JP2020160290A - Signal processing apparatus, signal processing system and signal processing method - Google Patents

Abstract

To achieve further improvement in suppressing a noise component of a signal.SOLUTION: A signal processing section 10A calculates an average of an observation signal in a predetermined period or longer to acquire a stationary estimated noise spectrum. The signal processing section 10A calculates, based upon an observation signal spectrum and the estimated noise spectrum, a filter coefficient for suppressing a noise component, and uses the filter coefficient to calculate an estimated desired signal of an observation signal. The signal processing section 10A calculates, based upon the observation signal and the estimated desired signal, a noise component of the observation signal in a current time frame, and calculates coherence indicative of correlation between the estimated desired signal spectrum and calculated noise signal spectrum. The signal processing section 10A calculates, according to the calculated coherence, a filter adaptation coefficient contributing to a filter coefficient for each frequency, and uses the filter adaptation coefficient to calculate a filter coefficient of a next time frame.SELECTED DRAWING: Figure 2

Description

The present disclosure relates to a signal processing device, a signal processing system, and a signal processing method that perform signal processing for suppressing a noise component of a signal.

As a technique for suppressing the noise component of the signal to be processed, for example, Patent Document 1 states that the estimated value of the noise spectrum of the input signal is smoothed according to the dispersion of the estimated value to perform spectrum subtraction processing. A method for reducing the generation of musical noise later is disclosed. Further, in Patent Document 2, a voice signal emphasizing the target sound and a voice signal emphasizing noise are calculated by beam forming processing, and when the input sound has a large number of components of the target sound, the degree of noise suppression is reduced. A method for reducing the generation of musical noise by increasing the degree of noise suppression when there are many noise components is disclosed.

Japanese Patent No. 5265056 Japanese Patent No. 5678445

Further improvement was required in the technique of suppressing the noise component of the signal such as the audio signal to be processed.

The present disclosure is devised in view of the above-mentioned conventional circumstances, and an object of the present disclosure is to provide a signal processing device, a signal processing system, and a signal processing method capable of further improving the suppression of noise components of a signal.

The present disclosure is, as one aspect, a signal processing device that performs signal processing on a signal to be processed, and is an observation signal acquisition unit that acquires an observation signal spectrum obtained by converting an observation signal input as the processing target into a frequency region. The estimated noise spectrum obtained by calculating the long-term average of the observed signal for a predetermined period or longer, estimating the stationary noise component in the observed signal from the calculated average value, and converting the estimated noise component into the frequency region. An estimated noise acquisition unit to be acquired, a filter coefficient calculation unit that calculates a filter coefficient for suppressing a noise component of the observed signal based on the observed signal spectrum and the estimated noise spectrum, and a filter using the filter coefficient. A desired signal estimation unit that calculates an estimated desired signal in the observed signal by processing, a signal output unit that outputs a signal obtained by converting the estimated desired signal into a time region as an observed signal after signal processing, and the observed signal and the estimation. A noise signal calculation unit that calculates a noise component in the current observation signal as a calculated noise signal based on the desired signal or the filter coefficient, an estimated desired signal spectrum in the frequency region of the estimated desired signal, and the calculated noise signal. Calculation of frequency region A coherence calculation unit that calculates coherence indicating the correlation with the noise signal spectrum, a filter adaptation coefficient calculation unit that calculates a filter adaptation coefficient that contributes to the filter coefficient for each frequency according to the coherence, and a filter adaptation coefficient calculation unit. The filter coefficient calculation unit provides a signal processing device that calculates the filter coefficient using the filter adaptation coefficient.

The present disclosure is, as one aspect, a signal processing system that performs signal processing on a signal to be processed, and is an input unit for inputting the signal to be processed and an observation signal input from the input unit as a processing target. On the other hand, the signal processing unit includes a signal processing unit that performs signal processing that suppresses the noise component of the observed signal, and an output unit that outputs a signal after signal processing by the signal processing unit. An observation signal acquisition unit that acquires an observation signal spectrum obtained by converting an observation signal into a frequency region, calculates a long-term average of the observation signal for a predetermined period or longer, and uses the calculated average value to generate stationary noise in the observation signal. To suppress the noise component of the observed signal based on the estimated noise acquisition unit that estimates the component and acquires the estimated noise spectrum obtained by converting the estimated noise component into the frequency region, and the observed signal spectrum and the estimated noise spectrum. A filter coefficient calculation unit that calculates a filter coefficient, a desired signal estimation unit that calculates an estimated desired signal in the observed signal by filter processing using the filter coefficient, and a signal that converts the estimated desired signal into a time region are signal processed. A signal output unit that outputs as a later observation signal, a noise signal calculation unit that calculates a noise component in the current observation signal as a calculation noise signal based on the observation signal and the estimated desired signal or the filter coefficient, and the above. A coherence calculation unit that calculates the coherence indicating the correlation between the estimated desired signal spectrum in the frequency region of the estimated desired signal and the calculated noise signal spectrum in the frequency region of the calculated noise signal, and the coherence calculation unit for each frequency according to the coherence. The filter coefficient calculation unit includes a filter adaptation coefficient calculation unit that calculates a filter adaptation coefficient that contributes to the filter coefficient, and the filter coefficient calculation unit provides a signal processing system that calculates the filter coefficient using the filter adaptation coefficient.

The present disclosure is, as one aspect, a signal processing method in a signal processing device that performs signal processing on a signal to be processed, and acquires an observation signal spectrum obtained by converting an observation signal input as the processing target into a frequency region. Estimated noise spectrum obtained by calculating the long-term average of the step and the observed signal for a predetermined period or longer, estimating the stationary noise component in the observed signal from the calculated average value, and converting the estimated noise component into the frequency region. A step of calculating a filter coefficient for suppressing a noise component of the observed signal based on the observed signal spectrum and the estimated noise spectrum, and a filter processing using the filter coefficient to obtain the observed signal. Based on the step of calculating the estimated desired signal in the above, the step of outputting the signal obtained by converting the estimated desired signal into a time region as an observed signal after signal processing, and the observed signal and the estimated desired signal or the filter coefficient. A step of calculating the noise component of the observed signal in the current time frame as a calculated noise signal, an estimated desired signal spectrum in the frequency region of the estimated desired signal, and a calculated noise signal spectrum in the frequency region of the calculated noise signal. A step of calculating the coherence showing the correlation, a step of calculating the filter adaptation coefficient contributing to the filter coefficient of the next time frame for each frequency according to the coherence, and a step of calculating the filter adaptation coefficient using the filter adaptation coefficient, and the next time frame using the filter adaptation coefficient. Provided is a signal processing method comprising the step of calculating the filter coefficient of the above.

According to the present disclosure, further improvement can be realized with respect to suppression of noise components of signals.

