interval, but if it is too large, it may be erroneously estimated as the noise interval even though there is speech. Therefore, the coefficient is about 2. By multiplying the noise amplitude estimate by this coefficient, erroneous estimation of a noise interval as a speech interval is halved.

  In the noise section estimated as described above, the noise bias value passed to the subsequent noise amplitude comparison unit 4 is reduced. This noise bias value is about 0.5 in the noise section and about 1.0 in the voice section. The usage of this value will be described later.

  In response to the noise bias value from the noise interval estimation unit 3, the noise amplitude comparison unit 4 compares the amplitude estimation value of the input signal with the noise amplitude estimation value. At this time, the input signal is compared with the noise bias value from the noise interval estimation unit 3 for comparison. This is to effectively perform noise suppression in the noise section. This will be described with reference to FIG.

  FIG. 3 simulates a change in the amplitude of an input signal with time. The shaded part of each bar graph is a noise component, and the white part represents an audio signal. Furthermore, the solid line is the average value of the estimated noise. If the average value of the noise is compared with the magnitude of the noise component, if the shaded bar is lower than the solid line, the noise can be completely suppressed, but if it is high, it cannot be suppressed. This is because the process using the Wiener filter is a method in which the average value of the noise amplitude is subtracted from the amplitude of the input signal, and noise that is larger than the average value cannot be completely suppressed.

  Therefore, when the noise section estimation unit 3 determines that it is a noise section, the average value of the noise amplitude is visually increased as shown in FIG. 4 and the Wiener filter processing is performed to completely suppress the noise. However, if processing is performed with a noise amplitude that is apparently increased in the voice section as well, the distortion of the voice increases. Therefore, the original average noise amplitude value as shown in FIG. 3 is used outside the noise section. In this case, the noise cannot be completely suppressed. However, since the noise is masked by the voice, the noise is hardly noticed in practice.

  The noise amplitude comparison unit 4 compares the estimated value of the input signal amplitude with the product of the noise bias value from the noise interval estimation unit 3 described above, and if the input signal amplitude value is small, the attenuation coefficient of that band. Is passed to the attenuation coefficient calculation unit 6. A method of using the maximum attenuation coefficient flag will be described later.

The signal / noise ratio calculation unit 5 divides the input signal amplitude estimated value As and the noise amplitude estimated value An to calculate a value of γ represented by the following equation.
γ = An / As
if An> As then γ = 1.0 (6)
It is. This γ is the reciprocal of SNR, and γ = 1.0 when the amplitude of the input signal is smaller than the estimated value of the noise amplitude. That is, if the input signal is larger than the noise amplitude estimation value, this value is small. This γ is used in the noise signal amplitude estimation unit 2.

The use of As for noise estimation is as described above, but in the non-noisy section, the accuracy of noise estimation is reduced by the speech component contained in As. Therefore, the estimation accuracy is increased by updating with γ × As. The first term δ × γ × As on the right side of Equation (5) is
δ × γ × As = δ × An / As × As ≒ δ × An (7)
And can be transformed. Therefore, the entire right side of the equation (5) can be updated with only the noise-free value. That is, it is possible to accurately estimate noise even in a non-noise section.

Next, an attenuation coefficient L, which is a main part of noise suppression, is obtained as an attenuation coefficient L. The attenuation coefficient is basically obtained by the Wiener filter formula, and when the obtained attenuation coefficient is L ′,
L '= (As-An) / As (8)
When this equation is expanded using γ in Equation (6),
L '= (As-An) / As = 1-γ (9)
It becomes. Here, when the maximum attenuation coefficient flag is transferred from the noise amplitude comparison unit 4, the value of L ′ is set to zero.

If the value of L ′ is used as it is as the attenuation coefficient L, the attenuation coefficient may become extremely large. For example, when γ = 0.9, L ′ = 0.1, and the input signal becomes 1/10. If the attenuation coefficient increases in this way, audio distortion is likely to occur. Therefore, the maximum attenuation coefficient ML is provided, and the attenuation coefficient L that is finally output is set.
L = L '
if ML> L 'then L = ML (10)
And The attenuation coefficient L obtained here is passed to the attenuation coefficient smoothing unit 7.

The attenuation coefficient L calculated by the above-described attenuation coefficient calculation unit 6 has been smoothed in the time axis direction by smoothing As of the input signal amplitude estimation unit 1, but it still means that the distortion of the speech is reduced. not enough. Therefore, the attenuation coefficient L is further smoothed on the time axis by leak integration. The attenuation coefficient smoothing unit 7 performs this.
SL (t) = δ × L + (1-δ) × SL (t-1) (11)
Expression (11) is an expression of the leakage coefficient leakage integral, and SL is the final attenuation coefficient. Here, it is assumed that δ is approximately 0.5 and it is easy to follow the instantaneous value of the attenuation coefficient. This is because if δ is too small, distortion occurs in the voice, which is a parameter for adjusting the trade-off relationship between the noise suppression performance and the voice distortion.

