JP4460256B2 - Noise reduction processing method, apparatus for implementing the method, program, and recording medium - Google Patents

Description

  The present invention relates to a noise reduction processing method, an apparatus, a program, and a recording medium for performing the method, and in particular, a voice noise mixture in which a target voice signal and an unnecessary noise signal are mixed in a voice communication apparatus such as a voice communication apparatus. The present invention relates to a noise reduction processing method for reducing a noise signal from a signal, an apparatus for implementing this method, a program, and a recording medium.

FIG. 3 shows a schematic diagram of a loudspeaker communication system when the transmitter side is the near end side and the receiver side is the far end side. In FIG. 3, 31 is a near-end microphone, 32 is a near-end speaker, 33 is a far-end microphone, 34 is a far-end speaker, 35 is a transmission line, 36 is a transmitter, and 37 is a receiver. . The transmitted voice uttered by the speaker 36 is transmitted to the receiver 37 via the near-end microphone 31, the transmission path 35, and the far-end speaker 34. This loudspeaker communication system is widely used for communication conferences, videophones, loudspeakers, and the like.
On the other hand, when ambient noise other than the target audio signal is mixed when receiving the sound by the near-end microphone 31, the clarity of the talker voice emitted from the far-end side speaker 34 is impaired, and the voice quality is significantly deteriorated. To do. For this reason, it is required to reduce ambient noise other than the target voice included in the transmitted signal by using a noise reduction device between the near-end microphone 31 and the far-end speaker 34.

As a conventional noise reduction processing method, for example, there is a method disclosed in Patent Document 1. Details of this technique are disclosed in the publication, and will be described briefly with reference to FIG. The microphone (not shown) includes a target utterance signal S (k) obtained by the utterance of a speaker located away from the microphone, an unnecessary noise signal N (k) generated due to air conditioning noise and other external sounds. ) Is collected as the input signal X (k) = S (k) + N (k). Here, k is an integer value representing the time representation of the signal as discrete time. The input signal X (k) is input to the first frequency domain converter 51, and is converted into the frequency domain signal X (ω) by, for example, discrete Fourier transform every short time. ω represents a frequency. The input signal power spectrum calculation unit 52 receives the frequency domain signal X (ω) and calculates the power spectrum Pav _x (ω). The noise power spectrum estimation unit 53 receives the frequency domain signal X (ω) and estimates the noise power spectrum Pav _N (ω) therein. The loss calculator 54 receives Pav _x (ω) and Pav _N (ω), calculates a loss value L (ω), and transfers the result to the loss inserter 55. The loss insertion unit 55 uses the loss value L (ω) calculated by the loss calculation unit 54 to calculate Y (ω) = L (ω) × X (ω), thereby reducing the output Y (ω). Is output. The time domain conversion unit 56 receives Y (ω), converts it into a time domain signal, and outputs a time domain signal Y (k) with reduced noise.
JP-A-9-258792

  As a conventional technique for reducing the noise of an input signal mixed with ambient noise, there is a method of estimating the noise power spectrum and inserting a loss value corresponding to the noise power spectrum into the input signal to reduce the noise. In this method, if the noise power spectrum is accurately estimated and the noise reduction amount is increased, the improved S / N ratio is increased, but noise estimation errors always exist, so that residual noise and excessive audio signals are generated. The audio signal is distorted. Since these change with time, there is a problem of generating unfavorable sound. On the other hand, if the estimated noise power spectrum is made small and the estimated noise reduction amount is made small in order to suppress the distortion of the audio signal due to processing, there is a problem that the improved SN ratio becomes low and the mixed noise cannot be reduced sufficiently.

  The present invention focuses on masking near-end side noise generated from a far-end speaker in a loudspeaker apparatus by using the near-end ambient noise as a masker and masking the amount of near-end noise. If the noise amount is equal to or greater than the threshold value, the difference between the noise amount and the masking threshold value is used as a corrected noise amount. The present invention provides a noise reduction processing method that suppresses signal quality degradation by suppressing signal distortion and maintaining auditory noise reduction, and an apparatus, a program, and a recording medium that implement this method.

  [Claim 1] In a noise reduction processing method of inputting a voice noise mixed signal in which a target voice signal and an unnecessary noise signal are mixed and outputting a noise reduced signal in which the noise signal is reduced, the noise reduced signal is reproduced. A noise reduction processing method for estimating a masking threshold value formed by ambient noise in a certain environment and reducing only an amount of mixed noise exceeding the estimated masking threshold value is constructed.