Block diagram showing an example of the configuration of the signal processing system according to the embodiment A block diagram showing an example of the configuration of the signal processing device according to the first embodiment. Block diagram showing a configuration example of the coherence calculation unit according to the first embodiment A block diagram showing an example of the configuration of the signal processing device according to the second embodiment. Block diagram showing a configuration example of the coherence calculation unit according to the second embodiment Block diagram showing an example of the configuration of the signal processing device according to the third embodiment.

(Findings that led to this disclosure)
As a method of suppressing noise such as Gaussian noise from a signal to be processed, a method based on spectrum subtraction such as a Wiener filter is known. As a problem of the noise suppression method based on the spectrum subtraction, there is a problem that toned noise called musical noise is generated at the time of noise suppression. This is because there is an error in the estimated value of the noise spectrum to be suppressed, so that an isolated peak is generated in the spectrum after the spectrum subtraction process.

In order to solve this problem, for example, in Patent Document 1, the estimated value of the noise spectrum of the input signal is smoothed according to the dispersion of the estimated value to generate musical noise after the spectrum subtraction processing. A method for reducing the noise has been proposed. However, there is a problem that the accuracy of noise suppression deteriorates in the entire band of the signal by smoothing the estimated value of the noise spectrum. Further, Patent Document 2 proposes a method using a beamforming process. In Patent Document 2, a voice signal emphasizing the target sound and a voice signal emphasizing noise are calculated by beam forming processing, the correlation between the signal emphasizing the target sound and the input sound, and the signal emphasizing noise and the input sound. By calculating the correlations of the above and taking the ratio of the calculated correlation values, the degree of mixing of the target sound component and the noise component is estimated from the input voice. Next, in the input sound, when the target sound component is large, the noise suppression degree is reduced, and when the noise component is large, the noise suppression degree is increased to reduce the generation of musical noise. However, since it is necessary to use a microphone array in estimating the degree of mixing of the target sound component and the noise component of the input voice, the number of microphones increases and the configuration becomes complicated, and it is necessary to introduce beamforming processing. Therefore, there is a problem that the amount of calculation and the hardware cost increase as a system.

Therefore, in the present disclosure, a configuration example of a signal processing device that reduces musical noise while maintaining the accuracy of noise suppression without complicating the configuration is shown.

Hereinafter, embodiments in which the signal processing apparatus, signal processing system, and signal processing method according to the present disclosure are specifically disclosed will be described in detail with reference to the accompanying drawings as appropriate. However, more detailed explanation than necessary may be omitted. For example, detailed explanations of already well-known matters and duplicate explanations for substantially the same configuration may be omitted. This is to avoid unnecessary redundancy of the following description and to facilitate the understanding of those skilled in the art. It should be noted that the accompanying drawings and the following description are provided for those skilled in the art to fully understand the present disclosure, and are not intended to limit the subject matter described in the claims.

In the following embodiment, as an example of the signal processing device, a configuration example of a signal processing system having a signal processing unit using a Wiener filter is shown.

(System configuration of the embodiment)
FIG. 1 is a block diagram showing an example of the configuration of the signal processing system according to the embodiment. In the present embodiment, for example, a configuration example of a signal processing system in which an audio signal (acoustic signal) is used as a signal to be processed and the audio signal to be processed is input as an observation signal to perform signal processing is shown. The signal processing system of the present embodiment performs signal processing for suppressing noise components by using a Wiener filter on the input observation signal, and outputs a signal after noise suppression. The signal processing system of the present embodiment is used, for example, in a hands-free communication device mounted in a vehicle interior of a vehicle, and is applied to a system that processes a voice signal (speaker or other party's voice signal) related to a call. It is possible.

The signal processing system includes a signal processing unit 10 that performs signal processing, a microphone 51 as an example of a signal input device, an input unit 52 including an AD converter and the like, an output unit 53 including a DA converter and the like, and an example of a signal output device. Includes the speaker 54 of. In the signal processing unit 10, the input unit 52 and the microphone 51 are connected to the input terminal 31, and the output unit 53 and the speaker 54 are connected to the output terminal 32.

The signal processing unit 10 includes a processor and a memory, has a function of a signal processing circuit including a Wiener filter, and executes digital signal processing for suppressing a noise component of an observed signal. The signal processing unit 10 realizes various functions by, for example, the processor executing a predetermined program held in the memory or the storage. The processor may include an MPU (Micro processing Unit), a CPU (Central Processing Unit), a DSP (Digital Signal Processor), a GPU (Graphical Processing Unit), and the like. The memory may include a RAM (Random Access Memory), a ROM (Read Only Memory), and the like. The storage may include an HDD (Hard Disk Drive), an SSD (Solid State Drive), an optical disk device, a memory card, and the like. Further, the signal processing unit 10 may be configured by a microcomputer or a hardware circuit that performs signal processing according to the present disclosure.

In the microphone 51, the ambient sound is picked up, converted into an audio signal of an electric signal, and input to the input unit 52. The input unit 52 converts the audio signal input from the microphone 51 into a digital signal by the AD converter, outputs the signal, and inputs the audio signal to the signal processing unit 10 via the input terminal 31. The signal processing unit 10 uses the audio signal of the input digital signal as the observation signal to be processed, executes digital signal processing described later on the observation signal, and suppresses the noise component of the signal. The signal processing unit 10 outputs an observed signal (audio signal) after noise-suppressed signal processing via the output terminal 32, and inputs it to the output unit 53. The output unit 53 converts the audio signal after signal processing into an analog signal by the DA converter and outputs it to the speaker 54. The input audio signal is reproduced and output from the speaker 54, and a desired noise-suppressed sound is emitted from the speaker 54. The input unit 52, the signal processing unit 10, and the output unit 53 may appropriately perform signal amplification, filtering, and the like.

Hereinafter, some examples of specific configurations and operations of the signal processing device including the signal processing unit 10 according to the present embodiment will be described.

(Embodiment 1)
FIG. 2 is a block diagram showing an example of the configuration of the signal processing device according to the first embodiment. The signal processing unit 10A includes an FFT (Fast Fourier Transform) unit 11, an average and signal coupling unit 12, an input average storage unit 13, an FFT unit 14, a winner filter calculation unit 15, an estimated desired signal spectrum calculation unit 16, and an IFFT (Inverse Fast). It has a Fourier Transform) unit 17, a subtraction unit 18A, a coherence calculation unit 19A, a winner filter adaptation coefficient calculation unit 20, and a winner filter adaptation coefficient storage unit 21.

The audio signal preprocessed by the input unit 52 is input to the signal processing unit 10A from the input terminal 31. The signal processing unit 10A executes signal processing by a Wiener filter, which will be described later, with respect to the observation signal to be processed to be input to the input end 31, and then outputs the processed noise-suppressed observation signal from the output end 32. Here, a processing example in which the observation signal is the observation signal vector x _t will be described.

