JP2005348173A - Noise reduction method, device for executing the same method, program and its recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a noise reduction device which reduces "muffling" of voice and sufficiently suppresses noise. <P>SOLUTION: The noise reduction device is equipped with a frequency band division part 22 which converts an inputted voice signal into a signal of a frequency region and divides the signal into a plurality of frequency bands, an inputted voice signal power calculation part 24 which calculates inputted voice signal power for each frequency from a frequency band signal of the inputted voice signal, a noise power estimation part 51 which estimates noise power for each frequency from the inputted voice signal power for each frequency, a base gain factor calculation part 61 which calculates a base gain factor from the inputted voice signal power for each frequency and the noise power for each frequency, a gain factor smoothing part 62 which smoothes the base gain factor, a gain factor insertion part 28 which superposes the smoothed gain factor on a frequency band signal of the inputted voice signal and calculates a frequency band signal of a noise reduction signal and a time region conversion part 29 which performs inverse conversion of the frequency band signal of the noise reduction signal into a time region and outputs it. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a noise reduction method, an apparatus for implementing the method, a program, and a recording medium thereof, and more particularly, to reduce a noise signal superimposed on a target voice and collected by a microphone in voice communication using a microphone. The present invention relates to a noise reduction method for improving quality, an apparatus for implementing the method, a program, and a recording medium thereof.

A conventional example of reducing a noise signal from an input voice signal in which a noise signal is superimposed on a target voice signal will be described with reference to FIG.
In FIG. 2, an input voice signal X (n) = S (n) + N (n) obtained by mixing unnecessary noise into the target voice collected by the microphone 11 is converted from analog to digital (A / D). The signal is digitized by the device 21 and input to the frequency band dividing unit 22. Note that S (n) indicates a target voice signal, and N (n) is an unnecessary noise signal mixed therein. The input audio signal X (n) after A / D conversion input to the frequency band dividing unit 22 is converted into a frequency domain signal and then divided into a plurality of frequency bands. Each frequency band signal divided by the frequency band dividing unit 22 is input to the input audio signal power calculating unit 24 and the gain factor inserting unit 28. Hereinafter, the flow of processing for the frequency band signal Xk (n) will be described using the kth frequency band signal of the input audio signal X (n) as a representative.

  The input voice signal power calculation unit 24 calculates the power level of the input frequency band signal Xk (n) and inputs the power level to the S / N ratio estimation unit 27 and the noise power estimation unit 51. The noise power estimation unit 51 estimates the noise power PNk (n) using the input voice signal power PXk (n) that has been input. In the S / N ratio estimation unit 27, the input speech signal power PXk (n), the estimated noise power PNk (n), and the signal Y′k (n) in which the noise obtained by the gain factor insertion unit 28 one processing frame before is reduced. ) Is used to estimate the target speech signal to noise signal ratio (S / N ratio) SNRk (n). The S / N ratio SNRk (n) estimated by the S / N ratio estimation unit 27 is input to the gain factor calculation unit 30 and the input audio signal addition rate determination unit 52. The gain factor calculation unit 30 determines the gain factor G (SNRk (n)) using the S / N ratio SNRk (n) input from the S / N ratio estimation unit 27. Here, the specific calculation of the gain factor G (SNRk (n)) is performed based on spectral subtraction, Wiener filter, ML estimation method, MMSE method, and other short-time spectral amplitude (STSA) estimation. The gain factor G (SNRk (n)) estimated by the gain factor calculation unit 30 is input to the gain factor insertion unit 28. The gain factor insertion unit 28 performs noise reduction using the gain factor G (SNRk (n)). Specifically, a calculation is performed to superimpose (multiply) the gain factor G (SNRk (n)) in the frequency domain on the frequency band signal Xk (n) input from the frequency band dividing unit 22. Then, the noise reduction signal Y′k (n) obtained by reducing the noise is input to the input audio signal adding unit 53 and the S / N ratio estimating unit 27. In parallel with this, the input sound signal addition rate determination unit 52 determines the input sound signal addition rate α based on the S / N ratio using the input S / N ratio SNRk (n), and the input sound. The signal is input to the signal adder 53.

