JP3309895B2 - Noise reduction method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice / audio device such as a voice conference device, a TV conference device, etc., which is a signal obtained by reducing noise from an input signal in which a signal of interest and a signal such as unnecessary noise are mixed. it relates to noise reduction how to output.

[0002]

2. Description of the Related Art In a voice communication system such as a voice conference or a TV conference, if ambient noise other than the target voice is mixed in a transmission signal received by a microphone and transmitted to a partner, the clarity of the voice is increased. And the call quality is significantly degraded. For this reason, there is a strong demand for reducing ambient noise other than the target voice included in the transmission signal.

[0003] The noise reduction method is a technique of outputting a noise-reduced signal from an input signal in which a target voice signal and a signal such as unnecessary ambient noise are mixed.

FIG. 6 shows a sound collection system, which will be used to explain a conventional noise reduction method and apparatus. 1
1 is a microphone, 12 is a speaker at a position distant from the microphone 11, 13 is a target audio signal uttered by the speaker 12, 14 is unnecessary ambient noise such as air-conditioning sound, and 15 is the microphone 11 16 shows a received input signal to the noise reduction device, 16 shows a noise reduction device, and 17 shows an output signal of the noise reduction device 16. In this specification, the time expression of a signal is represented by, for example, X (n) using an integer value n representing a discrete time.

[0005] Now, an audio signal 13 is converted to S (n) and ambient noise 14.
Is N (n), and the input signal 15 to the noise reduction device 16 is X (n).
If the output signal 17 of the noise reduction device 16 is Y (n), the input signal X (n) to the noise reduction device 16 includes ambient noise N (n) in addition to the target audio signal S (n). It is mixed. Ie

[0006]

(Equation 1) It is expressed as At this time, the ambient noise N in the input signal X (n)
(n), and extracts a signal close to the target audio signal S (n) as an output signal Y (n).
Call.

FIG. 7 is a block diagram showing a method called a spectrum subtraction method, which is conventionally used in the field of speech recognition. The conventional noise reduction method will be described with reference to FIG. In FIG. 7, common symbols are used for the same elements as those in FIG. 21 is an A / D converter, 22 is a frequency band divider, 23 is a noise discriminator, 24 is an input signal power calculator,
25 is an input signal phase calculator, 26 is a noise power calculator,
27 denotes a subtraction unit, 28 denotes a time domain conversion unit, and 29 denotes a D / A conversion unit.

First, an A / D converter 21 digitizes an input signal 15 mixed with a target signal and unnecessary noise received by a microphone 11 and transfers the digitized signal to a frequency band dividing unit 22 and a noise discriminating unit 23. . In the frequency band dividing unit 22, the transferred signal is divided into a plurality of frequency bands. Division into frequency bands is performed using, for example, a discrete Fourier transform or the like. Here, the band-divided signal is generally a complex number, but may be a real number depending on the division method. Here, discussion is generally made assuming a complex number, but the same discussion can be made for a real number. The signal of the k-th frequency band divided into frequency bands is

[0009]

(Equation 2) Then, X _k (n) is transferred to the input signal power calculator 24, the input signal phase calculator 25, and the noise power calculator 26. The input signal power calculator 24 calculates the power level of the input signal.

[0010]

(Equation 3) Is the phase in the input signal phase calculator 25.

[0011]

(Equation 4) Is calculated. Thereafter, P _{X, k} (n) is transferred to the subtraction unit 27, and Φ _k (n) is transferred to the time domain transformation unit 28. on the other hand,
The noise discriminating unit 23 first determines the power level of X (n) transferred from the A / D converting unit 21.

[0012]

(Equation 5) Is calculated. Next, for example, a predetermined threshold value P _th
Against

[0013]

(Equation 6) Is determined, and if the conditional expression is satisfied, it is determined that the noise is present. The noise power calculation unit 26 determines the power level of the noise only when the noise discrimination unit 23 determines that the input signal is noise.

[0014]

(Equation 7) And the time average Pav _{N, k} (n) is subtracted by the subtractor 27.
Transfer to The time average is, for example,

[0015]

(Equation 8) Is calculated. Where γ _m is, for example,

[0016]

(Equation 9) Is an exponentially weighted coefficient expressed as (γ <1),
A is

[0017]

(Equation 10) Is a constant for normalization. The subtraction unit 27 uses P _{X, k} (n) and Pav _{N, k} (n) transferred from the input signal power calculation unit 24 and the noise power calculation unit 26.

[0018]

[Equation 11] Is calculated. P _{Y, k} (n) is transferred to the time domain transforming unit 28 and using Φ _k (n) sent from the input signal phase calculating unit 25,

[0019]

(Equation 12) And combined into a full-band signal. This result is
The A / A converter 29 outputs Y (n) of the output signal 17 in which noise has been reduced into an analog signal.