The FFT unit 11 performs a fast discrete Fourier transform process on the observation signal vector x _t input from the input end 31, converts the signal from the time domain to the frequency domain, and calculates the observation signal spectrum X _t . The FFT unit 11 has a function as an example of an observation signal acquisition unit. The mean and the signal combiner 12, with respect to the observed signal vector x _t of the time series input from the input terminal 31, calculates the average value ¯x _t of the observation signals of each predetermined time frame averaged. Here, a character symbol with a bar indicating the average value is expressed as ￣ _{x t} . The input average storage unit 13 stores the calculated average value ￣ _{x t} of the observed signal vector.

The average and the signal coupling section 12 reads the average value for a predetermined period or more past observations signal from the input average storage unit 13, calculates a vector n _t combine with the mean value of the current observed signal. The calculated vector n _t is equivalent to a signal obtained by averaging the observed signals over a long period of time. When assuming as an observation signal a signal in which stationary noise such as automobile driving noise is superimposed on a non-stationary signal such as an audio signal, the long-term average of the observed signal is assumed as the signal to be processed. By taking the above, the level of a non-stationary signal such as an audio signal decreases according to the length of the average interval. On the other hand, the level of stationary noise such as the running noise of an automobile is constant regardless of the length of the average section. Thus, the vector n _t bound the average value of the past and current observation signal, since it presumed to be equivalent to the noise signal, to be regarded to as estimated noise signal vector n _t. The predetermined period for averaging the observed signals over a long period of time is, for example, about 10 to 20 seconds. The FFT unit 14 performs FFT processing on the estimated noise signal vector n _t to convert it from the time domain to the frequency domain, and calculates the estimated noise signal spectrum N _t . The average and signal coupling unit 12, the input average storage unit 13, and the FFT unit 14 have a function as an example of the estimated noise acquisition unit.

The Wiener filter calculation unit 15 calculates the Wiener filter coefficient H _t by using the calculated observation signal spectrum X _t and the estimated noise signal spectrum N _t and the Wiener filter adaptation coefficient α _t-1, which will be described later. The Wiener filter calculation unit 15 has a function as an example of the filter coefficient calculation unit. The estimated desired signal spectrum calculation unit 16 calculates the estimated desired signal spectrum ^ Y _t from the observed signal spectrum X _t by the filter processing using the calculated Wiener filter coefficient H _t . Here, a character symbol with a hat indicating an estimated value is expressed as ^ Y _t . The estimated desired signal spectrum calculation unit 16 has a function as an example of the desired signal estimation unit. The IFFT unit 17 performs a high-speed discrete inverse Fourier transform process on the estimated desired signal spectrum ^ Y _t , converts the signal from the frequency domain to the time domain, and calculates the estimated desired signal vector ^ y _t . The IFFT unit 17 has a function as an example of a signal output unit.

The subtraction unit 18A subtracts the estimated desired signal vector ^ y _t from the observation signal vector x _t to calculate the calculated noise signal vector ^ n _t . The subtraction unit 18A has a function as an example of a noise signal calculation unit. By taking the difference between the observed signal and the estimated desired signal containing a noise component, it is possible to obtain a noise signal, is calculated noise signal vector ^ n _t is determined by the subtraction unit 18A. Coherence calculation unit 19A uses the calculated noise signal vector ^ n _t and the estimated desired signal vector ^ y _t calculated above, calculates a coherence (correlation value) gamma _t showing the correlation between the estimated desired signal and calculating a noise signal To do.

Here, the ratio of the desired signal component included in the calculated noise signal can be inferred from the coherence γ _t showing the correlation between the estimated desired signal and the calculated noise signal. When the coherence γ _t is small and the correlation is weak, it is considered that the desired signal component is small in the calculated noise signal. On the other hand, when the coherence γ _t is large and the correlation is strong, it is considered that a large amount of desired signal components are mixed in the calculated noise signal. When the coherence γ _t is large, if the Wiener filter is applied as it is, not only the noise component but also the desired signal component is suppressed, which may lead to deterioration of sound quality and generation of musical noise. Therefore, in the present embodiment, the Wiener filter coefficient Ht is varied for each frequency according to the magnitude of the coherence γ _t , and the operation of the Wiener filter is optimized.

The Wiener filter adaptation coefficient calculation unit 20 calculates the Wiener filter adaptation coefficient α _t to be used for the next time frame for each frequency based on the coherence γ _t calculated by the coherence calculation unit 19A. The Wiener filter adaptation coefficient calculation unit 20 has a function as an example of the filter adaptation coefficient calculation unit. The Wiener filter adaptation coefficient storage unit 21 stores the calculated Wiener filter adaptation coefficient α _t . The Wiener filter adaptation coefficient storage unit 21 has a function as an example of the filter coefficient storage unit. In the present embodiment, the Wiener filter adaptation coefficient α _t is used as the intensity coefficient of the Wiener filter that varies depending on the coherence value. For example, in a frequency band where the coherence γ _t is small (the calculated noise signal has few desired signal components), the Wiener filter adaptation coefficient α _t is increased to increase the degree of application of the Wiener filter and increase the degree of suppression of the noise component. As a result, a large amount of noise components contained in the observation signal are suppressed. Further, in the frequency band where the coherence γ _t is large (the calculated noise signal has many desired signal components), the Wiener filter adaptation coefficient α _t is reduced to weaken the degree of application of the Wiener filter and reduce the degree of suppression of the noise component. As a result, it is less likely that the desired signal component is suppressed together with the noise component, and the generation of musical noise when the noise component is suppressed is reduced. In this way, by changing the strength of the Wiener filter for each frequency, deterioration of sound quality and generation of musical noise are suppressed while maintaining the accuracy of noise suppression as much as possible.

The Wiener filter calculation unit 15 calculates the Wiener filter coefficient H _t using the Wiener filter adaptation coefficient α _t with respect to the observed signal spectrum X _t and the estimated noise signal spectrum N _t in the next time frame.

Next, the processing of the signal processing unit 10A in the first embodiment will be described more specifically.

The FFT unit 11 converts the signal from the time domain to the frequency domain by, for example, the discrete Fourier transform operation shown in the following equation (1). Actually, the high-speed discrete Fourier transform process is executed.