In the input voice signal adding unit 53, the frequency band signal Xk (n) is added (added) to the noise reduction signal Y′k (n) at a rate according to the input voice signal addition rate α, and the following frequency band signal Yk is added. (N) is output (see Patent Document 1).
Yk (n) = αXk (n) + (1−α) Y′k (n)
The frequency band signal Yk (n) is input to the time domain conversion unit 29, and all the bands are combined, converted back to a time domain signal, and input to the digital / analog (D / A) converter 34. Here, the noise-reduced output signal Y (n) converted into an analog signal by the digital / analog converter 34 is output.
Japanese Patent No. 3454402

  In the above conventional example, the gain factor G (SNRk (n)) is calculated using the S / N ratio. The S / N ratio corresponds to the power ratio between the target voice signal and the noise signal, but the target voice signal S (n) and the noise signal N (n) are mixed in the input voice signal X (n). Therefore, each power cannot be measured independently. Therefore, in Patent Document 1, a method of estimating each power and S / N ratio is adopted, but an estimation error occurs in the estimated power and S / N ratio. Due to the influence of this estimation error, the calculated gain factor is not an ideal value based on the true S / N ratio. That is, the gain factor adjacent in the frequency domain becomes intermittent (jumps) compared with the ideal value based on the true S / N ratio, and is distorted into the processed output audio signal Y (n). It will cause to generate.

  From the above, the present invention provides a noise reduction method that suppresses the occurrence of distortion of the output audio signal after processing and realizes sufficient noise reduction even if there is an estimation error in the estimated power and S / N ratio, An apparatus, a program, and a recording medium for implementing the method are provided.

  [Claim 1] A noise reduction method for reducing a noise signal from an input voice signal, converting the input voice signal into a frequency domain signal, calculating an input voice signal power for each frequency from a frequency band signal of the input voice signal, The noise power for each frequency is estimated from the input audio signal power for each frequency, the prime gain factor is calculated from the input voice signal power for each frequency and the noise power for each frequency, the prime gain factor is smoothed, and the smoothed gain factor is input. A noise reduction method is constructed in which a frequency band signal of a noise reduction signal is calculated by being superimposed on a frequency band signal of an audio signal, and the frequency band signal of the noise reduction signal is inversely converted into a time domain signal and output.

[2] The noise reduction method according to [1], wherein a noise reduction method for performing a smoothing process of weighting and averaging the prime gain factors is configured.
[3] The noise reduction method according to any one of [1] and [2], wherein the smoothing gain factor is emphasized, and the emphasis gain factor is superimposed on the frequency band signal of the input speech signal to reduce the noise. A noise reduction method for calculating the frequency band signal of the reduced signal was constructed.
According to a fourth aspect of the present invention, in the noise reduction method according to the third aspect of the present invention, a noise reduction method is implemented in which an enhancement process is performed to bring the smoothing gain factor close to 0 or 1 depending on the value of the smoothing gain factor.

  5. A noise reduction apparatus for reducing a noise signal from an input audio signal, wherein the frequency band dividing unit 22 converts the input audio signal into a frequency domain signal and divides the signal into a plurality of frequency bands, and the frequency band signal of the input audio signal. The input voice signal power calculation unit 24 that calculates the input voice signal power for each frequency from the input, the noise power estimation unit 51 that estimates the noise power for each frequency from the input voice signal power for each frequency, and the input voice signal power for each frequency And a prime gain factor calculation unit 61 for calculating a prime gain factor from noise power for each frequency, a gain factor smoothing unit 62 for smoothing the prime gain factor, and a smoothing gain factor superimposed on the frequency band signal of the input audio signal. The gain factor insertion unit 28 for calculating the frequency band signal of the noise reduction signal and the frequency band signal of the noise reduction signal. To constitute a noise reduction apparatus comprising a time domain conversion unit 29 to output the inverse transformation in the region.

Claim 6: The noise reduction apparatus according to claim 5, wherein a noise reduction apparatus that performs a smoothing process that performs weighted addition averaging on the prime gain factor is configured.
[7] The noise reduction apparatus according to any one of [5] and [6], further comprising a gain factor emphasizing unit 63 for emphasizing the smoothing gain factor, and using the enhanced gain factor as the input audio signal. A noise reduction device is constructed that calculates the frequency band signal of the noise reduction signal superimposed on the frequency band signal.
Claim 8: The noise reduction apparatus according to claim 7, wherein the noise reduction apparatus is configured to perform enhancement processing for making the smoothing gain factor close to 0 or 1 depending on the value of the smoothing gain factor.