In the noise reduction processing method, the input voice noise mixed signal is converted into a frequency domain signal, and the signal power spectrum of the voice noise mixed signal is calculated with respect to the voice noise mixed signal converted into the frequency domain signal. Calculates the noise power spectrum of the mixed audio and noise signal using the mixed audio and noise signal converted to the frequency domain signal, while the frequency of the reproduced environment collected signal collected in the environment where the noise-reduced signal is reproduced. The noise power spectrum of the reproduction environment sound collection signal is calculated using the reproduction environment sound collection signal converted into the frequency domain signal and converted into the frequency domain signal, and the frequency domain masking is performed from the noise power spectrum of the reproduction environment sound collection signal. Calculate the threshold, compare the masking threshold with the noise power spectrum of the mixed speech and noise signal, and mix the speech and noise mixed signal exceeding the masking threshold Calculate the corrected noise power spectrum, use the input noise power spectrum of the mixed voice noise signal and the corrected noise power spectrum of the mixed voice noise signal to predict the ratio of the corrected mixed noise signal noise and calculate the loss value based on this. A noise reduction process that calculates and converts a noise-reduced signal into a frequency-domain signal, inserts a loss value, outputs a noise-reduced signal, and converts the noise-reduced signal into a time-domain audio signal Configured the method.
Claim 2 : In a noise reduction processing apparatus for inputting a voice noise mixed signal in which a target voice signal and an unnecessary noise signal are mixed and outputting a noise reduced signal in which the noise signal is reduced, the noise reduced signal The noise reduction processing apparatus is configured to estimate the masking threshold value formed by the ambient noise in the environment where the noise is reproduced, and to reduce only the amount of mixed noise exceeding the masking threshold value.

In the noise reduction processing apparatus, a first frequency domain conversion unit 51 that converts an input voice noise mixed signal into a frequency domain signal, and voice noise with respect to the voice noise mixed signal converted into the frequency domain signal. The noise power spectrum of the voice noise mixed signal is estimated using the input signal power spectrum calculation unit 52 of the voice noise mixed signal for calculating the signal power spectrum of the mixed signal and the voice noise mixed signal converted into the frequency domain signal. A noise power spectrum estimator 11 for mixed audio noise signals, a second frequency domain converter 12 for converting a reproduction environment sound collected in an environment where a noise-reduced signal is reproduced, to a frequency domain signal, The noise power spectrum of the reproduction environment sound collection signal is estimated by using the reproduction environment sound collection signal converted into the frequency domain signal. The tone estimation unit 13, the masking threshold estimation unit 14 for estimating the masking threshold in the frequency domain from the reproduction environment sound pickup signal noise power spectrum, the masking threshold and the noise power spectrum of the voice noise mixed signal are compared, and the masking threshold is exceeded. The corrected noise power is calculated using the corrected noise power spectrum calculation unit 15 of the mixed voice noise signal 15 for calculating the mixed noise power spectrum of the voice noise and the input signal power spectrum of the mixed voice noise signal and the corrected noise power spectrum of the mixed voice noise signal. A loss calculation unit 16 that predicts a ratio of mixed signal noise and determines a loss value based on the ratio, and a signal in which a loss value is inserted into a voice noise mixed signal converted into a frequency domain signal to reduce noise A loss insertion unit 55 that outputs a signal, and a time domain variable that converts a noise-reduced signal into a time domain signal To constitute a noise reduction processing device including a part 56.

Further, according to claim 3: Convert the audio noise mixed signal inputted into a frequency domain signal, to calculate the signal power spectrum of the audio noise mixed signal to the audio noise mixed signal converted into a frequency domain signal, the frequency Calculates the noise power spectrum of the mixed audio and noise signal using the mixed audio and noise signal converted to the signal in the domain, and uses the frequency environment signal as the playback environment collected signal collected in the environment where the noise-reduced signal is reproduced. The noise power spectrum of the playback environment sound pickup signal is calculated using the playback environment sound pickup signal converted into the frequency domain signal, and the frequency domain masking threshold is calculated from the noise power spectrum of the playback environment sound pickup signal. The masking threshold is compared with the noise power spectrum of the voice noise mixed signal, and the corrected noise power spectrum of the voice noise mixed signal exceeding the masking threshold is compared. Calculate the spectrum, predict the ratio of the corrected voice noise mixed signal noise using the input noise power spectrum of the voice noise mixed signal and the corrected noise power spectrum of the voice noise mixed signal, calculate the loss value based on this, Inserts a loss value into a voice / noise mixed signal converted to a frequency domain signal, outputs a noise-reduced signal, and converts the noise-reduced signal into a time-domain voice signal. Configured a noise reduction processing program.
Furthermore, according to claim 4: forming the recording medium noise reduction processing program is recorded as described in claim 3.