In this method, when the power of the target sound included in the input signal 15 in which the audio signal serving as the target signal and unnecessary noise are mixed is P _S and the power of the unnecessary sound is P _N ,

[0021]

(Equation 13) When the S / N ratio defined by the above is about 25 dB or more, noise can be effectively reduced. However, the S / N ratio is 2
In the case of about 0 dB or less, the noise is reduced, but the sound signal is distorted accordingly. This is because the spectrum subtraction method performs a non-linear process of subtraction at the stage of a signal divided into bands.

On the other hand, when the subtraction is performed using a value lower than the measured noise level in order to reduce the distortion of the voice, for example, the P estimated by the noise power calculator 26 is used.
for av _{N, k} (n)

[0023]

[Equation 14] Is performed, the Pav _{N, k} (n) is transferred to the subtraction unit 27, and when the subtraction is performed, the noise cannot be completely reduced, and the remaining noise of Y (n) of the output signal 17 of the noise reduction device 16 is not removed. Produces an unpleasant sound due to temporal changes. That is, in the spectrum subtraction method, there is a problem that the output signal after processing becomes undesirably audible regardless of the subtraction amount. For this reason, audio conference devices and TVs
This method cannot be directly applied to sound collection such as a conference apparatus in which sound quality is important for listening.

[0024]

As a method for reducing the noise of an input signal mixed with ambient noise, a conventional method of subtracting the amplitude of noise from a speech signal results in a large amount of subtraction resulting in large distortion of speech. If the amount of subtraction is small, there is a problem in that the noise left undesirably sounds due to the temporal change of the remaining noise.

[0025] It is an object of the present invention is to provide a noise reduction how little deterioration of sound quality of audibility.

[0026]

Noise reduction <br/> how according to the present invention, in order to solve the problems] divides the received sound signals in a mixed microphone such as voice signals and ambient noise are the object into a plurality of bands, each Estimate the noise power for the signal of each band, compare the estimated noise power with the input signal power actually input, estimate the ratio of the voice signal and the noise signal, and based on this, This is a technique for suppressing noise by giving a loss to a signal.

As the loss value, a value determined based on the estimated value of the ratio of the voice signal to the noise signal, and after averaging between the bands and averaging with the past loss value is used. .

The noise power is estimated using a histogram of the power level distribution of the input signal. The power level of an audio signal greatly fluctuates and its distribution varies. However, since the power level value of stationary noise is almost constant, a strong peak appears at the noise average level position when a histogram of the power level distribution is taken. . In the present invention, using this property, the peak section of the histogram is detected and the power of the noise is estimated.

[0029]

The present invention has the following features. First, by dividing the band into bands and processing, even if a loss insertion error occurs in some bands, the overall effect is small. In addition, since the linearity of insertion of loss gives processing to the conventional spectrum subtraction method using non-linear processing called subtraction, distortion to a signal is small. In addition, in the rising portion and the falling portion of the voice, the power of the voice fluctuates drastically in a short time, so that the ratio between the voice signal and the noise signal greatly changes, and accordingly, the loss value before averaging changes sharply. As a result, if this loss value is used, a sense of discontinuity occurs in the processed signal. In the present invention, the discontinuity is eliminated by averaging the loss value with the band and time. Further, in particular, by using a histogram for noise power estimation required for calculating the noise contribution ratio, noise power can be estimated even in a section where speech and noise are mixed.

[0030]

DETAILED DESCRIPTION FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention, and used to explain the processing procedure of the present invention. 6 and 7, the same symbols are used. In FIG. 1, reference numeral 31 denotes a noise power estimator, 32 denotes a loss value calculator, and 33 denotes a loss inserter.

The operation will be described. First, an input signal 15 mixed with a target signal and unnecessary noise received by the microphone 11 is digitized in the A / D converter 21 and transferred to the frequency band divider 22. In the frequency band dividing section 22, the transferred signal is divided into frequency bands. Each of the divided band signals is transferred to an input signal power calculation unit 24, a noise power estimation unit 31, and a loss insertion unit 33. Hereafter, let the k-th band signal be X _k (n),
The flow of processing for _k (n) will be described.

The input signal power calculator 24 calculates the power level of the transferred band signal X _k (n).

[0033]

(Equation 15) Is calculated. Next, the power level P _{X, k} (n) is averaged for a predetermined time, and transferred to the loss value calculation unit 32 as Pav _{x, k} (n). The time average is, for example,

[0034]

(Equation 16) Is calculated. Where γ _m is, for example,

[0035]

[Equation 17] Is an exponentially weighted coefficient expressed as (γ <1),
A is

[0036]

(Equation 18) Is a constant for normalization. If the average time m is too short, the temporal variation in discrimination between speech and noise increases. On the other hand, if the length is too long, the performance of following a temporal change is deteriorated. As an average time, for example, 0.5 to 1 mse
Take c.