Claim 9: The noise reduction program which described the noise reduction method in any one of Claim 1 thru | or 4 by the code | symbol written in and read out to a computer was comprised.
Claim 10: A recording medium on which the noise reduction program according to claim 9 is recorded is configured.

  In the present invention, by performing smoothing of the gain factor, the discontinuity in the frequency domain of the gain factor is reduced and the occurrence of distortion of the output audio signal is suppressed, and by emphasizing the gain factor after smoothing, By avoiding the loss of part of the frequency component of the processed voice, the effect of reducing the “clouding” of the voice and suppressing the noise sufficiently can be achieved.

The best mode for carrying out the present invention will be described with reference to FIG.
FIG. 1 is a diagram for explaining an embodiment of a noise reduction apparatus. This embodiment is a gain factor calculation unit as compared with the noise reduction apparatus described with reference to FIG. Only the configuration of 30 is different, and the other configurations are common. From the above, the calculation of each value in the part other than the gain factor calculation unit 30 in this embodiment can be executed following the method described in Patent Document 1.
An input voice signal X (n) = S (n) + N (n) obtained by mixing unnecessary noise in the target voice collected by the microphone 11 is digitally converted by an analog / digital (A / D) converter 21. It is converted into a signal and input to the frequency band dividing unit 22. Note that S (n) indicates a target voice signal, and N (n) is an unnecessary noise signal mixed therein. In the frequency band dividing unit 22, the input audio signal X (n) is converted into a frequency domain signal and divided into a plurality of frequency bands. Each divided frequency band signal is input to the input audio signal power calculation unit 24 and the gain factor insertion unit 28. Hereinafter, the process flow for the kth frequency band signal Xk (n) will be described with the kth frequency band signal of the input audio signal as a representative. The input voice signal power calculation unit 24 calculates a power level from the input frequency band signal Xk (n), and calculates the input voice signal power PXk (n) as a calculation result from the S / N ratio estimation unit 27, noise power. Input to the estimation unit 51. The noise power estimation unit 51 estimates the noise power PNk (n) using the input voice signal power PXk (n) that has been input. In the S / N ratio estimation unit 27, the input voice signal power PXk (n), the estimated noise power PNk (n), and the frequency band signal Y′k in which the noise obtained by the gain factor insertion unit 28 before one processing frame is reduced. (N) is used to estimate the S / N ratio SNRk (n). The S / N ratio SNRk (n) estimated by the S / N ratio estimation unit 27 is input to the gain factor calculation unit 30 and the input audio signal addition rate determination unit 52.

  Here, in the embodiment of the present invention, the gain factor calculation unit 30 includes a prime gain factor calculation unit 61, a gain factor smoothing unit 62, and a gain factor enhancement unit 63. In the gain factor calculation unit 30, first, the prime gain factor G (SNRk (n)) is used by using the S / N ratio SNRk (n) input from the S / N ratio estimation unit 27 in the prime gain factor calculation unit 61. ) Is calculated. The specific calculation of the elementary gain factor is performed based on spectral subtraction, Wiener filter, ML estimation method, MMSE method and other short-time spectral amplitude (STSA) estimation. The gain factor is calculated based on the S / N ratio. Specifically, the gain factor means the ratio of each frequency domain in the frequency domain of the target voice in the input voice signal, and is calculated as it is based on the S / N ratio. The gain factor is referred to as a prime gain factor.

  Next, the gain factor smoothing unit 62 performs smoothing with a weight applied to the prime gain factor G (SNRk (n)). The prime gain factor G (SNRk (n)) calculated by the prime gain factor calculation unit 61 is ideal due to the influence of the estimation error of the S / N ratio as described in the section “Problems to be Solved by the Invention”. And a discontinuity occurs in each elementary gain factor in the frequency domain. For this reason, in this embodiment, smoothing is performed to smooth the values of the respective gain factors by relaxing the discontinuity of the elementary gain factors adjacent on the frequency axis. Hereinafter, the smoothing method will be described in detail.