  The present invention uses the above-described correction noise amount to reduce only the near-end side mixed noise amount exceeding the masking threshold for masking the far-end side ambient noise, thereby distorting the audio signal due to the noise estimation error, The problem in noise reduction processing that caused audio quality degradation on hearing was solved. The near-end mixed noise that is lower than the masking threshold is not reduced, so the improved SN ratio cannot be increased, but the near-end mixed noise at the far end is not audible below the detection limit. There is an effect of sufficiently reducing noise. Further, since the near-end side mixed noise that is equal to or less than the masking threshold is not reduced, the amount of noise reduction is reduced and the noise estimation error portion is not reduced, so that the occurrence of distortion of the audio signal can be reduced. That is, according to the present invention, it is possible to suppress deterioration in voice quality due to noise reduction processing while reducing noise in terms of hearing.

The best mode for carrying out the invention will be described with reference to the embodiment of FIG.
Reference numeral 11 denotes a noise power spectrum estimation unit, 12 denotes a second frequency domain conversion unit, 13 denotes a noise power spectrum estimation unit, 14 denotes a masking threshold estimation unit, and 15 denotes a corrected noise power spectrum calculation unit. For the rest, the same common symbols as in FIG. 5 were used.
The operation of the noise reduction device 10 according to the present invention will be described. First, an input signal X (k) mixed with a target signal on the near end side and unnecessary ambient noise or the like is converted into a frequency domain signal X (ω by a discrete Fourier transform, for example, every short time in the first frequency domain conversion unit 51. ). Here, the signal converted into the frequency domain is generally a complex number, and X (ω) = X _r (ω) + jX _i (ω) (where X _r and X _i are respectively the real part of X (ω) and Imaginary part). The frequency domain signal X <ω) is transferred to the input signal power spectrum calculation unit 52, the noise power spectrum estimation unit 11, and the loss insertion unit 55. Hereinafter, the flow of processing for X (ω) will be described.

The input signal power spectrum calculation unit 52 calculates the power spectrum P _x (ω) = (X _r (ω)) ² + (X _i (ω)) ^{2 of} the transferred frequency domain signal X (ω). . Next, the power spectrum P _x (ω) is averaged over a predetermined time and transferred to the loss value calculation unit 16 as Pav _x (ω). For example, in the processing for each frame, if the current processing frame is (·) _n and the average time is m frames,
Pav _x , _n (ω) = (1 / A) Σ _m γ _m P _x , _nm (ω)
Is calculated. Here, γ _m is an exponential weighting coefficient expressed as, for example, γ _m = (γ) ^m , (γ <1), and A is a normalization of (1 / A) Σ _m γ _m = 1. Is a constant for

The noise power spectrum estimation unit 11 estimates the noise power spectrum Pav _N (ω) using the transferred frequency domain signal X (ω) and sends it to the corrected noise power spectrum calculation unit 15. For the noise power spectrum estimation, for example, a method of updating the minimum value and estimating the value as the noise power spectrum is used. In this method, first, the power spectrum P _x (ω) of the frequency domain signal X (ω) transferred from the first frequency domain converter 51 is obtained. Next, the power spectrum P _x (ω) is averaged over a predetermined time to obtain Pav ′ _x (ω). Next, the average power spectrum Pav ′ _x , _n (ω) of the current processing frame obtained is compared with Pav ′ _x , _n−1 (ω) one frame before, and Pav ′ _x , _n (ω) is Noise power spectrum Pav _N , _n (ω) = Pav ′ _x , _n (ω) if small, Pav _N , _n (ω) = Pav ′ _x , _n−1 (ω) Pav _N , _n (ω ) Is estimated as the noise power spectrum Pav _N , _n (ω).

On the other hand, the input signal X _r (k), which is received by the far-end microphone and mixed with unnecessary ambient noise or the like of the far-end target signal, is transferred to the second frequency domain converter 12 of the noise processing device. Is done. In the second frequency domain transforming unit 12, for example, the frequency domain signal X _r (ω) is transformed by discrete Fourier transform every short time. X _r (ω) is sent to the noise power spectrum estimation unit 13.
The noise power spectrum estimation unit 13 estimates the noise power spectrum Pavr _N (ω) using the transferred X _r (ω). For the noise power spectrum estimation, for example, the same minimum value as that of the noise power spectrum estimation unit 11 is updated and the value is estimated as the noise power spectrum. The obtained noise power spectrum Pavr _N (ω) is sent to the masking threshold value estimation unit 14.