  As described above, the attenuation coefficient SL in each frequency band is obtained. Noise can be suppressed by multiplying the signal of each frequency band by this attenuation coefficient SL. However, when the attenuation coefficient SL is multiplied as it is, if the difference between the attenuation coefficients between the bands is very large, the distortion of the sound becomes very large. Therefore, in the present invention, an interband attenuation coefficient smoothing unit 8 is introduced.

  The function of the interband attenuation coefficient smoothing unit 8 will be described with reference to FIGS. First, FIG. 5 shows a model of an input signal in each frequency band at a certain time. Here, the horizontal axis of the graph is frequency, and the vertical axis is amplitude. Further, the hatched portion of the graph represents a voice component, and the white portion represents a noise component.

  Assuming that such an input signal is input and noise is accurately estimated, the attenuation coefficient obtained by the Wiener filter theory is as shown in FIG. The horizontal axis of this graph is the frequency as in FIG. 5, and corresponds to FIG. The vertical axis represents the calculated multiplication value for attenuation. It can be seen from FIG. 6 that there are combinations in which the difference in attenuation coefficient is large between adjacent bands. The purpose of the inter-band attenuation coefficient smoothing unit 8 is to correct this extremely large attenuation coefficient difference.

  As a technique for smoothing the attenuation coefficient between existing bands, a technique described in Patent Document 4 can be cited. In this document, smoothing is realized by weighting and averaging the attenuation coefficients of an arbitrary number of bands around the band for correcting the attenuation coefficient. When the process of using the average of the attenuation coefficients of the three adjacent bands as an example in Patent Document 4 as the attenuation coefficient is applied to the attenuation coefficient in FIG. 6, the attenuation coefficient is as shown in FIG.

  Certainly, looking at FIG. 7, the attenuation coefficient between adjacent bands is smoothed, and there is no extreme change. Therefore, the output of the input signal shown in FIG. 5 is calculated using the attenuation coefficient shown in FIG. The result is shown in FIG. In this figure, the audio component (shaded portion) included in FIG. 5 (input signal) and the noise (white portion) generated by the attenuation process are shown. Here, the point to be noted is a white noise component, and noise in a negative direction is generated depending on the band. That is, it can be seen that the input audio component is cut off and the audio is distorted. That is, in this smoothing method, although the continuity between the bands is improved by smoothing the attenuation coefficient of each band, noise is newly added to cause distortion and noise of the voice. With this, high-quality sound pickup cannot be expected.

  In the present invention, the inter-band attenuation coefficient smoothing unit 8 checks the attenuation coefficient of adjacent bands (hereinafter referred to as adjacent bands) of each band, and the ratio (hereinafter referred to as MD) with the attenuation coefficient of the adjacent bands is constant. It has a function of correcting so as not to be above. Correction is performed only in the direction of decreasing the attenuation coefficient. Thereby, the attenuation coefficient between adjacent bands is connected smoothly, and the distortion of the voice can be largely eliminated.

A specific correction flow will be described with reference to FIGS. FIG. 9 shows, in FIG. 6, a band in which the MD of the adjacent “high” frequency band is a certain level or more is black. In this case, the band to be corrected is a band drawn with polka dots, and the attenuation coefficient of this band is corrected to be equal to or less than a certain value. Specifically, for example, when the attenuation coefficient 0.2 (polka dots) and 0.8 (black) are next to each other,
MD = 0.2 / 0.8 = 0.25 (12)
It becomes. Here, when the minimum value of MD is set to 0.5, correction is performed because MD is 0.25. The correction is the attenuation coefficient of the band that was 0.2 attenuation coefficient,
0.2 → 0.8 * Minimum value of MD = 0.8 × 0.5 = 0.4 (13)
Correct as follows. Here, the point to be noted is that the attenuation coefficient in a band where the attenuation coefficient is large (the multiplication value is small) is corrected. Also, correction is performed from the attenuation coefficient in the highest frequency band. As a result, correction is possible even when correction is required continuously in two bands, such as the 2000 Hz and 2250 Hz bands in FIG. 9.

  Next, the correction in FIG. 10 will be described. FIG. 10 shows the result of performing the correction of FIG. 9 (correction focusing on adjacent high frequency bands). In contrast to FIG. 9, the color coding in FIG. 10 is compared with adjacent “low” frequency bands, and a band where MD is a certain value or more is expressed by polka dots. With respect to the band of the polka dots, the smoothed attenuation coefficient as shown in FIG. 11 can be obtained by performing the correction of the equations (12) and (13) which are the correction processes of FIG. FIG. 12 shows the result of processing the input signal with the attenuation coefficient of FIG. From this result, it can be seen that since there is no noise in the negative direction, the noise component can be suppressed without removing the voice component.

  However, a noise component that cannot be completely suppressed is observed in the lowest band. Regarding this noise, since the voice signal in the adjacent band is large, the noise is hardly noticed by the masking effect. The masking effect is one of human auditory characteristics, and when there is a large component in a certain frequency component, it is a phenomenon that it becomes difficult to hear the sound in the vicinity.

