JPH05165495A - Noise suppressing device - Google Patents

Abstract

PURPOSE:To improve the quality of processing sounds at the noise suppressing device to suppress a noise component from an acoustic signal mixing constant noise into a required signal. CONSTITUTION:A Wiener filter estimating means 8 estimates a Wiener filter transmission function from the estimated values of an input signal power spectrum and a noise power spectrum, a filter transmission function correcting means 9 sets the value of the frequency component of the Wiener filter transmission function smaller than a threshold value equally to the threshold value so as to prevent a signal component from being missed by adding/subtracting the noise component, an impulse response control means 11 provides a time constant in an impulse response intermittently outputted from an inverse FFT 10 for every (m) sample cycles so as to prevent foreign sounds from being generated in the processing sounds by continuous changes for every samples, and a high frequency emphasizing means 3 is provided in the preceding step of the FFT 5 so as to improve accuracy for calculating a high frequency band when calculating the fixed decimal of a finite word length.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise suppressing device for suppressing a noise component from a signal in which stationary noise is mixed in an acoustic signal.

[0002]

2. Description of the Related Art An example of a conventional noise suppressor will be described below with reference to the drawings. (FIG. 2) is a block diagram of a conventional noise suppression device. In FIG. 2, 1 is an A / D converter. A shift register 2 holds the past N samples of the output signal from the A / D converter 1. Reference numeral 4 denotes a noise section detection means, which is provided in the subsequent stage of the shift register 2 and determines whether or not the current signal is only a noise component. Reference numeral 5 denotes an FFT, which is provided at the subsequent stage of the shift register 2 and converts a time series signal into a frequency spectrum. 6
Is a power spectrum conversion means for converting the output signal of the FFT 5 into a power spectrum. Reference numeral 7 is a noise power spectrum estimation means, which estimates the noise power spectrum using the output signals from the noise section detection means 4 and the power spectrum conversion means 6 as inputs. Reference numeral 8 is a Wiener filter estimation means, which estimates the transfer function of the Wiener filter using the output signals from the power spectrum conversion means 6 and the noise power spectrum estimation means 7 as inputs. An inverse FFT 10 is provided after the Wiener filter estimator 8 and outputs the impulse response of the Wiener filter. Reference numeral 12 is a convolution operation means, which performs a convolution operation on the signals from the shift register 2 and the inverse FFT 10. Reference numeral 13 denotes a D / A converter, which is a convolution operation means 12
It is provided in the subsequent stage and converts a digital signal into an analog signal.

The Wiener filter is a filter used for predicting a future value of a signal or extracting a signal buried in noise based on a past observed value of the signal (for example, "Toru Katayama: Applied Kalman filter, Asakura Shoten, P.57
~ P.70)), the conventional noise suppressor uses the FFT to estimate the impulse response of the Wiener filter that suppresses the noise from the power spectrum of the signal, and convolves the estimated impulse response with the input signal. Noise is suppressed by filtering with.

First, the A / D converter 1 converts the input signal into a digital value, and the input signal is stored in the shift register 2 having a register length N as N time-series data x (n). The noise section detecting means 4 examines the signal level, the periodicity of the signal, and the shape of the signal spectrum, and determines whether the time series data x (n) stored in the shift register 2 is only a stationary noise component. Is detected. In the FFT 5, the Fourier transform is performed on the M pieces of data from the newest data of the time series data stored in the shift register 2 and a complex spectrum of the number of data M is output (where M ≦ N). In the power spectrum conversion means 6, FFT5
Power spectrum X (ω) of the complex spectrum output from
Is required. The noise power spectrum estimating means 7 outputs X from the power spectrum converting means 6 when the noise section detecting means 4 detects that the input signal is a noise component only.
The estimated noise spectrum N (ω) is obtained by averaging (ω). In the Wiener filter estimation means 8, from the power spectrum X (ω) of the input signal which is the output signal from the power spectrum conversion means 6 and the average noise power spectrum N (ω) which is the output signal from the noise power spectrum estimation means 7. The transfer function W (ω) of the Wiener filter is obtained. When the power spectra of the signal component and the noise component of the input signal are S ₀ (ω) and N ₀ (ω), the Wiener filter is generally given by (Equation 1) (for example, “Toru Katayama: Applied Kalman filter, Asakura Shoten, P.60, Formula (4.19) ”).