The masking threshold value estimation unit 14 obtains the amount that the ambient noise on the receiving side masks. That is, the noise power spectrum Pavr _N (ω) transferred from the noise power spectrum estimation unit 13 becomes a masker, and a threshold value P _t (ω) for masking is obtained. Assuming that the ambient noise is stationary noise such as an air conditioner operating sound, the masking threshold value is calculated by, for example, “Hearing and Acoustic Psychology” edited by Hisao Sakai p. The pure tone masking curve described in 111-113 can be used. Thereby, the noise power spectrum Pavr _N (ω) becomes a masker, and the threshold value Pt (ω) for masking can be obtained. The obtained masking threshold value P _t (ω) is sent to the corrected noise power spectrum calculation unit 15.

The corrected noise power spectrum calculation unit 15 obtains a corrected noise power spectrum P′av _N (ω) to be reduced using the masking threshold value Pt (ω). A flow chart of the processing is shown in FIG. First, the masking threshold value Pt (ω) transferred from the masking threshold value estimation unit 14 is compared with the noise power spectrum Pav _N (ω) transferred from the noise power spectrum estimation unit 11. When the noise power spectrum Pav _N (ω) is large, the correction noise power spectrum is set to P′av _N (ω) = αPav _N (ω) −Pt (ω). Here, α is a coefficient for correcting the variance (estimation error) of the estimated noise pow spectrum.
When the noise power spectrum Pav _N (ω) is small, the mixed noise is considered to be masked, so the corrected noise power spectrum is set to P′av _N (ω) = 0. The corrected noise power spectrum P′av _N (ω) is sent to the loss calculator 16.

Loss in the calculation unit 16, the auxiliary power spectrum of positive noise power spectrum calculating unit 15 corrects the noise power spectrum the sent from P'av _N (ω) and the input signal power spectrum calculating unit input signal the sent from 52 Pav A loss value L (ω) for reducing noise is obtained from _N (ω). The loss value L (ω) is calculated by, for example, a simple spectral subtraction method: L (ω) = √ (( Pav _x (ω) ) ² − ( P′av _N (ω) ) ² ) / ( Pav _x (ω ) ) Calculated in ² . The loss value L (ω) is sent to the loss insertion unit 55.
Loss interpolation join the club 55 multiplies the loss calculation unit 16 loss value has been transferred from the L (omega) the frequency domain signals X (omega), and outputs reduced signal Y (omega) noise. Next, the time domain converter 56 receives the signal Y (ω) with reduced noise, converts it into a time domain signal, and outputs it.

FIG. 4 is a diagram showing the application of the noise reduction apparatus of the present invention to a loudspeaker. In the present invention, noise mixed on the near end side is reduced on the far end side in consideration of the masking effect on the far end side. Since this is true even if the near end side and the far end side are interchanged, the noise on both sides can be reduced by using the noise reduction device according to the present invention also on the near end side.
The processing of each block of the noise reduction device according to the present invention can be performed by a DSP (Digital Signal Processor). The program can be executed by causing the computer to execute the program. In this case, the program is recorded by being recorded on a CD-ROM, flexible disk, magnetic disk or other recording medium into a program memory in the computer. The program memory may be downloaded by communication.

The figure explaining an Example. The flowchart explaining the derivation method of the correction noise power spectrum by this invention. The figure explaining a loudspeaker apparatus. The figure explaining the loudspeaker apparatus to which the noise reduction apparatus by this invention is applied. The figure explaining a prior art example.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Noise reduction apparatus 11 Noise power spectrum estimation part 12 Frequency domain conversion part 13 Noise power spectrum estimation part 14 Masking threshold value estimation part 15 Correction noise power spectrum calculation part 16 Loss calculation part 31 Near end side microphone 32 Near end side speaker 33 Far end Side microphone 34 Far end speaker 35 Transmission path 36 Speaker 37 Receiver 50 Noise reduction device 51 Frequency domain converter 52 Input signal power spectrum calculator 53 Noise power spectrum calculator 54 Loss calculator 55 Loss inserter 56 Time domain Conversion unit

Claims (4)