  Therefore, the minimum value of MD is preferably set to a limit value at which noise is masked by the sound of the adjacent band. This means that each frequency band is divided into critical band bands (general Bark Scale can be used) that serve as masking indices, and the minimum value of MD is determined. However, since the amount of calculation increases, It may be a constant value in the band. In this case, the minimum value of MD is set to about 0.5.

  As described above, the multiplier 9 multiplies the input signal of each band by the determined attenuation coefficient in each band. This signal is synthesized by the band synthesizing unit 40 to obtain a final output signal. The band synthesizing unit 40 can use a band synthesizing method using SSB modulation as in the band dividing unit 20.

  The above is the best embodiment of the present invention. Up to now, one configuration has been described as a model, but it is not limited to the description of parameters and the like, and can be changed within the scope of maintaining the gist.

The block diagram which shows the noise suppression apparatus of this invention Distribution diagram of absolute amplitude of background noise in a room Figure showing examples of noise average and noise / speech input models Figure showing an example of noise / speech input model when the average noise value is shown to be large The figure which shows the amplitude of the audio signal and noise signal of a certain band The figure which shows the attenuation coefficient calculated in the case of the input of FIG. The figure which shows the smoothing result of the attenuation coefficient by the technique of patent document 4 The figure which shows the amplitude of the output signal by the attenuation coefficient of FIG. The figure which shows the smoothing result of the attenuation coefficient by the high region comparison in this invention The figure which shows the smoothing result of the attenuation coefficient by the low region comparison in this invention The figure which shows the smoothing result of the high region comparison and low region comparison attenuation coefficient in this invention The figure which shows the amplitude of the output signal by the attenuation coefficient of FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Input signal amplitude estimation part 2 Noise signal amplitude estimation part 3 Noise area estimation part 4 Noise amplitude comparison part 5 Signal / noise ratio calculation part 6 Attenuation coefficient calculation part 7 Attenuation coefficient smoothing part 8 Interband attenuation coefficient smoothing part 9 Multiplier 10 Microphone 20 Band division unit 30 Processing unit 40 Band synthesis unit

Claims (4)

In the noise suppressor that uses the spectral subtraction method,
Band dividing means for dividing an input signal in a time domain in which an audio signal and a stationary noise signal are mixed into a signal in a frequency band limited by using SSB modulation;
A processing unit corresponding to a frequency band that performs processing to suppress noise in the input signal of the divided frequency band;
A band synthesis unit that outputs one signal in which noise is suppressed by synthesizing the signals processed by the respective processing units,
Each of the processing units
Amplitude estimation means for suppressing the difference in attenuation coefficient between the frames of the input signal by leak integration and smoothing on the time axis to estimate the current input signal amplitude value and noise amplitude value,
An attenuation coefficient determining means for obtaining a value γ obtained by dividing the noise amplitude estimated value by the input signal amplitude estimated value and using (1−γ) as an attenuation coefficient ;
Attenuation coefficient correction means for correcting the attenuation coefficient determined by the attenuation coefficient determination means by performing smoothing on the time axis and between bands ;
Multiplying means for multiplying the input signal by the attenuation coefficient obtained from the attenuation coefficient correction means;
The noise suppression device according to claim 1, wherein the amplitude estimation unit does not stop noise estimation even in a speech section by multiplying the value of the input signal amplitude estimation value in the leakage integral equation by the value γ . The noise amplitude estimation value is compared with the input signal amplitude estimation value obtained by the amplitude estimation means to determine the noise interval. At this time, the noise amplitude estimation value is multiplied by a coefficient of about 2, and the noise interval is about 0.5. Noise interval estimation means for outputting a noise bias value of about 1.0 in the speech interval;
The noise amplitude estimated value obtained by the amplitude estimating means and the input signal amplitude estimated value are multiplied by the noise bias value and compared, and if the input signal amplitude estimated value is small, the maximum attenuation coefficient of the band is maximized. 2. The noise suppression apparatus according to claim 1, further comprising noise amplitude comparison means for outputting an attenuation coefficient flag to the attenuation coefficient determination means. The attenuation coefficient determination means includes
While calculating the value gamma, if the input signal amplitude estimate is less than the noise amplitude estimate signal-to-noise ratio calculating means for 1.0 the gamma,
Wherein said While calculating the attenuation coefficient, if the maximum damping coefficient flag is passed is characterized in that it is composed of a damping coefficient calculation means for output to the damping coefficient correcting unit the attenuation coefficient as 0 Item 3. The noise suppression device according to Item 2 . The attenuation coefficient correction means includes
Attenuation coefficient leveling means for smoothing on the time axis by leak integration for the attenuation coefficient from the attenuation coefficient determining means ,
Concerning the attenuation coefficient in the frequency band, it is composed of inter-band attenuation coefficient leveling means that checks the attenuation coefficient of the adjacent band and corrects only in the direction of decreasing the attenuation coefficient so that the ratio with the attenuation coefficient does not exceed a certain value. The noise suppression device according to any one of claims 1 to 3, wherein

Images