JP4542790B2 - Noise suppressor and voice communication apparatus provided with noise suppressor - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To realize high-speed and highly accurate noise suppression processing by making it possible to estimate a noise suppression coefficient with high accuracy in a small amount of operation. <P>SOLUTION: A noise suppression coefficient calculating part 30 approximates occurrence probability for each of a real part and an imaginary part of voice spectrum by a Laplace distribution, and also approximates the occurrence probability for each of a real part and an imaginary part of noise spectrum by a Gauss distribution, and further prepares voice spectrum estimation formula derived by using MAP estimation and Bayes' theorem, and is made to calculate a noise suppression coefficient according to this estimation formula. <P>COPYRIGHT: (C)2005,JPO&amp;NCIPI

Description

  The present invention relates to a noise suppressor provided to suppress a noise component contained in an input audio signal, and an audio communication apparatus such as a fixed telephone, a mobile phone, an IP telephone, and a video conference apparatus provided with the noise suppressor.

  Voice communication systems such as CELP (Code Excited Linear Prediction) systems are used in voice communication apparatuses such as mobile phones. When this type of device is used in an environment with a large amount of background noise, the background noise is captured and encoded together with the speech, resulting in a decrease in the clarity of the speech. For this reason, various techniques (noise suppressors) for performing speech encoding by removing or suppressing background noise and approaching a speech-only signal have been studied.

  For example, as a first prior art, a signal-to-noise ratio of an input signal is obtained for each frequency band, a noise suppression coefficient is determined based on the signal-to-noise ratio, and this coefficient is applied to the input signal for each frequency band. Therefore, a method for suppressing noise is known (for example, see Non-Patent Document 1). According to this method, it is possible to obtain an audio signal in which noise is suppressed. However, since the noise suppression coefficient is calculated using the modified Bessel function, there is a problem that the amount of calculation becomes extremely large.

  In addition, as a second prior art, the amplitude of a speech frequency spectrum is approximated by a gamma distribution and a noise suppression coefficient estimation formula is derived using MAP estimation and Bayes' theorem, and noise suppression is performed using this estimation formula. There is a method for calculating a coefficient (see, for example, Non-Patent Document 2). In this method, since the noise suppression coefficient estimation formula is a simple formula, the amount of calculation can be reduced as compared with the first prior art. However, it is not a mathematically complete derivation. For this reason, there is a problem that parameters must be determined through trial and error.

  Further, as a third prior art, there is a method in which the calculation amount of the noise suppression coefficient is reduced by performing the same equation modification as the second prior art (see, for example, Non-Patent Document 3). In this method, similarly to the technology described in the second prior art, the noise suppression coefficient estimation formula becomes a simple formula, so that the amount of calculation can be reduced. However, since the speech frequency spectrum is approximated by a Gaussian distribution, there is a problem that the estimation accuracy of the noise suppression coefficient is low.

Further, as a fourth prior art, there is a method obtained by improving the signal-to-noise ratio estimation method in the first prior art (see, for example, Patent Document 1). In this method, the estimation accuracy of the noise suppression coefficient can be increased by increasing the estimation accuracy of the signal-to-noise ratio. However, since the modified Bessel function is used to calculate the noise suppression coefficient, there is still a problem that the amount of calculation is large.
JP 2003-140700 A Y. Ephraim et al., “Speech enhancement using a minimum mean-square error short-time spectral amplitude estimator”, ASSP-32 (6), pp.1109-1121, 1984 T.Lotter et al., “Noise reduction by maximum a posteriori spectral amplitude estimation with supergaussian speech modeling”, IWAENC 2003, pp.83-86 P. Wolfe et al., “Simple alternatives to the Ephraim and Malah suppression rule for speech enhancement”, Proc. IEEE Workshop on SSP, pp.496-499, Aug, 2001

As described above, the techniques described in the respective prior art documents have a problem that the amount of calculation for calculating the noise suppression coefficient is large or the estimation accuracy of the noise suppression coefficient is low.
The present invention has been made paying attention to the above circumstances, and the object of the present invention is to make it possible to estimate a noise suppression coefficient with high accuracy with a small amount of computation, thereby enabling high-speed and high-accuracy noise suppression processing. An object of the present invention is to provide a noise suppressor and a voice communication apparatus including the noise suppressor.