This is a noise reduction processing method using a microphone and a speaker that inputs a mixed audio / noise signal that includes the target audio signal and unnecessary noise signal, and outputs a noise-reduced signal with reduced noise signal. And
Converts input audio noise mixed signal to frequency domain signal,
Calculate the signal power spectrum of the mixed audio and noise signal for the mixed audio and noise signal converted to the frequency domain signal,
Calculate the noise power spectrum of the mixed audio and noise signal using the mixed audio and noise signal converted to the frequency domain signal,
On the other hand, the reproduction environment sound collection signal collected in the environment where the noise-reduced signal is reproduced is converted into a frequency domain signal,
Calculate the noise power spectrum of the reproduction environment sound signal using the reproduction environment sound signal converted into the frequency domain signal,
Calculate the masking threshold in the frequency domain from the noise power spectrum of the reproduction environment sound signal,
Comparing the masking threshold and the noise power spectrum of the voice noise mixed signal, calculating the corrected noise power spectrum of the voice noise mixed signal exceeding the masking threshold,
Using the input signal power spectrum of the voice noise mixed signal and the corrected noise power spectrum of the voice noise mixed signal, the ratio of the corrected voice noise mixed signal noise is predicted, and based on this, the loss value is calculated.
Insert a loss value into the audio / noise mixed signal converted to the frequency domain signal, and output a signal with reduced noise,
A noise reduction processing method characterized by converting a noise-reduced signal into a time-domain audio signal. This is a noise reduction processing device using a microphone and a speaker, which inputs a voice noise mixed signal in which the target voice signal and unnecessary noise signal are mixed and outputs a noise-reduced signal with reduced noise signal. And
A first frequency domain converter that converts an input audio noise mixed signal into a frequency domain signal;
A voice noise mixed signal input signal power spectrum calculation unit for calculating a signal power spectrum of the voice noise mixed signal with respect to the voice noise mixed signal converted into the frequency domain signal;
A noise power spectrum estimator for a voice noise mixed signal that estimates a noise power spectrum of the voice noise mixed signal using the voice noise mixed signal converted into a frequency domain signal;
A second frequency domain conversion unit that converts a reproduction environment sound collection signal collected in an environment where a noise-reduced signal is reproduced, into a frequency domain signal;
A noise power spectrum estimator for a reproduction environment sound collection signal for estimating a noise power spectrum using the reproduction environment sound collection signal converted into a frequency domain signal;
A masking threshold estimator for estimating a masking threshold in the frequency domain from the reproduction environment sound pickup signal noise power spectrum;
Comparing the masking threshold and the noise power spectrum of the voice noise mixed signal and calculating the voice noise mixed signal noise power spectrum exceeding the masking threshold;
A loss calculation unit that predicts the ratio of the corrected voice noise mixed signal noise using the input signal power spectrum of the voice noise mixed signal and the corrected noise power spectrum of the voice noise mixed signal, and determines a loss value based on this,
A loss insertion unit that inserts a loss value into a voice noise mixed signal converted into a frequency domain signal and outputs a signal with reduced noise;
A noise reduction processing apparatus comprising: a time domain conversion unit that converts a noise-reduced signal into a time domain signal. A computer as a noise reduction processing device for a loudspeaker system using a microphone and a speaker, which inputs a voice / noise mixed signal in which the target voice signal and unnecessary noise signal are mixed and outputs a noise-reduced signal with reduced noise signal A noise reduction processing program for operating
Converts input audio noise mixed signal to frequency domain signal,
Calculate the signal power spectrum of the mixed audio and noise signal for the mixed audio and noise signal converted to the frequency domain signal,
Calculate the noise power spectrum of the mixed audio and noise signal using the mixed audio and noise signal converted to the frequency domain signal,
On the other hand, the reproduction environment sound collection signal collected in the environment where the noise-reduced signal is reproduced is converted into a frequency domain signal,
Calculate the noise power spectrum of the reproduction environment sound signal using the reproduction environment sound signal converted into the frequency domain signal,
Calculate the masking threshold in the frequency domain from the noise power spectrum of the reproduction environment sound signal,
Comparing the masking threshold and the noise power spectrum of the voice noise mixed signal, calculating the corrected noise power spectrum of the voice noise mixed signal exceeding the masking threshold,
Using the input signal power spectrum of the voice noise mixed signal and the corrected noise power spectrum of the voice noise mixed signal, the ratio of the corrected voice noise mixed signal noise is predicted, and based on this, the loss value is calculated.
Insert a loss value into the audio / noise mixed signal converted to the frequency domain signal, and output a signal with reduced noise,
Convert noise-reduced signals to time-domain audio signals,
A noise reduction processing program that gives commands to a computer. A recording medium on which the noise reduction processing program according to claim 3 is recorded.

Images