[0005]

[Equation 1]

Here, the information that can be actually obtained from the input signal is the input signal power spectrum X (ω) (= S
₀ (ω) + N ₀ (ω)) and the estimated noise spectrum N (ω) (m ² N ₀
Since (ω)), (Equation 1) is approximately obtained by (Equation 2).

[0007]

[Equation 2]

However, β is a constant for controlling the amount of noise suppression and 0 <β. When the noise power spectrum N (ω) can be accurately obtained at each time, if β = 1, the optimal Wiener filter W (ω) is obtained, but the noise component included in the actual acoustic signal is non-stationary. On the other hand, since the estimated noise power spectrum N (ω) is a representative value obtained by averaging over a long period of time, it cannot be said that N (ω) is always accurate. To be selected. For example, if the signal component is voice, increasing β increases the noise suppression effect but also decreases the voice intelligibility, and in the case where the intelligibility is emphasized in the relationship between the noise suppression amount and the voice intelligibility, β is set small. In the inverse FFT 10, the inverse Fourier transform of the Wiener filter transfer function W (ω) is performed, and the impulse response h (n) of the Wiener filter is obtained. In the convolution operation means 12, the impulse response h (n)
And the input time series signal x (n) stored in the shift register 2
And a convolution operation is performed and the result is output to the D / A converter 13 and converted into an analog signal.

[0009]

However, in the above-mentioned configuration, the optimum Wiener filter characteristics change from moment to moment in accordance with changes in the spectrum of the signal component and the noise component, so that the impulse response h (n) is sampled. It is desirable to be updated every cycle. However, the process from the noise detection means 4 for obtaining the impulse response h (n) of the Wiener filter to the FFT 5, the power spectrum conversion means 6, the noise power spectrum estimation means 7, the Wiener filter estimation means 8 and the inverse FFT 10 is one sample. The calculation time is considerably longer than the cycle, and in the current real-time processing, the impulse response h
Since (n) cannot be updated, impulse response h
When the change amount of (n) becomes large, the signal waveform is distorted, resulting in abnormal noise in the processed sound, and when the noise component has non-stationarity, the noise power spectrum N Since the estimation error of (ω) becomes large, the error of the impulse response h (n) also increases, and the abnormal noise of the processed sound becomes larger.

In view of the above-mentioned problems, the present invention provides a noise suppression apparatus for obtaining a processed sound of high quality in which no abnormal noise is generated in a processed signal.

[0011]

In order to solve the above problems, a noise suppressing device of the present invention is an A / D converter for converting an input signal into a digital signal, and an output signal from the A / D converter. Of the past N samples, noise section detection means for judging whether or not the input signal is in the noise section from the time series data held in the shift register, and held in the shift register. An FFT for obtaining a frequency spectrum of time series data, and the FF
Power spectrum conversion means for converting the output of T into a power spectrum; noise power spectrum estimation means for estimating a noise power spectrum using the output signals from the power spectrum conversion means and the noise section detection means as inputs; and the power spectrum conversion. Means and an output signal from the noise power spectrum estimating means as an input, a Wiener filter estimating means for estimating a Wiener filter transfer function, an inverse FFT for obtaining an impulse response from the output of the Wiener filter estimating means by an inverse Fourier transform, and the inverse Impulse response control means for outputting an impulse response that changes with a time constant with respect to the temporal change of the impulse response output from the FFT, and convolution operation of the time series data of the shift register and the output from the impulse response control means. A convolution operator When the output of the convolution arithmetic means is obtained and a D / A converter for converting the analog signal.