When the prime gain factor G (SNRk (n)) of the kth frequency band is replaced with G (k) and the smoothed gain factor Ge (k) after smoothing is performed, an example of the smoothing process is as follows: It can be represented by the following formula.
Ge (k) = Σ _{i, j} a (i) × G (j) / Σ _i a (i)
This equation obtains an average value of a plurality of elementary gain factors G (j) adjacent to the kth frequency band indicated by the index j, and performs smoothing as a smoothing gain factor Ge (k) of the kth frequency band. Indicates processing. The total number of i and j when taking the sum is the same, and the total is at most equal to or less than the number of frequency analysis points. The weighting factor a (i) controls the influence of each elementary gain factor when calculating the average value, that is, the rate at which the discontinuity is relaxed. After the above processing, a smoothing gain factor Ge (k), that is, Ge (SNRk (n)) is output.

  Next, in the gain factor emphasizing unit 63, the smoothed gain factor Ge (k) already smoothed by the gain factor smoothing unit 62 is enhanced. The smoothing gain factor Ge (k) has no discontinuity due to the smoothing, but has a negative effect of “rounding” due to a trade-off. Specifically, when the gain factor of the kth frequency band is 1, the gain factor of the k−1 frequency band is 0.92, and the gain factor of the k + 2 frequency band is 0.93, the three When the smoothing gain coefficient of the kth frequency band from the average value in the frequency band is determined, 0.95 is obtained. In this example, the weighting factors a (i) are all set to 1.

As described above, the gain factor is calculated on the basis of the S / N ratio. Specifically, the gain factor is the ratio of each frequency domain in the frequency domain of the target speech occupied in the input speech signal. That is, when the calculated gain factor is close to 1, it means that the noise is small in the input voice signal and the ratio of the target voice is high, and when it is close to 0, the target voice in the input voice signal is small and the ratio of the noise is high. It means that there are many states. As explained above, when the smoothing gain factor is “rounded”, the value will deviate from 0 and 1. When moving away from 1, for example, when smoothed to 0.95 as in the previous example, the frequency domain component in which only the target speech originally exists is 95%, so a 5% loss occurs in the target speech. When the distance from 0 is, for example, 0.05, when only noise is present and should be reduced by 100%, the noise is reduced by 95% and the speech quality is adversely affected. Therefore, the following smoothing “smoothed” smoothing gain factor Ge (k) is enhanced. When the gain factor is Gg (k) by emphasizing the k-th in the frequency domain, the enhancement process is a process of making each gain coefficient close to 0 or 1 depending on the magnitude of the value of the smoothing gain factor Ge (k). is there. That is, when the smoothing gain factor Ge (k) is close to 1, it is closer to 1 to make it easier to pass the target speech, and when the smoothing gain factor Ge (k) is close to 0, it is more close to 0. The smoothing gain factor Ge (k) is emphasized so that the noise is greatly reduced by approaching. A specific example of this enhancement processing is shown by the following formula.
When Ge (k) is larger than th1: Gg (k) = th1 × (Ge (k) / th1) ^v1
When Ge (k) is smaller than th2:
Gg (k) = 1- (1-th2) {(1-Ge (k)) / (1-th2)} ^v2
Here, v1 (k) and v2 (k) are integers of 1 or more. Further, th1 and th2 are integers of 0 or more and 1 or less that satisfy the relationship of th1 ≧ th2. Since Ge (k) has a value in the range of 0 to 1, this expression realizes processing closer to 1 when it is larger than th1, and closer to 0 when it is smaller than th2. After the above processing, the enhanced gain factor Gg (k), that is, Gg (SNRk (n)) is output.

  The enhanced gain factor G (SNRk (n)) calculated by the gain factor emphasizing unit 63 in the gain factor calculating unit 30 is input to the gain factor inserting unit 28. The gain factor insertion unit 28 performs noise reduction using the enhanced gain factor Gg (SNRk (n)). Specifically, calculation is performed to superimpose (multiply) the enhancement gain factor Gg (SNRk (n)) in the frequency domain on the frequency band signal Xk (n) input from the frequency band dividing unit 22. Then, the noise-reduced signal Y′k (n) is input to the input audio signal adding unit 53 and the S / N ratio estimating unit 27. In parallel with this, the input sound signal addition rate determination unit 52 determines the input sound signal addition rate α based on the S / N ratio using the input S / N ratio SNRk (n), and the input sound. The signal is input to the signal adder 53.