In the above equation, N is the period of the signal, n is the index indicating the discrete time, f is the index indicating the discrete frequency, x _n is the time domain signal, and X _f is the frequency domain signal. The discrete time by n is the time in units of the sampling period of the observed signal. For example, when the sampling frequency = 48 kHz, (1/48000) seconds ≈ 0.021 ms is one unit. When the number of observation signal samples is N = 1024, one time frame is (1024/48000) ≈21.3 ms. The discrete frequency according to f is a frequency in units of (sampling frequency / signal period) of the observed signal. For example, when the sampling frequency = 48 kHz and the number of samples N = 1024, (48000/1024) Hz ≈ 46.9 Hz is one unit.

The FFT unit 11 multiplies the entire signal of the time frame having the frame width L by a window function such as a Hanning window with respect to the input time domain signal, and regards the input time domain signal as a periodic signal having a period L. In the above example, the time frame has a period L = 21.3 msec. Then, the FFT unit 11 performs a fast discrete Fourier transform with a reduced amount of calculation based on the discrete Fourier transform shown in the above equation (1) on the periodic signal of the observation signal having a period L, and converts the observation signal into the frequency domain. To do.

Here, the observation signal in the time domain input to the signal processing unit 10A is represented by, for example, the observation signal vector x _t as shown in the following equation (2).

x _t = [x (L × t), x (L × (t + 1)), ..., x (L × N (t + 1))] ^T
… (2)

In the above equation, the equation on the right indicates each element x (n) of the vector, n is a parameter representing a time, t is a time frame, and T is a transpose. The same applies to the following equation.

The FFT unit 11 converts the observed signal vector x _t from the time domain to the frequency domain by the discrete Fourier transform operation based on the above equation (1), and the observed signal spectrum X _t as shown in the following equation (3). Is calculated.
X _t = [X _t (0), X _t (1), ..., X _t (Ω)] ^T ... (3)

In the above equation, the equation on the right indicates each element X _t (ω) of the vector, ω represents the frequency bin, and ω = 0, 1, 2, ..., Ω. The same applies to the following equation.

The average and signal coupling unit 12 averages the observed signals by, for example, the calculation shown in the following equation (4), and calculates the average value of the observed signals for each time frame.

In the above equation, L is the frame width of the time frame, n represents an index indicating the discrete time, x _t is the observed signal, ¯x _t represents the average value of the observed signal.

Further, the average and signal coupling unit 12 reads out the average value of the observation signals for the past N-1 time frames from the input average storage unit 13, and the average value of the observation signals of the current time frame calculated above ￣ _{x t.} Combine with to calculate the vector representing the long-term average of the observed signals. When the noise signal is stationary, the noise signal can be estimated by taking the long-term average of the observed signal xt. In this case, the mean and the signal combiner 12 is, for example, as shown in the following equation (5), it calculates a vector representing the long-term average of the observed signal as an estimated noise signal vector n _t.

FFT unit 14, to the estimated noise signal vector n _t, converted from the time domain to the frequency domain by a discrete Fourier transform operation based on the above formula (1), the estimated noise signal spectrum as shown in the following equation (6) Calculate N _t .

N _t = [N _t (0), N _t (1), ..., N _t (Ω)] ^T ... (6)

Wiener filter calculating section 15, and the observed signal spectrum and the estimated noise signal spectrum, based on the Wiener filter adaptation coefficient, to calculate the Wiener filter coefficients H _t in the frequency domain, for example by calculation shown in the following equation (7).

In the above equation, X _t (ω) is the observed signal spectrum in the time frame t, N _t (ω) is the noise signal spectrum in the time frame t, and α _t-1 (ω) is the estimation calculated in the previous time frame t-1. The Wiener filter adaptation coefficient, H _t (ω), which fluctuates depending on the coherence value of the desired signal and the calculated noise signal, represents the Wiener filter coefficient in the time frame t.

Here, the Wiener filter calculation unit 15 performs a calculation based on the above equation (7) using the observation signal spectrum X _t shown in the above equation (3) and the estimated noise signal spectrum N _t shown in the above equation (6). to calculate the Wiener filter coefficients H _t as shown in the following equation (8).

H _t = [H _t (0), H _t (1), ..., H _t (Ω)] ^T ... (8)

The Wiener filter adaptation coefficient α _t-1 calculated in the previous time frame t- ₁ is represented by the following equation (9).

α _t-1 = [α _t-1 (0), α _t-1 (1), ..., α _t-1 (Ω)] ^T … (9)

Estimating the desired signal spectrum calculating unit 16, by using the calculated Wiener filter coefficients H _t, for example, calculates the estimated desired signal spectrum from the observed signal by the filter processing operation shown in the following equation (10).

In the above equation, H _t (ω) represents the Wiener filter coefficient in the time frame t, X _t (ω) represents the observed signal spectrum in the time frame t, and ^ Y _t (ω) represents the estimated desired signal spectrum in the time frame t. There is.

The estimated desired signal spectrum ^ Y _t (ω) of a desired signal such as an audio signal estimated by the winner filter is obtained by multiplying the amplitude spectrum of the observed signal by the calculated winner filter coefficient and further multiplying the phase spectrum of the observed signal. It can be calculated by doing.

Here, the estimated desired signal spectrum calculating unit 16, Wiener filter coefficients shown in the equation (8) H _t, using the observed signal spectrum X _t shown in the equation (3), a calculation based on the equation (10) Then, the estimated desired signal spectrum ^ Y _t as shown in the following equation (11) is calculated.

The IFFT unit 17 converts the signal from the frequency domain to the time domain by, for example, the discrete inverse Fourier transform operation shown in the following equation (12). Actually, the high-speed discrete inverse Fourier transform process is executed.

In the above equation, N is the period of the signal, n is the index indicating the discrete time, f is the index indicating the discrete frequency, x _n is the time domain signal, and X _f is the frequency domain signal. The sampling period of the discrete time, the sampling frequency, and the number of samples of the observed signal are the same as those of the discrete Fourier transform operation shown in the above equation (1).

The IFFT unit 17 regards the input frequency domain signal as a frequency domain signal having L frequency components as a periodic signal having a period L. Then, the IFFT unit 17 performs a high-speed discrete inverse Fourier transform with a reduced amount of calculation based on the discrete inverse Fourier transform shown in the above equation (12) on the periodic signal of the estimated desired signal having a period L, and obtains the estimated desired signal. Convert to time domain.

Here, the IFFT unit 17 converts the estimated desired signal spectrum ^ Y _t from the frequency domain to the time domain by the discrete inverse Fourier transform operation based on the above equation (12), and is shown in the following equation (13). The estimated desired signal vector ^ y _t is calculated.

Subtraction unit 18A performs a subtraction process of subtracting the estimated desired signal vector ^ y _t from the observation signal vector x _t, and calculates the calculated noise signal vector ^ n _t as shown in the following equation (14).

FIG. 3 is a block diagram showing a configuration example of the coherence calculation unit 19A according to the first embodiment. The coherence calculation unit 19A includes FFT units 191 and 192, power spectrum calculation units 193 and 194, a cross spectrum calculation unit 195, and a coherence calculation unit 196.