[0012]

According to the present invention, by properly selecting the time constant of the impulse response control means, the impulse response of the Wiener filter, which has been intermittently changed by a large amount of change, continuously changes every sample. As a result, the discontinuity of the processed signal disappears and abnormal noise does not occur. Further, even if there is a large estimation error in the impulse response output from the inverse FFT, if some preceding and following impulse responses are estimated almost correctly, the smoothing effect according to the time constant may cause deterioration in the quality of the processed sound. It will be alleviated.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A noise suppressing device according to an embodiment of the present invention will be described below with reference to the drawings. (FIG. 1) is a block diagram of a noise suppressing apparatus according to an embodiment of the present invention. In FIG. 1, the A / D converter 1 and the shift register 2 are the same as those of the conventional noise suppression device described above. Reference numeral 3 denotes a high frequency emphasizing means for emphasizing the high frequency component of the time series data stored in the shift register 2. Reference numeral 4 denotes a noise section detecting means, which is provided after the high frequency emphasizing means 3 and determines whether or not the current signal is only a noise component. Reference numeral 5 denotes an FFT, which is provided after the high frequency emphasizing means 3 and converts the time series signal into a frequency spectrum. Power spectrum converting means 6, noise power spectrum estimating means 7 and Wiener filter estimating means 8
Is similar to the above-described conventional noise suppression device. Reference numeral 9 denotes a filter transfer function correction means, which is provided at the latter stage of the Wiener filter estimation means 8 and sets the amplitude value to the threshold value α for the frequency component whose amplitude value of the Wiener filter transfer function is smaller than the threshold value α. An inverse FFT 10 is provided after the filter transfer function correcting means 9 and outputs the impulse response of the Wiener filter. An impulse response control unit 11 controls the impulse response so as to follow the impulse response output from the inverse FFT 10 with a time constant. Reference numeral 12 denotes a convolution operation means, which performs a convolution operation on the impulse response of the Wiener filter from the impulse response control means 11 and the time series data of the input signal stored in the shift register 2. A D / A converter 13 converts the output signal of the convolution operation means 12 into analog.

The operation of the noise suppressor configured as above will be described below (FIG. 1). In FIG. 1, the A / D converter 1 and the shift register 2 operate similarly to the above-described conventional noise suppression device. The high-frequency emphasis unit 3 emphasizes the high-frequency range of the latest M data among the N time-series data stored in the shift register 2 and outputs M data strings. When the present embodiment is implemented by fixed point arithmetic, the high-frequency emphasizing means 3 described above generally obtains a power spectrum and a filter transfer function because the spectrum of the acoustic signal generally has a spectrum inclination in which the high frequency side is lowered. This is provided in order to solve the problem that a high-frequency component having a small amplitude is cut off during the calculation of, and a sufficient calculation accuracy cannot be obtained. However, if the frequency characteristic of the high-frequency emphasizing means 3 is C (ω), the Wiener filter transfer function W (ω) becomes as in (Equation 3), and the numerator and denominator C (ω) cancel each other out. Therefore, it does not affect W (ω).

[0015]

[Equation 3]

The noise section detecting means 4, the FFT 5, the power spectrum converting means 6, the noise power spectrum estimating means 7 and the Wiener filter estimating means 8 operate in the same manner as the above-mentioned conventional noise suppressing apparatus. Filter transfer function correction means 9
Is for the frequency component smaller than the threshold value α such that the amplitude value of the Wiener filter transfer function W (ω), which is the output from the Wiener filter estimator 8, satisfies the expression (4). Limit the value.