In order to achieve the above object, the present invention divides a frequency spectrum obtained from an input signal into a plurality of bands, estimates a noise spectrum for each of the divided bands, and generates a signal pair from the noise spectrum and the frequency spectrum. A noise ratio is calculated, a noise suppression coefficient is set based on the calculated signal-to-noise ratio, the frequency spectrum is weighted according to the set noise suppression coefficient, and then converted into a time domain signal. In the suppressor, the speech spectrum estimation formula derived by approximating the appearance probability for each real part and imaginary part of the speech spectrum included in the frequency spectrum by a statistical distribution model is partially differentiated and set to zero, and when the phase spectrum was φ | cosφ | + | sinφ | prepared arithmetic expression is approximated as a constant, the noise in accordance with the calculation equation It is obtained so as to set to calculate the pressure coefficient.

  Therefore, according to the present invention, the frequency spectrum of speech is approximated by a statistical distribution model, and an estimation expression of a noise suppression coefficient is derived using the MAP estimation and Bayes' theorem, and the noise suppression coefficient is calculated using this estimation expression. As a result, the noise suppression coefficient estimation formula becomes relatively simple, thereby reducing the amount of calculation. In addition, since the appearance probability for each real part and imaginary part of the speech spectrum is approximated by a statistical distribution model, the approximation accuracy can be improved, and further mathematically complete derivation is performed, so the parameters to be adjusted are As a result, the estimation accuracy of the noise suppression coefficient can be increased.

When deriving the estimation formula of the speech spectrum, the appearance probability for each real part and imaginary part of the speech spectrum is approximated by the first statistical distribution model, and the appearance probability for each real part and imaginary part of the noise spectrum. May be approximated by a second statistical distribution model. In this way, it is possible to derive a speech spectrum estimation formula in consideration of the noise spectrum in addition to the speech spectrum, thereby further improving the accuracy of the speech spectrum estimation formula.
Specifically, a Laplace distribution or a gamma distribution may be used as the first statistical distribution model for approximating the speech spectrum, while a Gaussian distribution may be used as the second statistical distribution model for approximating the noise spectrum.

  Furthermore, it is preferable to derive and set an arithmetic expression for the noise suppression coefficient for each of the divided bands, and calculate the noise suppression coefficient for each of the divided bands using these arithmetic expressions. In this way, the noise suppression coefficient is calculated by an arithmetic expression set independently for each frequency band. For this reason, compared to the case where the noise suppression coefficient is calculated using an arithmetic expression commonly set for all frequency bands, the distribution of the spectral probability density function (SPDF) of the speech differs for each frequency band. Considering the points, it is possible to perform approximation with higher accuracy.

In short, in the present invention, in the means for setting the noise suppression coefficient, the estimation formula of the speech spectrum derived by approximating the appearance probability for each real part and imaginary part of the speech spectrum included in the frequency spectrum by a statistical distribution model is provided. The noise suppression coefficient is calculated according to an arithmetic expression approximated by using | cosφ | + | sinφ | as a constant when the partial differentiation is set to zero and the phase spectrum is φ.
Therefore, according to the present invention, it is possible to estimate the noise suppression coefficient with high accuracy with a small amount of computation, thereby enabling high-speed and high-accuracy noise suppression processing, and a voice communication apparatus including the noise suppressor. Can be provided.

FIG. 1 is a block diagram showing a configuration of a mobile phone which is a first embodiment of a voice communication apparatus provided with a noise suppressor according to the present invention.
A radio signal transmitted from a base station (not shown) is received by the antenna 1 and then input to the receiving circuit (RX) 3 through the antenna duplexer (DUP) 2. The receiving circuit 3 mixes the received radio signal with the local oscillation signal output from the frequency synthesizer (SYN) 4 and converts the frequency into an intermediate frequency signal (down-conversion). Then, the down-converted received intermediate frequency signal is orthogonally demodulated, and a received baseband signal generated thereby is supplied to a code division multiple access (CDMA) signal processing unit 6. The frequency of the local oscillation signal generated from the frequency synthesizer 4 is specified by the control signal SYC from the control unit 18.

  The CDMA signal processing unit 6 includes a RAKE receiver. In the RAKE receiver, the plurality of paths included in the received baseband signal are each subjected to despreading processing using spreading codes. Then, the signals of the respective paths subjected to the despreading process are combined after being matched in phase. As a result, received encoded data is reproduced, and this received encoded data is input to a speech code decoding processing unit (hereinafter referred to as speech codec (SP-COD)) 7.