The FFT unit (an example of the first frequency conversion unit) 191 converts the estimated desired signal vector ^ y _t from the time domain to the frequency domain by the discrete Fourier transform operation based on the above equation (1), and the following equation The estimated desired signal spectrum ^ Y _t as shown in (15) is calculated.

FFT section (second example of the frequency conversion section) 192 for calculating the noise signal vector ^ n _t, converted from the time domain to the frequency domain by a discrete Fourier transform operation based on the above formula (1), the following equation calculating the calculated noise signal spectrum ^ _{N t} as shown in (16).

The power spectrum calculation units 193 and 194 calculate the power spectrum from the signal spectrum by, for example, the calculation shown in the following equation (17).

_PXX (ω) = | X (ω) | ² ... (17)
In the above equation, _PXX (ω) represents the power spectrum.

Here, the power spectrum calculation unit (first power spectrum calculation unit) 193 performs an calculation based on the above equation (17) using the estimated desired signal spectrum ^ Y _t, and is shown in the following equation (18). The power spectrum P _{^ Yt ^ Yt} of the desired estimated signal is calculated.

The power spectrum calculating unit (second power spectrum calculation section) 194, using the calculated noise signal spectrum ^ N _t, performs a calculation based on the equation (17), as shown in the following equation (19) Calculation The power spectrum P _{^ Nt ^ Nt} of the noise signal is calculated.

The cross spectrum calculation unit 195 calculates the cross spectrum from the signal spectrum by, for example, the calculation shown in the following equation (20).

P _XY (ω) = X ^* (ω) Y (ω)… (20)

In the above equation, _PXY (ω) represents the cross spectrum and X ^* represents the complex conjugate of X.

Here, the cross spectrum calculation unit 195 performs an calculation based on the above equation (20) using the estimated desired signal spectrum ^ Y _t and the calculated noise signal spectrum ^ N _t, and is shown in the following equation (21). The cross spectrum P _{^ Nt ^ Yt} of the estimated desired signal and the calculated noise signal is calculated.

The coherence calculation unit 196 calculates coherence from the power spectrum and the cross spectrum by, for example, the calculation shown in the following equation (22).

In the above equation, γ (ω) represents coherence for each frequency.

Here, the coherence calculation unit 196 uses the power spectrum P _{^ Yt ^ Yt} of the estimated desired signal, the power spectrum P _{^ Nt ^ Nt of the} calculated noise signal, and the cross spectrum P _{^ Nt ^ Yt} to use the above equation (22). The coherence γ _t as shown in the following equation (23) is calculated by performing the calculation based on. The coherence γ _t calculated in this way becomes the coherence γ _t of the calculation result of the coherence calculation unit 19A.

γ _t = [γ _t (0), γ _t (1), ..., γ _t (Ω)] ^T ... (23)

The Wiener filter adaptation coefficient calculation unit 20 calculates the Wiener filter adaptation coefficient used for the next time frame t + 1 based on the coherence γ _t in the time frame t calculated as described above, for example, by the following equation (24). Calculated for each frequency index.

α _t (ω) = (1-γ _t (ω)) ² … (24)

In the above equation, γ _t (ω) represents the coherence value for each frequency index, and α _t (ω) represents the Wiener filter adaptation coefficient for each frequency index. However, 0 ≦ α _t (ω) ≦ 1 and t ≧ 0.

When referring to the past values, the Wiener filter adaptation coefficient calculation unit 20 calculates the Wiener filter adaptation coefficient by, for example, the calculation shown in the following equation (25).

α _t (ω) = (ηα _t-1 (ω)) (1-γ _t (ω)) ² … (25)
In the above equation, η represents the forgetting coefficient. However, 0 <η <1.

Here, the Wiener filter adaptation coefficient calculation unit 20 performs an operation based on the above equation (24) or equation (25) using the coherence γ _t shown in the above equation (23), and is shown in the following equation (26). The Wiener filter adaptation coefficient α _t is calculated.

α _t = [α _t (0), α _t (1),…, α _t (Ω)] ^T … (26)

Then, the Wiener filter calculation unit 15 uses the Wiener filter adaptation coefficient α _t calculated for each frequency as described above, and uses the Wiener filter for the observation signal of the next time frame t + 1 as in the previous time frame t. The filter coefficient H _t is calculated (see the above equations (7) and (8)).

As described above, in the present embodiment, the Wiener filter coefficient varied for each frequency is calculated according to the magnitude of the coherence between the estimated desired signal and the calculated noise signal, and the Wiener filter coefficient is used to obtain the observed signal. An estimated desired signal with suppressed noise components is acquired. Then, the Wiener filter adaptation coefficient is sequentially updated for each time frame of the input observation signal, and the Wiener filter coefficient is calculated to appropriately select the noise suppression intensity for each frequency band and select the noise component. Suppress. As a result, the generation of musical noise can be reduced while maintaining the strength of noise suppression as much as possible, and the noise suppression of the observed signal can be performed with high accuracy.

(Embodiment 2)
FIG. 4 is a block diagram showing an example of the configuration of the signal processing device according to the second embodiment. The signal processing unit 10B includes an FFT unit 11, an average and signal coupling unit 12, an input average storage unit 13, an FFT unit 14, a Wiener filter calculation unit 15, an estimated desired signal spectrum calculation unit 16, an IFFT unit 17, a subtraction unit 18B, and coherence. It has a calculation unit 19B, a Wiener filter adaptation coefficient calculation unit 20, and a Wiener filter adaptation coefficient storage unit 21.

The second embodiment is an example in which the configurations of the coherence calculation unit 19B and the subtraction unit 18B are different from those of the first embodiment, and the coherence calculation method is changed. Hereinafter, the parts different from those of the first embodiment will be mainly described, and the description of the same configuration and processing as that of the first embodiment will be omitted.

The subtraction unit 18B subtracts the estimated desired signal spectrum ^ Y _t from the observed signal spectrum X _t in the frequency domain to calculate the calculated noise signal spectrum ^ N _t . The subtraction unit 18B has a function as an example of a noise signal calculation unit. The coherence calculation unit 19B uses the estimated desired signal spectrum ^ Y _t and the calculated noise signal spectrum ^ N _t calculated by the subtraction unit 18B to show the correlation between the estimated desired signal and the calculated noise signal in the frequency domain. The coherence (correlation value) γ _t is calculated.

FIG. 5 is a block diagram showing a configuration example of the coherence calculation unit 19B according to the second embodiment. The coherence calculation unit 19B includes power spectrum calculation units 193 and 194, a cross spectrum calculation unit 195, and a coherence calculation unit 196.