[0017]

[Equation 4]

[0018]

[Equation 5]

The above-mentioned filter transfer function correction means 9 is
It has a meaning to prevent the loss of information of the signal component generated by the noise suppression processing. The missing information of the signal component is the estimated value N of the noise power spectrum in the subtraction part of the numerator part of (Equation 2).
This occurs when (ω) multiplied by a constant β is larger than the instantaneous true noise power spectrum N ₀ (ω). N (ω) is an estimated value obtained by averaging the input signal power spectrum X (ω) in the noise section, and if the true value N ₀ (ω) has strong stationarity, the error is small, but the noise nonstationarity is strong. When the error becomes large N
_{The state of 0} (ω) << βN (ω) frequently occurs and the spectrum of the signal component is scraped. In the past, the value of β was adjusted to prevent the loss of signal components, but it is necessary to change it according to the signal-to-noise ratio of the input signal or reset the value according to the strength of non-stationarity of noise. However, since the processing is complicated and the stability is poor, W at the frequency ω such that N ₀ (ω) << βN (ω) is satisfied.
A process in which the value of (ω) becomes very small is limited by the threshold value α to prevent the loss of the signal component.

The inverse FFT 10 obtains the impulse response h (n) of the Wiener filter by inverse Fourier transform of the transfer function W (ω) corrected by the filter transfer function correction means 9. Here, the processing from the high frequency emphasizing means 3 to the inverse FFT 10 requires a large amount of calculation, and it is difficult to find the optimum impulse response h (n) for each sampling period. Therefore, the impulse response h (n) is calculated every m samples. Ask for.

The impulse response control means 11, m samples each on the obtained impulse response h (n) as an input, when the impulse response changes as constant closer to each sample h (n) with h _a (n) output To do. The above-mentioned processing of the impulse response control means 11 can be performed with a small calculation amount as in (Equation 6).

[0022]

[Equation 6]

Where h _a (n) _k is the output impulse response h _a (n) of the impulse response control means 11 at time k, h (n) is the impulse response output from the inverse FFT 10 every m sample periods, and γ Is _a constant that determines the update time constant of h _a (n).

The above-mentioned impulse response control means 11 is
A function that continuously changes the impulse response, which has changed every m sample periods, and further absorbs the estimation error of the sudden impulse response by smoothing to obtain a processed sound with no distortion and less distortion. With.

The convolution operation means 12 performs a convolution operation of the impulse response ha (n) which is the output of the impulse response control means 11 and the input time series signal x (n) stored in the shift register 2, and D / A The converter 13 converts the signal from the convolution operation means 12 into an analog signal and outputs it.

As described above, according to this embodiment, an A / D converter for converting an input signal into a digital signal, and the A / D converter
A shift register that holds the past N samples of the output signal from the D converter, and noise section detection means that determines whether or not the input signal is in the noise section from the time series data held in the shift register, The high frequency emphasizing means for extracting M samples from the newest time of the time series data held in the shift register and emphasizing the high frequency range to output M data strings, and the high frequency emphasizing means. M
The FFT for obtaining the frequency spectrum of the number M of points for each data string, the power spectrum conversion means for converting the output of the FFT into the power spectrum, and the output signals from the power spectrum conversion means and the noise section detection means. Noise power spectrum estimating means for estimating a noise power spectrum as an input, Wiener filter estimating means for estimating a Wiener filter transfer function using the output signals from the power spectrum converting means and the noise power spectrum estimating means as input, and the Wiener filter A filter transfer function correcting means for setting a frequency component having an amplitude smaller than the threshold value α of the Wiener filter transfer function output from the estimating means to the threshold value α, and the Wiener filter transfer function corrected by the filter transfer function correcting means are reversed. Inverse of Fourier transform to obtain impulse response An FFT, an impulse response control means for outputting an impulse response that changes with a time constant with respect to a temporal change of the impulse response output from the inverse FFT, time series data of the shift register, and the impulse response control means. By providing a convolution operation means for performing a convolution operation with the output and a D / A converter for converting the output of the convolution operation means into an analog signal, the impulse response is continuously changed and addition / subtraction of noise components is limited. In addition, by sufficiently obtaining the calculation accuracy, it is possible to obtain a processed sound of high quality in which the information of the signal component in which abnormal noise does not easily occur is not lost.

Although the noise section detecting means 4 is provided after the high frequency emphasizing means 3, the noise section detecting means 4 may be provided after the shift register 2.