  The speech codec 7 performs speech decoding on the received encoded data output from the CDMA signal processing unit 6. The PCM codec 8 performs PCM decoding on the digital reception signal output from the speech codec 7 and outputs an analog reception signal. The analog reception signal is amplified by the reception amplifier 9 and then outputted as sound from the speaker 10.

  On the other hand, a speaker's transmission voice signal input to the microphone 11 is amplified to an appropriate level by the transmission amplifier 12 and then subjected to PCM encoding processing by the PCM codec 8, whereby a digital transmission signal and Become. This digital transmission signal is input to the speech codec 7 via a noise suppressor (NS) 20 described later. The speech codec 7 encodes the digital transmission signal output from the PCM codec 8 according to a predetermined speech encoding method. Then, the transmission encoded data generated thereby is supplied to the CDMA signal processing unit 6.

  The CDMA signal processing unit 6 performs spread spectrum processing on the transmission encoded data output from the speech codec 7 using a spreading code assigned to the transmission channel. Then, the signal subjected to the spread spectrum processing is supplied to the transmission circuit (TX) 5. The transmission circuit 5 modulates the spectrum spread signal using a digital modulation method such as a QPSK (Quadrature Phase Shift Keying) method. Then, the transmission signal generated by this modulation is combined with the local oscillation signal generated from the frequency synthesizer 4 and frequency-converted into a radio signal. At the same time, the transmission circuit 5 amplifies the radio signal with a power amplifier at a high frequency so that the transmission power level indicated by the control unit 12 is obtained. The amplified radio signal is supplied to the antenna 1 via the antenna duplexer 2 and transmitted from the antenna 1 to a base station (not shown).

  The control unit 18 uses, for example, a microcomputer, and manages all control related to the communication operation of the mobile phone. The storage unit 13 uses a RAM and a flash memory. The flash memory stores a phone book, transmitted / received electronic mail, various control information, and the like. In addition to the dial keys, the input unit 14 is provided with function keys such as a transmission key, an end key, a power key, a volume adjustment key, and a mode designation key. The display unit 15 is provided with an LCD and an LED. Of these, the LCD displays information indicating the operating state of the device such as the remaining battery level and received electric field strength, e-mails to be transmitted / received, telephone numbers of terminals used by the telephone book and other users, and transmission / reception history Used for. Further, the LED is used for notifying incoming calls and displaying the state of charge of the battery 16. The power supply circuit (POW) generates a predetermined operating power supply voltage Vcc based on the output of the battery 17 and supplies it to each circuit unit.

By the way, the noise suppressor 20 is constituted by a DSP (Digital Signal Processor), for example, and has the following functions. FIG. 2 is a functional block diagram showing the configuration.
The noise suppressor 20 includes a fast Fourier transform unit (FFT) 21, a spectrum amplitude suppression unit 22, an inverse fast Fourier transform unit (IFFT) 23, a first band division unit 24, a voice detection unit 25, and a noise level. An estimator 26, a posteriori SNR estimator 27, a second band divider 28, an a priori SNR estimator 29, and a noise suppression coefficient calculator 30 are provided.

  The fast Fourier transform unit (FFT) 21 converts an input signal from a time domain signal to a frequency domain signal. For example, 128 input digital transmission signals x (t) are divided into frames for each predetermined time length, and fast Fourier transform processing is performed for each of these frames, whereby the amplitude spectrum X (k) (k = 0). ˜N−1, N is the frame length).

  Prior to the fast Fourier transform process, a pre-emphasis process may be performed on the input digital transmission signal x (t) for the purpose of flattening the spectral envelope. Further, the frame length and the shift width of the fast Fourier transform process need not be the same. For example, when the frame length is 128 and the shift width is 80, the input digital transmission signal x (t) for 80 samples is obtained. In order to eliminate the discontinuity of the boundary after storing in the first half of the frame and setting the remaining 48 samples to 0 (zero), windowing of a sine wave characteristic may be performed. More specific methods of pre-emphasis and windowing are described in detail in TIA / EIA IS-127 EVRC, 1997-01, which is a standard for an encoding method standardized by the US TIA.