In the input audio signal adding unit 53, the frequency band signal Xk (n) is added (added) to the signal Y′k (n) whose noise is reduced at a rate according to the input audio signal addition rate α, and the conventional example is obtained. The following frequency band signal Yk (n) as described above is output.
Yk (n) = αXk (n) + (1−α) Y′k (n)
The frequency band signal Yk (n) is input to the time domain conversion unit 29, where the entire band is synthesized and inversely converted into a time domain signal. The inversely converted time domain signal is input to the digital / analog converter 34, converted to an analog signal, and output as an output signal Y (n) with reduced noise.

  By the way, the noise reduction device of the present invention can be configured by a DSP (Digital Signal Processor). Moreover, you may make it function by making a computer run a program. In this case, the program is recorded on a CD-ROM, a floppy (registered trademark) disk, a magnetic disk, or the like by being loaded into a program memory in the computer. The program memory may be downloaded by communication.

The figure explaining an Example. The figure explaining a prior art example.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 Microphone 21 Analog / digital converter 22 Frequency band division part 24 Input sound signal power calculation part 27 S / N ratio estimation part 28 Gain factor insertion part 29 Time domain conversion part 30 Gain factor calculation part 34 Digital / analog converter 51 Noise Power estimation unit 52 Input speech signal addition rate determination unit 53 Input speech signal addition unit 61 Elemental gain factor calculation unit 62 Gain factor smoothing unit 63 Gain factor enhancement unit

Claims (10)

In a noise reduction method for reducing a noise signal from an input voice signal,
Convert the input audio signal into a frequency domain signal, calculate the input audio signal power for each frequency from the frequency band signal of the input audio signal, estimate the noise power for each frequency from the input audio signal power for each frequency, and for each frequency Calculate the raw gain factor from the input voice signal power and noise power at each frequency, smooth the prime gain factor, and superimpose the smoothed gain factor on the frequency band signal of the input voice signal to generate the frequency band signal of the noise reduction signal. A noise reduction method comprising: calculating, inversely converting a frequency band signal of a noise reduction signal into a time domain signal, and outputting the signal. The noise reduction method according to claim 1,
A noise reduction method, characterized by performing a smoothing process of weighting and averaging an elementary gain factor. In the noise reduction method according to any one of claims 1 and 2,
A noise reduction method characterized by emphasizing a smoothing gain factor and calculating a frequency band signal of a noise reduction signal by superimposing the enhanced gain factor on a frequency band signal of an input voice signal. The noise reduction method according to claim 3,
A noise reduction method characterized by performing an emphasis process for making a smoothing gain factor close to 0 or 1 depending on a value of a smoothing gain factor. In a noise reduction device that reduces a noise signal from an input voice signal,
A frequency band dividing unit that converts an input audio signal into a frequency domain signal and divides the signal into a plurality of frequency bands; an input audio signal power calculation unit that calculates an input audio signal power for each frequency from the frequency band signal of the input audio signal; A noise power estimator for estimating noise power for each frequency from input voice signal power for each frequency; a prime gain factor calculator for calculating a prime gain factor from input voice signal power for each frequency and noise power for each frequency; Gain factor smoothing unit that smoothes the gain factor, gain factor insertion unit that calculates the frequency band signal of the noise reduction signal by superimposing the smoothing gain factor on the frequency band signal of the input audio signal, and the frequency band of the noise reduction signal A noise reduction apparatus comprising: a time domain conversion unit that converts a signal back to the time domain and outputs the signal. The noise reduction device according to claim 5, wherein
A noise reduction apparatus characterized by performing a smoothing process of weighted addition averaging on an elementary gain factor. In the noise reduction device according to any one of claims 5 and 6,
A noise reduction device comprising a gain factor emphasizing unit for emphasizing a smoothing gain factor, and calculating a frequency band signal of a noise reduction signal by superimposing the enhancement gain factor on a frequency band signal of an input voice signal . The noise reduction device according to claim 7,
A noise reduction apparatus characterized by performing an emphasis process for making a smoothing gain factor close to 0 or 1 depending on a value of a smoothing gain factor.   A noise reduction program in which the noise reduction method according to any one of claims 1 to 4 is described by codes written to and read from a computer.   A recording medium on which the noise reduction program according to claim 9 is recorded.

Images