Further, the filter transfer function correction means 9 sets the frequency component smaller than the threshold value α such that the amplitude value of the Wiener filter transfer function W (ω) which is the output from the Wiener filter estimation means 8 satisfies (Equation 4). On the other hand, although the amplitude value is limited as in (Equation 5), the filter transfer function correcting unit 9 determines that the amplitude value of the Wiener filter transfer function W (ω) output from the Wiener filter estimating unit 8 is (Equation 5). 7)
The amplitude value may be limited as in (Equation 8) for frequency components smaller than the threshold function α (ω) that satisfies

[0029]

[Equation 7]

[0030]

[Equation 8]

[0031]

As described above, according to the present invention, an A / D converter that converts an input signal into a digital signal, a shift register that holds the past N samples of the output signal from the A / D converter, Noise interval detection means for determining whether or not the input signal is in a noise interval from the time series data held in the shift register, and FFT for obtaining the frequency spectrum of the time series data held in the shift register,
A power spectrum converting means for converting the output of the FFT into a power spectrum; and a noise power spectrum estimating means for estimating a noise power spectrum using the output signals from the power spectrum converting means and the noise section detecting means as inputs.
Wiener filter estimating means for estimating a Wiener filter transfer function using the output signals from the power spectrum converting means and the noise power spectrum estimating means as inputs, and an inverse FFT for obtaining an impulse response from the output of the winner filter estimating means by inverse Fourier transform. And the inverse FFT
An impulse response control means for outputting an impulse response that changes with a time constant with respect to a temporal change of the impulse response output from, and a convolution operation of the time series data of the shift register and the output from the impulse response control means. By providing the convolution operation means for performing and the D / A converter for converting the output of the convolution operation means into an analog signal, noise that does not generate abnormal noise in the processed signal component and obtains an output signal with good clarity A suppression device can be realized.

[Brief description of drawings]

FIG. 1 is a block diagram of a noise suppression device according to an embodiment of the present invention.

FIG. 2 is a block diagram of a conventional noise suppression device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 A / D converter 2 Shift register 3 High-frequency emphasis means 4 Noise area detection means 5 FFT 6 Power spectrum conversion means 7 Noise power spectrum estimation means 8 Wiener filter estimation means 9 Filter transfer function correction means 10 Inverse FFT 11 Impulse response control Means 12 Convolution operation means 13 D / A converter

Claims (4)

[Claims] 1. A for converting an input signal into a digital signal
Whether or not the input signal is in the noise section from the / D converter, the shift register that holds the past N samples of the output signal from the A / D converter, and the time-series data held in the shift register Noise section detection means for determining
The FFT for obtaining the frequency spectrum of the time-series data held in the shift register, the power spectrum conversion means for converting the output of the FFT into a power spectrum, and the output signals from the power spectrum conversion means and the noise section detection means Noise power spectrum estimating means for estimating a noise power spectrum as an input, Wiener filter estimating means for estimating a Wiener filter transfer function using the output signals from the power spectrum converting means and the noise power spectrum estimating means as input, and the Wiener filter An inverse FFT for obtaining an impulse response from the output of the estimating means by an inverse Fourier transform, and an impulse response control means for outputting an impulse response that changes with a time constant with respect to the temporal change of the impulse response output from the inverse FFT, Shift register A noise suppressor comprising: a convolution calculation means for performing a convolution calculation of sequence data and an output from the impulse response control means; and a D / A converter for converting the output of the convolution calculation means into an analog signal. .. 2. A filter transfer function correcting means for setting a value smaller than a threshold value α to α with respect to the amplitude value of the transfer function from the Wiener filter estimating means is provided between the Wiener filter estimating means and the inverse FFT. Item 1. The noise suppression device according to Item 1. 3. The noise suppressing device according to claim 2, wherein the threshold value α of the filter transfer function correcting means differs depending on the frequency. 4. A high-frequency emphasizing means for emphasizing frequency components in the high frequency range is provided between the shift register and the FFT.
The noise suppression device described.