  The first band dividing unit 24 divides the amplitude spectrum X (k) into, for example, eight frequency bands from a low band to a high band, and takes the average for each frequency band to represent the band power representing each frequency band. Xd (i) (i = 0 to Ni, Ni is, for example, 16 in frequency band) is obtained.

  The voice detection unit 25 compares the band power Xd (i) obtained by the first band dividing unit 24 with the threshold value q (i), and the band power Xd (i) is compared with the threshold value θ (i). If it is larger, it is detected as voice. Here, the threshold value θ (i) may be a fixed value, or may be adaptively updated by taking a weighted average with the past threshold value θ (i) or the band power Xd (i). .

The noise level estimator 26 estimates the noise level N (i) for each band from the band power Xd (i) of the frame that is not detected by the voice detector 24. Here, the current band power Xd (i) may be used as the noise level N (i) as it is, or the noise level N (() is adaptively obtained by taking a weighted average of the past noise level and the band power Xd (i). i) may be sought.
The a posteriori SNR estimator 27 calculates the ratio of the band power Xd (i) obtained by the band divider 24 to the noise level N (i) obtained by the noise level estimator 26, that is, SNR (i). Ask.

The second band dividing unit 28 divides the amplitude spectrum Y (k) after the noise is suppressed by the spectrum amplitude suppressing unit 22 to be described later into, for example, 16 bands from a low frequency range to a high frequency range. The band power Yd (i) (i = 0 to Ni, Ni is, for example, 16 is the number of bands) representing each band.
The prior SNR estimation unit 29 obtains the ratio SNR ′ (i) between the band power Yd (i) obtained by the second band dividing unit 28 and the current noise level N (k).

  The amplitude suppression coefficient calculator 30 is based on the SNR (i) obtained by the a posteriori SNR estimator 27 and the SNR ′ (i) obtained by the a priori SNR estimator 29 according to a previously prepared arithmetic expression. Then, a noise suppression coefficient W (i) is obtained. A method for deriving an arithmetic expression for the noise suppression coefficient W (i) will be described in detail later.

  The spectrum amplitude suppression unit 22 multiplies the amplitude spectrum X (k) obtained by the fast Fourier transform unit (FFT) 21 by the noise suppression coefficient W (i) obtained by the amplitude suppression coefficient calculation unit 30, To obtain an amplitude spectrum Y (k) in which noise is suppressed. Here, the amplitude spectrum X (k) in the same band is multiplied by the same noise suppression coefficient W (i).

  The inverse fast Fourier transform unit (IFFT) unit 23 calculates the amplitude spectrum Y (k) noise-suppressed by the spectrum amplitude suppression unit 22 and the phase spectrum P (k) obtained by the fast Fourier transform unit (FFT) 21. Convert to time domain signal y (t). The converted digital transmission signal y (t) in the time domain is supplied to the speech code decoding processing unit (SP-COD) 7.

If the input digital transmission signal x (t) is pre-emphasized prior to the fast Fourier transform (FFT) process, it is de-emphasized and restored. Further, when the frame length and the shift width of the fast Fourier transform (FFT) process are not the same, the frame boundaries are overlapped to eliminate the discontinuity. More specific methods of de-emphasis and overlap are detailed in TIA / EIA 127 EVRC, 1997-01 mentioned above.
By the way, the arithmetic expression of the noise suppression coefficient W (i) used in the amplitude suppression coefficient calculator 30 is derived as follows.
First, an arithmetic expression of the noise suppression coefficient W (i) is shown below.

Next, a process for deriving an arithmetic expression for the noise suppression coefficient W (i) will be described. In the first embodiment of the present invention, it is assumed that the real part and the imaginary part of a spectral probability density function (SPDF) of speech can be approximated by a statistical distribution model. Here, if a Laplace distribution is used as the statistical distribution model, the speech spectral probability density function (SPDF) is expressed as follows.

Here, φ represents the phase spectrum, and σX represents the dispersion of X. Here, for simplicity of description, the description of the frequency band number i is omitted (hereinafter also omitted). The estimated value X ^ (read as X-hat) of the speech amplitude spectrum X is expressed as follows from the MAP estimation and Bayes' theorem.