Power spectrum calculating unit (first power spectrum calculating unit) 193, using the estimated desired signal spectrum ^ Y _t, as in the first embodiment, calculates a power spectrum P _{^ Yt ^ Yt} estimated desired signal ( (See equations (17) and (18) above). The power spectrum calculating unit (second power spectrum calculating unit) 194 calculates using the calculated noise signal spectrum ^ N _t, as in the first embodiment, the power spectrum P _{^ Nt ^ Nt} calculated noise signal (Refer to the above equations (17) and (19)). The cross spectrum calculation unit 195 uses the estimated desired signal spectrum ^ Y _t and the calculated noise signal spectrum ^ N _t , and similarly to the first embodiment, the cross spectrum P _{^ Nt ^ Yt} of the estimated desired signal and the calculated noise signal. (See equations (20) and (21) above).

The coherence calculation unit 196 uses the power spectrum P _{^ Yt ^ Yt} of the estimated desired signal, the power spectrum P _{^ Nt ^ Nt of the} calculated noise signal, and the cross spectrum P _{^ Nt ^ Yt} in the same manner as in the first embodiment. The coherence γ _t is calculated (see the above equations (22) and (23)). The coherence γ _t calculated in this way becomes the coherence γ _t of the calculation result of the coherence calculation unit 19B.

The Wiener filter adaptation coefficient calculation unit 20 calculates the Wiener filter adaptation coefficient α _t for each frequency based on the coherence γ _t calculated by the coherence calculation unit 19B (the above equations (24), (2). 25), (26)).

Then, the Wiener filter calculation unit 15 uses the Wiener filter adaptation coefficient α _t calculated for each frequency as described above, and uses the Wiener filter for the observation signal of the next time frame t + 1 as in the previous time frame t. The filter coefficient H _t is calculated (see the above equations (7) and (8)).

In the second embodiment, by calculating the coherence in the frequency domain, the configuration of the coherence calculation unit can be simplified and the amount of calculation can be reduced. Therefore, it is possible to shorten the processing time of the signal processing unit and improve the noise suppression accuracy, which is more preferable.

(Embodiment 3)
FIG. 6 is a block diagram showing an example of the configuration of the signal processing device according to the third embodiment. The signal processing unit 10C includes an FFT unit 11, an average and signal coupling unit 12, an input average storage unit 13, an FFT unit 14, a Wiener filter calculation unit 15, an estimated desired signal spectrum calculation unit 16, an IFFT unit 17, and a noise signal calculation unit 22. , The coherence calculation unit 19B, the Wiener filter adaptation coefficient calculation unit 20, and the Wiener filter adaptation coefficient storage unit 21.

The third embodiment is an example in which the noise signal calculation unit 22 is provided instead of the subtraction unit 18B as compared with the second embodiment, and the calculation method of the calculated noise signal spectrum is changed. Hereinafter, the parts different from those of the first embodiment and the second embodiment will be mainly described, and the same configurations and processes as those of the first embodiment and the second embodiment will be omitted.

The noise signal calculation unit 22 is calculated based on the Wiener filter coefficient H _t calculated by the Wiener filter calculation unit 15 and the observation signal spectrum X _t , for example, by the calculation shown in the following equation (27). Noise signal spectrum ^ N _t Is calculated.

The formula (27), as shown in the following equation (28) is obtained by deforming the equation for calculating the calculated noise signal spectrum ^ N _t by subtraction unit 18B in the second embodiment.

The coherence calculation unit 19B uses the estimated desired signal spectrum ^ Y _t and the calculated noise signal spectrum ^ N _t calculated by the noise signal calculation unit 22 to determine the correlation between the estimated desired signal and the calculated noise signal in the frequency domain. The indicated coherence (correlation value) γ _t is calculated.

The Wiener filter adaptation coefficient calculation unit 20 calculates the Wiener filter adaptation coefficient α _t for each frequency based on the coherence γ _t calculated by the coherence calculation unit 19B, as in the first and second embodiments (the above equation (24)). , (25), (26)).

Then, the Wiener filter calculation unit 15 uses the Wiener filter adaptation coefficient α _t calculated for each frequency as described above, and uses the Wiener filter for the observation signal of the next time frame t + 1 as in the previous time frame t. The filter coefficient H _t is calculated (see the above equations (7) and (8)).

In the third embodiment, the calculation amount can be reduced and the processing time of the signal processing unit can be shortened by calculating the calculated noise signal spectrum using the calculation result of the Wiener filter coefficient, which is more preferable. ..

In the present embodiment described above, it is assumed that the noise signal is stationary in the stationary observation signal, the generation of musical noise is reduced under this condition, and the noise component of the observation signal is suppressed. The signal processing unit of the present embodiment calculates the coherence between the estimated desired signal and the calculated noise signal in the frequency domain. Then, the Wiener filter coefficient is calculated according to the magnitude of coherence for each frequency, and the noise component of the observed signal is suppressed by the Wiener filter using this Wiener filter coefficient.

At this time, the intensity of noise suppression is maintained in the frequency band with good noise estimation accuracy (small coherence), and the intensity of noise suppression is reduced in the frequency band with poor noise estimation accuracy (large coherence). As a result, the frequency band that deteriorates the accuracy of noise suppression can be adaptively narrowed according to the observed signal, and musical noise can be reduced while maintaining the accuracy of noise suppression as much as possible. Further, in the present embodiment, since processing for a single input signal is premised, the generation of musical noise is reduced by a simple system configuration without using a plurality of input means such as a plurality of microphones and a microphone array. It is possible to suppress the noise. In addition, beamforming processing in the previous stage is not required, and the amount of calculation can be reduced and the hardware cost can be reduced.

As described above, the present embodiment shows an example of a signal processing device, a signal processing system, and a signal processing device method that perform signal processing on a signal to be processed. The signal processing system is a signal processing unit 10 that performs signal processing that suppresses the noise component of the observed signal with respect to the input unit 52 that inputs the signal to be processed and the observation signal that is input from the input unit 52 as the processing target. And an output unit 53 that outputs a signal after signal processing by the signal processing unit 10. In the signal processing unit 10, the FFT unit 11 acquires an observation signal spectrum obtained by converting an observation signal input as a processing target into a frequency domain. The average and signal coupling unit 12, the input average storage unit 13, and the FFT unit 14 calculate the long-term average of the observation signal for a predetermined period or longer, and estimate the stationary noise component in the observation signal from the calculated average value. , The estimated noise spectrum obtained by converting the estimated noise component into the frequency domain is acquired. The Wiener filter calculation unit 15 calculates the Wiener filter coefficient for suppressing the noise component of the observation signal based on the observation signal spectrum and the estimated noise spectrum. The estimated desired signal spectrum calculation unit 16 calculates the estimated desired signal spectrum in the observed signal by filtering using the Wiener filter coefficient. The IFFT unit 17 outputs a signal obtained by converting the estimated desired signal spectrum into a time domain as an observation signal after signal processing. The subtraction unit 18A, 18B, or the noise signal calculation unit 22 calculates the noise component in the observation signal of the current time frame as the calculation noise signal based on the observation signal and the estimated desired signal or filter coefficient. The coherence calculation units 19A and 19B calculate the coherence showing the correlation between the estimated desired signal spectrum in the frequency domain of the estimated desired signal and the calculated noise signal spectrum in the frequency domain of the calculated noise signal. The Wiener filter adaptation coefficient calculation unit 20 calculates the Wiener filter adaptation coefficient that contributes to the filter coefficient for each frequency according to the coherence. Then, the Wiener filter calculation unit 15 calculates the Wiener filter coefficient to be used in the next time frame by using the Wiener filter adaptation coefficient.