Here, p (X) is the output probability of X, and R is the amplitude spectrum of speech with noise . Since p (R) is irrelevant to X, X that maximizes p (R | X) p (X) may be obtained. The conditional probability p (R | X) and the prior probability p (X) are expressed as follows according to Non-Patent Document 2 described above.

Here, σN represents the variance of the amplitude spectrum of noise. Substituting Equations (2) and (3) into Equation (6) and approximating | cosφ | + | sinφ | = a (a is a constant)

Get. Taking the product of Equation (5) and Equation (7), differentiating the logarithm with X and setting it to 0, the optimal solution for X ^ can be derived. Finally, equation (1) is derived as a formula for calculating the noise suppression coefficient W (i).

  With the above configuration, when the digital transmission signal x (t) output from the PCM code decoding processing unit 8 is input to the noise suppressor 20, the digital transmission signal x (t) is first subjected to fast Fourier transform. The unit (FFT) 21 converts each frame into a frequency domain signal, whereby an amplitude spectrum X (k) is obtained. The amplitude spectrum X (k) is divided into, for example, 16 frequency bands by the first band dividing unit 24, and then averaged for each of these frequency bands to obtain band power Xd (i). The band power Xd (i) is compared with a preset threshold value q (i) in the voice detector 25, thereby detecting a voice frame.

  The noise level estimator 26 estimates the noise level N (i) for each band based on the band power Xd (i) of the frame that has not been detected by the voice detector 24 and the estimated noise. Based on the level N (i) and the band power Xd (i) obtained by the band dividing unit 24, the a posteriori SNR estimating unit 27 calculates SNR (i).

  On the other hand, the amplitude spectrum Y (k) after the noise is suppressed in the spectrum amplitude suppression unit 22 is divided into 16 bands after being divided into 16 bands, similar to the amplitude spectrum X (k) before the suppression described above. The band power Yd (i) is averaged for each band. Then, based on the band power Yd (i) and the current noise level N (k), the prior SNR estimation unit 29 calculates SNR ′ (i).

  Now, in the amplitude suppression coefficient calculation unit 30, the SNR (i) obtained by the a posteriori SNR estimation unit 27 and the a priori are calculated in accordance with the estimation equation of the noise suppression coefficient W (i) derived as described above. The noise suppression coefficient W (i) is calculated based on the SNR ′ (i) obtained by the SNR estimation unit 29.

  At this time, the estimation calculation formula of the noise suppression coefficient W (i) is obtained by approximating the appearance probability for each real part and imaginary part of the speech spectrum by a Laplace distribution as shown in the formula (1), and at the same time the real part of the noise spectrum. It is a relatively simple arithmetic expression that approximates the appearance probability for each imaginary part by a Gaussian distribution and is derived using MAP estimation and Bayes' theorem. For this reason, the noise suppression coefficient W (i) is calculated with a relatively small amount of calculation. Moreover, since the appearance probabilities for the real and imaginary parts of the speech spectrum and noise spectrum are approximated by the Laplace distribution and Gaussian distribution, respectively, high-precision approximation is possible, and further mathematically complete derivation is performed. There are few parameters to be adjusted, so that the noise suppression coefficient W (i) can be calculated with high accuracy.

  The noise suppression coefficient W (i) calculated by the amplitude suppression coefficient calculation unit 30 is given to the spectrum amplitude suppression unit 22. The spectrum amplitude suppression unit 22 multiplies the amplitude spectrum X (k) obtained by the fast Fourier transform unit (FFT) 21 by the noise suppression coefficient W (i) calculated by the amplitude suppression coefficient calculation unit 30 for each band. As a result, a noise-suppressed amplitude spectrum Y (k) is obtained. The noise-suppressed amplitude spectrum Y (k) is input to the inverse fast Fourier transform unit (IFFT) unit 23 together with the phase spectrum P (k) obtained by the fast Fourier transform unit (FFT) 21. The inverse fast Fourier transform unit (IFFT) unit 23 converts the noise-suppressed amplitude spectrum Y (k) and phase spectrum P (k) into a time domain signal y (t), and the converted digital transmission. The signal y (t) is subjected to speech encoding processing by the speech encoding / decoding processing unit 7.

  Accordingly, the voice code decoding processing unit 7 performs voice coding processing on the digital transmission signal y (t) in which the noise component is suppressed, and after the transmission coded data generated thereby is modulated, the other party of the call Sent to the mobile phone. As a result, the other party's speaker can hear clear voice with less noise, thereby improving call quality.

  As described above, in the first embodiment, the appearance probability for each real part and imaginary part of the speech spectrum is approximated by a Laplace distribution, and the appearance probability for each real part and imaginary part of the noise spectrum is approximated by a Gaussian distribution. In addition, the noise suppression coefficient W (i) is calculated using an arithmetic expression derived using MAP estimation and Bayes' theorem. Therefore, it is possible to calculate the noise suppression coefficient W (i) with a relatively small amount of computation and with high accuracy, thereby enabling high-speed and high-accuracy noise suppression processing, and voice communication including the noise suppressor. An apparatus can be provided.

  FIG. 3 shows the improvement in the segmental SNR when the noise data is subjected to noise suppression processing by the noise suppressor 20 according to the first embodiment of the present invention, and the second prior art document ( Non-Patent Document 2) and third prior art document (Non-Patent Document 3), the degree of improvement of the segmental SNR when the noise suppression processing is performed by the technique described in the first prior art document (Non-Patent Document 1). The improvement degree of the segmental SNR by the technique described in the above is shown as a baseline, respectively.

  Here, the voice data is a sample of the Japanese part of NTT Advanced Technology, “Multi-lingual Speech database for telephonometry 1994” downsampled to 8kHz, and 4 men and women (total 8 people) uttered 4 sentences. It is. There are three types of noise data: Babble (crowded), Car (in the car), and Street (sidewalk) described in the Japan Electronics Industry Promotion Association, “Electronic Collaborative Noise Database” 1990. Was superimposed on the audio data on the computer so that the SNR was 9 dB and 18 dB.

  FIG. 3 shows that the improvement of the SNR by the noise suppressor 20 according to the first embodiment of the present invention is the best, and a high noise suppression effect can be obtained. In particular, when the SNR is small, that is, when the noise level is large, a remarkable effect can be obtained.

(Second Embodiment)
In the second embodiment of the present invention, a noise suppression coefficient arithmetic expression is derived and set for each divided frequency band, and a noise suppression coefficient is calculated for each of the divided frequency bands using these arithmetic expressions. It is what you do.

  FIG. 4 is a functional block diagram showing a configuration of a noise suppression coefficient calculation unit 30 'which is a main part of a noise suppressor according to the second embodiment of the present invention. As shown in the figure, the noise suppression coefficient calculation unit 30 ′ includes noise suppression coefficient calculation units 31 to 3 n (for example, 16) corresponding to the number of division bands set in the band division unit 24. Each of these noise suppression coefficient calculation units 31 to 3n has an arithmetic expression of the noise suppression coefficient W (i) derived and set independently for each band. That is, each of the noise suppression coefficient calculators 31 to 3n is provided with an arithmetic expression of the noise suppression coefficient W (i) in which λ in the proviso of the expression (1) is set independently for each frequency band.

  Therefore, in the noise suppression coefficient calculation unit 30 ′, the noise suppression coefficient calculation units 31 to 3n calculate the noise suppression coefficient W (i) (i = 1 to n) independently for each frequency band. Each of the spectrum amplitude suppression units 22 performs a suppression process of the noise component included in the amplitude spectrum X (k) for each band according to the noise suppression coefficient W (i) (i = 1 to n) calculated for each band. Is called.

  Here, the reason why λ is set independently for each frequency band is that the distribution of the spectral probability density function (SPDF) of speech differs for each band. FIG. 5 and FIG. 6 are diagrams showing examples of the distribution of the spectral probability density function (SPDF) of speech in the low frequency range and the high frequency range, respectively. Here, the 0 Hz to 250 Hz band and the 250 Hz to 500 Hz band are selected as the low band, and the 2000 Hz to 2250 Hz band and the 2250 Hz to 2500 Hz band are selected as the high band.

  As is apparent from FIGS. 5 and 6, the low-frequency SPDFs are sparsely distributed, while the high-frequency SPDFs are concentrated on small values. It is possible to approximate the SPDFs having different distributions with higher accuracy by approximating with the optimum λ for each band than approximating with the same λ.