In the above configuration, the Wiener filter coefficient is calculated according to the magnitude of the coherence between the estimated desired signal obtained for each frequency and the calculated noise signal, and the noise component of the observed signal is suppressed by the Wiener filter using this Wiener filter coefficient. .. As a result, the intensity of noise suppression can be appropriately selected for each frequency band to execute noise suppression, and musical noise can be reduced while maintaining the accuracy of noise suppression as much as possible.

Further, the signal processing unit 10 includes a subtraction unit 18B for calculating the difference between the observed signal spectrum and the estimated desired signal spectrum in the frequency domain, and the subtracting unit 18B calculates by the difference between the observed signal spectrum and the estimated desired signal spectrum. Calculate the noise signal spectrum. The coherence calculation unit 19B calculates coherence from the estimated desired signal spectrum and the calculated noise signal spectrum. As a result, the coherence between the estimated desired signal and the calculated noise signal for each frequency bin is calculated using the calculated noise signal spectrum calculated in the frequency domain, and the winner whose noise suppression intensity is varied for each frequency according to the coherence. The filter coefficient can be calculated. Therefore, in addition to the effect of reducing musical noise while maintaining the accuracy of noise suppression as much as possible, it is possible to simplify the configuration of the coherence calculation unit and reduce the amount of calculation by calculating coherence in the frequency domain. It will be possible.

Further, the signal processing unit 10 includes a noise signal calculation unit 22 that calculates a noise signal spectrum calculated by the observed signal spectrum and the winner filter coefficient in the frequency domain. The coherence calculation unit 19B calculates coherence from the estimated desired signal spectrum and the calculated noise signal spectrum. As a result, the coherence between the estimated desired signal and the calculated noise signal for each frequency bin is calculated using the calculated noise signal spectrum calculated in the frequency domain, and the winner whose noise suppression intensity is varied for each frequency according to the coherence. The filter coefficient can be calculated. Therefore, in addition to the effect of reducing musical noise while maintaining the accuracy of noise suppression as much as possible, the amount of calculation can be further reduced by calculating the calculated noise signal spectrum using the calculation result of the Wiener filter coefficient. , It is possible to shorten the processing time of the signal processing unit.

Further, in the coherence calculation unit 19B, the power spectrum calculation unit 193 calculates the power spectrum of the estimated desired signal spectrum, and the power spectrum calculation unit 194 calculates the power spectrum of the calculated noise signal spectrum. The cross spectrum calculation unit 195 calculates the cross spectrum between the estimated desired signal spectrum and the calculated noise signal spectrum. Then, the coherence calculation unit 196 calculates the coherence from the power spectrum of the estimated desired signal spectrum, the power spectrum of the calculated noise signal spectrum, and the cross spectrum. This makes it possible to calculate the coherence between the estimated desired signal and the calculated noise signal for each frequency bin in the frequency domain.

Further, the signal processing unit 10 includes a subtraction unit 18A for calculating the difference between the observed signal and the estimated desired signal in the time domain, and the subtracting unit 18A calculates the calculated noise signal based on the difference between the observed signal and the estimated desired signal. To do. The coherence calculation unit 19A calculates the coherence from the estimated desired signal and the calculated noise signal. As a result, the coherence between the estimated desired signal and the calculated noise signal for each frequency bin is calculated using the calculated noise signal calculated in the time domain, and the noise suppression intensity is varied for each frequency according to the coherence. The coefficient can be calculated. Therefore, it is possible to reduce musical noise while maintaining the accuracy of noise suppression as much as possible.

Further, in the coherence calculation unit 19A, the FFT unit 191 acquires the estimated desired signal spectrum obtained by converting the estimated desired signal into the frequency domain, and the FFT unit 192 acquires the calculated noise signal spectrum obtained by converting the calculated noise signal into the frequency domain. To do. The power spectrum calculation unit 193 calculates the power spectrum of the estimated desired signal spectrum, and the power spectrum calculation unit 194 calculates the power spectrum of the calculated noise signal spectrum. The cross spectrum calculation unit 195 calculates the cross spectrum between the estimated desired signal spectrum and the calculated noise signal spectrum. Then, the coherence calculation unit 196 calculates the coherence from the power spectrum of the estimated desired signal spectrum, the power spectrum of the calculated noise signal spectrum, and the cross spectrum. Thereby, the estimated desired signal and the calculated noise signal can be converted from the time domain to the frequency domain, and the coherence between the estimated desired signal and the calculated noise signal for each frequency bin can be calculated.

Although various embodiments have been described above with reference to the drawings, it goes without saying that the present disclosure is not limited to such examples. It is clear that a person skilled in the art can come up with various modification examples, modification examples, replacement examples, addition examples, deletion examples, and equal examples within the scope of claims. It is understood that it naturally belongs to the technical scope of the present disclosure. In addition, each component in the various embodiments described above may be arbitrarily combined as long as the gist of the invention is not deviated.

The present disclosure is useful as a signal processing device, a signal processing system, and a signal processing method capable of further improving the suppression of noise components of a signal.