(Other embodiments)
In each of the above embodiments, the speech spectrum is approximated using the Laplace distribution, but may be approximated using the gamma distribution. In the above embodiment, the case where the noise suppressor 20 is realized by a DSP has been described as an example. However, the noise suppressor 20 may be realized by causing a microprocessor to execute a noise suppression processing program. In addition, the functional configuration of the noise suppressor, the processing procedure and processing contents of each function, the type and configuration of the voice communication device, and the like can be variously modified without departing from the scope of the present invention.

  In short, the present invention is not limited to the above-described embodiments as they are, and can be embodied by modifying the components without departing from the scope of the invention in the implementation stage. Moreover, various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the above embodiments. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, you may combine suitably the component covering different embodiment.

1 is a block diagram showing a configuration of a CDMA mobile phone which is a first embodiment of a voice communication apparatus including a noise suppressor according to the present invention. The block diagram which shows the function structure of the noise suppressor provided in the audio | voice communication apparatus shown in FIG. The figure which contrasts and shows the noise suppression effect according to use environment of the noise suppressor concerning this invention and the noise suppressor described in prior art literature. The block diagram which shows the structure of the noise suppression count calculation part of the noise suppressor concerning the 2nd Embodiment of this invention. The figure which shows an example of the probability density function (SPDF) of the audio | voice spectrum in a low region. The figure which shows an example of the probability density function (SPDF) of the audio | voice spectrum in a low region.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Antenna, 2 ... Antenna duplexer (DUP), 3 ... Reception circuit (RX), 4 ... Frequency synthesizer (SYN), 5 ... Transmission circuit (TX), 6 ... CDMA signal processing part, 7 ... Speech code decoding process (SP-COD), 8 ... PCM code processing unit (PCM codec), 9 ... receiver amplifier, 10 ... speaker, 11 ... microphone, 12 ... transmitter amplifier, 13 ... storage unit, 14 ... input unit, 15 ... display , 16 ... Power supply circuit (POW), 17 ... Battery, 18 ... Control unit, 20 ... Noise suppressor (NS), 21 ... Fast Fourier transform unit (FFT), 22 ... Spectral amplitude suppression unit, 23 ... Inverse fast Fourier transform Unit (IFFT), 24 ... band division unit, 25 ... voice detection unit, 26 ... noise level estimation unit, 27 ... post-hoc SNR estimation unit, 28 ... band division unit, 29 ... pre-SNR estimation unit, 30, 3 ~3n ... noise suppression coefficient calculation unit.

Claims (8)

Means for obtaining the frequency spectrum from the input signal;
Means for dividing the determined frequency spectrum into a plurality of bands;
Means for estimating a noise spectrum based on the frequency spectrum for each of the divided bands;
Means for calculating a signal-to-noise ratio from the frequency spectrum and the estimated noise spectrum for each of the divided bands;
Means for setting a noise suppression coefficient based on the calculated signal-to-noise ratio;
Means for weighting the determined frequency spectrum according to the set noise suppression coefficient;
Means for converting the weighted frequency spectrum into a time domain signal;
The means for setting the noise suppression coefficient is:
The speech spectrum estimation formula derived by approximating the appearance probability for each real and imaginary part of the speech spectrum included in the frequency spectrum by a statistical distribution model is partially differentiated to be zero, and the phase spectrum is φ and A noise suppressor, wherein the noise suppression coefficient is calculated according to an arithmetic expression approximated by | cosφ | + | sinφ | The means for setting the noise suppression coefficient is:
The appearance probability for each real part and imaginary part of the speech spectrum included in the frequency spectrum is approximated by the first statistical distribution model, and the appearance probability for each real part and imaginary part of the estimated noise spectrum is set to the second According to the arithmetic expression approximated by | cosφ | + | sinφ | as a constant when the differential equation of the speech spectrum derived by approximating with the statistical distribution model is set to zero and the phase spectrum is φ. The noise suppressor according to claim 1, wherein the noise suppression coefficient is calculated. The means for setting the noise suppression coefficient is:
The noise suppressor according to claim 1, further comprising the arithmetic expression derived for each of the divided bands, and calculating the noise suppression coefficient for each of the divided bands according to the arithmetic expression. The means for setting the noise suppression coefficient is:
The speech spectrum estimation formula is derived by using a Laplace distribution or a gamma distribution as a statistical distribution model of the speech spectrum, and using MAP estimation (Maximum A posteriori Estimation) and Bayes' theorem. Or the noise suppressor of 3. The means for setting the noise suppression coefficient is:
Using Laplace distribution or gamma distribution as the first statistical distribution model, using Gaussian distribution as the second statistical distribution model, and using MAP estimation (Maximum A posteriori Estimation) and Bayes' theorem, The noise suppressor according to claim 2 or 3, wherein an estimation formula of a speech spectrum is derived. An audio input unit that inputs audio and outputs a digital audio signal corresponding to the input audio;
A noise suppressor for suppressing a noise component included in the digital audio signal output from the audio input unit and outputting a digital audio signal in which the noise component is suppressed;
A digital audio signal output from the noise suppressor is converted into a transmission signal, and includes a transmission unit that transmits the transmission signal to a transmission line,
The noise suppressor is
Means for obtaining the frequency spectrum from the digital audio signal output from the audio input unit;
Means for dividing the determined frequency spectrum into a plurality of bands;
Means for estimating a noise spectrum based on the frequency spectrum for each of the divided bands;
Means for calculating a signal-to-noise ratio from the frequency spectrum and the estimated noise spectrum for each of the divided bands;
Means for setting a noise suppression coefficient based on the calculated signal-to-noise ratio;
Means for weighting the obtained frequency spectrum according to the set noise suppression coefficient;
Means for converting the weighted frequency spectrum into a time domain signal;
The means for setting the noise suppression coefficient is:
The speech spectrum estimation formula derived by approximating the appearance probability for each real and imaginary part of the speech spectrum included in the frequency spectrum by a statistical distribution model is partially differentiated to be zero, and the phase spectrum is φ and The noise suppression coefficient is calculated according to an arithmetic expression approximated by | cosφ | + | sinφ | Means for obtaining the frequency spectrum from the input signal;
Means for dividing the determined frequency spectrum into a plurality of bands;
Means for estimating the past and present noise spectra based on the frequency spectrum for each of the divided bands;
For each of the divided bands, a subsequent signal-to-noise ratio SNR (i) (i is the number of divided bands) is calculated from the frequency spectrum before noise suppression and the estimated past noise spectrum; Means for calculating a prior signal-to-noise ratio SNR ′ (i) (i is the number of divided bands) from the frequency spectrum after noise suppression and the current noise spectrum;
Means for setting a noise suppression coefficient W (i) (i is the number of divided bands) based on the calculated posterior signal-to-noise ratio SNR (i) and a priori signal band noise ratio SNR ′ (i) When,
Means for weighting the obtained frequency spectrum according to the set noise suppression coefficient W (i);
Means for converting the weighted frequency spectrum into a time domain signal;
The means for setting the noise suppression coefficient W (i) is a formula for calculating the noise suppression coefficient W (i) when λ is a set value for approximation.

A noise suppressor that calculates a noise suppression coefficient W (i) using An audio input unit that inputs audio and outputs a digital audio signal corresponding to the input audio;
A noise suppressor for suppressing a noise component included in the digital audio signal output from the audio input unit and outputting a digital audio signal in which the noise component is suppressed;
A digital audio signal output from the noise suppressor is converted into a transmission signal, and includes a transmission unit that transmits the transmission signal to a transmission line,
The noise suppressor is
Means for obtaining the frequency spectrum from the digital audio signal output from the audio input unit;
Means for dividing the determined frequency spectrum into a plurality of bands;
Means for estimating the past and present noise spectra based on the frequency spectrum for each of the divided bands;
For each of the divided bands, a subsequent signal-to-noise ratio SNR (i) (i is the number of divided bands) is calculated from the frequency spectrum before noise suppression and the estimated past noise spectrum; Means for estimating a prior signal-to-noise ratio SNR ′ (i) (i is the number of divided bands) from the frequency spectrum after noise suppression and the current noise spectrum;
Means for setting a noise suppression coefficient W (i) (i is the number of divided bands) based on the calculated posterior signal-to-noise ratio SNR (i) and a priori signal band noise ratio SNR ′ (i) When,
Means for weighting the obtained frequency spectrum according to the set noise suppression coefficient W (i);
Means for converting the weighted frequency spectrum into a time domain signal;
The means for setting the noise suppression coefficient W (i) is a formula for calculating the noise suppression coefficient W (i) when λ is a set value for approximation.

A voice communication apparatus that calculates a noise suppression coefficient W (i) using

Images