10, 10A, 10B, 10C Signal processing unit 11, 14, 191, 192 FFT unit 12 Average and signal coupling unit 13 Input average storage unit 15 Wiener filter calculation unit 16 Estimated desired signal spectrum calculation unit 17 IFFT unit 18A, 18B Subtraction unit 19A, 19B Coherence calculation unit 20 Wiener filter adaptation coefficient calculation unit 21 Wiener filter adaptation coefficient storage unit 22 Noise signal calculation unit 193, 194 Power spectrum calculation unit 195 Cross spectrum calculation unit 196 Coherence calculation unit

Claims (8)

A signal processing device that performs signal processing on a signal to be processed.
An observation signal acquisition unit that acquires an observation signal spectrum obtained by converting an observation signal input as a processing target into a frequency domain.
A long-term average of the observed signal for a predetermined period or longer is calculated, a stationary noise component in the observed signal is estimated from the calculated average value, and an estimated noise spectrum obtained by converting the estimated noise component into a frequency domain is obtained. Estimated noise acquisition unit and
A filter coefficient calculation unit that calculates a filter coefficient for suppressing a noise component of the observation signal based on the observation signal spectrum and the estimated noise spectrum.
A desired signal estimation unit that calculates an estimated desired signal in the observed signal by filter processing using the filter coefficient, and
A signal output unit that outputs a signal obtained by converting the estimated desired signal into a time domain as an observation signal after signal processing.
A noise signal calculation unit that calculates a noise component in the current observation signal as a calculation noise signal based on the observation signal and the estimation desired signal or the filter coefficient.
A coherence calculation unit that calculates the coherence indicating the correlation between the estimated desired signal spectrum in the frequency domain of the estimated desired signal and the calculated noise signal spectrum in the frequency domain of the calculated noise signal.
It has a filter adaptation coefficient calculation unit that calculates a filter adaptation coefficient that contributes to the filter coefficient for each frequency according to the coherence.
The filter coefficient calculation unit calculates the filter coefficient using the filter adaptation coefficient.
Signal processing device. The signal processing device according to claim 1.
The noise signal calculation unit
In the frequency domain, a subtraction unit for calculating the difference between the observed signal spectrum of the observed signal and the estimated desired signal spectrum of the estimated desired signal is included, and the calculated noise signal spectrum of the calculated noise signal is calculated from the difference.
The coherence calculation unit
The coherence is calculated from the estimated desired signal spectrum and the calculated noise signal spectrum.
Signal processing device. The signal processing device according to claim 1.
The noise signal calculation unit
In the frequency domain, the calculated noise signal spectrum of the calculated noise signal is calculated from the observed signal spectrum of the observed signal and the filter coefficient.
The coherence calculation unit
The coherence is calculated from the estimated desired signal spectrum of the estimated desired signal and the calculated noise signal spectrum.
Signal processing device. The signal processing device according to claim 2 or 3.
The coherence calculation unit
A first power spectrum calculation unit for calculating the power spectrum of the estimated desired signal spectrum, and
A second power spectrum calculation unit that calculates the power spectrum of the calculated noise signal spectrum, and
A cross spectrum calculation unit that calculates a cross spectrum between the estimated desired signal spectrum and the calculated noise signal spectrum,
It has a power spectrum of the estimated desired signal spectrum, a power spectrum of the calculated noise signal spectrum, and a coherence calculation unit for calculating the coherence based on the cross spectrum.
Signal processing device. The signal processing device according to claim 1.
The noise signal calculation unit
In the time domain, the calculated noise signal is calculated by including a subtracting unit for calculating the difference between the observed signal and the estimated desired signal.
The coherence calculation unit
The coherence is calculated from the estimated desired signal and the calculated noise signal.
Signal processing device. The signal processing device according to claim 5.
The coherence calculation unit
A first frequency conversion unit that acquires an estimated desired signal spectrum obtained by converting the estimated desired signal into a frequency domain, and
A second frequency conversion unit that acquires a calculated noise signal spectrum obtained by converting the calculated noise signal into a frequency domain, and
A first power spectrum calculation unit for calculating the power spectrum of the estimated desired signal spectrum, and
A second power spectrum calculation unit that calculates the power spectrum of the calculated noise signal spectrum, and
A cross spectrum calculation unit that calculates a cross spectrum between the estimated desired signal spectrum and the calculated noise signal spectrum,
It has a power spectrum of the estimated desired signal spectrum, a power spectrum of the calculated noise signal spectrum, and a coherence calculation unit for calculating the coherence based on the cross spectrum.
Signal processing device. A signal processing system that performs signal processing on the signal to be processed.
An input unit for inputting the signal to be processed and
A signal processing unit that performs signal processing that suppresses the noise component of the observation signal with respect to the observation signal input as a processing target from the input unit.
It has an output unit that outputs a signal after signal processing by the signal processing unit.
The signal processing unit
An observation signal acquisition unit that acquires an observation signal spectrum obtained by converting the observation signal into a frequency domain,
A long-term average of the observed signal for a predetermined period or longer is calculated, a stationary noise component in the observed signal is estimated from the calculated average value, and an estimated noise spectrum obtained by converting the estimated noise component into a frequency domain is obtained. Estimated noise acquisition unit and
A filter coefficient calculation unit that calculates a filter coefficient for suppressing a noise component of the observation signal based on the observation signal spectrum and the estimated noise spectrum.
A desired signal estimation unit that calculates an estimated desired signal in the observed signal by filter processing using the filter coefficient, and
A signal output unit that outputs a signal obtained by converting the estimated desired signal into a time domain as an observation signal after signal processing.
A noise signal calculation unit that calculates a noise component in the current observation signal as a calculation noise signal based on the observation signal and the estimation desired signal or the filter coefficient.
A coherence calculation unit that calculates the coherence indicating the correlation between the estimated desired signal spectrum in the frequency domain of the estimated desired signal and the calculated noise signal spectrum in the frequency domain of the calculated noise signal.
It has a filter adaptation coefficient calculation unit that calculates a filter adaptation coefficient that contributes to the filter coefficient for each frequency according to the coherence.
The filter coefficient calculation unit calculates the filter coefficient using the filter adaptation coefficient.
Signal processing system. A signal processing method in a signal processing device that performs signal processing on a signal to be processed.
The step of acquiring the observation signal spectrum obtained by converting the observation signal input as the processing target into the frequency domain, and
The average of the observed signal for a long period of time or longer is calculated, the stationary noise component of the observed signal is estimated from the calculated average value, and the estimated noise spectrum obtained by converting the estimated noise component into the frequency domain is acquired. Steps and
A step of calculating a filter coefficient for suppressing a noise component of the observed signal based on the observed signal spectrum and the estimated noise spectrum, and
A step of calculating an estimated desired signal in the observed signal by a filter process using the filter coefficient, and
A step of outputting the signal obtained by converting the estimated desired signal into the time domain as an observation signal after signal processing, and
A step of calculating a noise component in the observed signal in the current time frame as a calculated noise signal based on the observed signal and the estimated desired signal or the filter coefficient.
A step of calculating the coherence indicating the correlation between the estimated desired signal spectrum in the frequency domain of the estimated desired signal and the calculated noise signal spectrum in the frequency domain of the calculated noise signal.
A step of calculating a filter adaptation coefficient that contributes to the filter coefficient of the next time frame for each frequency according to the coherence.
It comprises a step of calculating the filter coefficient of the next time frame using the filter adaptation coefficient.
Signal processing method.

Images