The noise power estimator 31 estimates the noise power Pav _{N, k} (n) using the transferred band signal X _k (n). FIG. 2 shows a flowchart of the processing performed by the noise power estimating unit 31. Detailed description of the processing will be made using this figure. S21 to S23 in FIG. 2 indicate each step.

First, in step S21, the power level P _{X, k} (n) of the transferred X _k (n) is calculated using the equation (12). Next, the power levels P _{X, k} (n) are averaged for a predetermined time to obtain Pav ′ _{x, k} (n). Pav' x, the average time m'of _k (n) is, Pav x in the input signal power calculation unit 24, as compared to the average time m of _k (n), it is desirable to lengthen the average time is too long No longer follows noise fluctuations. For example, it takes several msec. Next, in step S22, a histogram of the level distribution of Pav ' _{x, k} (n) is _obtained . That is _, the number of power sections to which Pav ' _{x, k} (n) belongs is added by 1, that is,

[0039]

[Equation 19] I do. int (*) indicates that the value after the decimal point is truncated to an integer. Further, in step S23, the histogram h
The peak section of _k (i) is detected and stored. That is, for the value before and after

[0040]

(Equation 20) Is obtained. The smallest i among these i is assumed to be noise power Pav _{N, k} (n). This noise power Pav _{N, k} (n)
Is transferred to the loss value calculator 32.

The loss value calculator 32 uses the Pav _{x, k} (n) calculated by the input signal power calculator 24 and the noise power Pav _{N, k} (n) estimated by the noise power estimator 31 to perform a loss. The value is calculated. FIG. 3 shows a flow chart of the processing for the loss value calculation unit 32. Detailed description of the processing will be made using this figure.

[0042] First, the noise power Pav N, _k a value obtained by adding a predetermined margin P _th to (n) threshold P _{N ',} and _k (n) in step S31. Ie

[0043]

(Equation 21) And Here, if the margin P _th is too small, the effect will be small, and if it is too large, a sense of discontinuity will appear. For example, 5 × Pav _{N, k} (n) is taken as the margin P _th . Next, in step S32, Pav _{X, k} (n) is compared with Pav ′ _{N, k} (n). If Pav _{X, k} (n) is smaller than Pav ′ _{N, k} (n), the loss value is calculated. Is smaller than 1 and the loss value is set to 1 if it is equal to or greater than 1.

[0044]

(Equation 22) Then, in step S33, with respect to the loss value L _k determined in step S32 (n), the loss values of neighboring bands, _{L k-2 (n),} L k-1 (n) ......, L k + Averaging by ₂ (n) to reduce the difference between each band. Also, the past loss values L _k (n-1),
Using L _k (n−2)..., The amount of loss is averaged in a time of, for example, about 30 msec. If the averaging time is short, the loss value fluctuates with time, causing a sense of discontinuity. The value obtained by the averaging is newly set as L _k (n) and transferred to the loss insertion unit 33.

The loss insertion unit 33 performs loss control using the loss value L _k (n) calculated by the loss value calculation unit 32. That is, for the band signal X _k (n) transferred from the frequency band division unit 22,

[0046]

(Equation 23) And outputs a band output Y _k (n) with reduced noise. The band output Y _k (n) is transferred to the time domain converter 28, where it is converted to the time domain. The result is converted into an analog signal by the D / A converter 29, and the Y (n) output signal 17 in which noise is reduced is output.

In the above description, each section in FIG. 1 is shown as hardware, but these can be configured as software by a computer program. Therefore, in the description of claims, parts are expressed by means instead of parts.

Next, the noise reduction method of the present invention will be described. Basically, the operation procedure is the same as that of FIG. 1, but when collectively shown, it is as shown in the flowchart of FIG. In this figure, S41
S47 indicate each step.

When an input signal X (n) in which speech and noise are mixed is input, the speech and noise mixed signal is divided into a plurality of bands in the process of step S41. In step S42, an average power level Pav _{x, k} (n) of the input signal is calculated for each band signal X _k (n) of the plurality of divided frequency bands. Wherein to the process of step S43 in parallel with this noise power Pav N of each band signals X _k (n), the estimation of _k (n) is performed. Average power level Pav _{x, k} (n) and noise power Pa
From v _{N, k} (n), the ratio of noise in each band in the band signal X _k (n) is determined in the process of step S44. Next, in the process of step S45, a loss value for each band is determined based on the ratio. Next, a loss value determined for each band is inserted into each band signal of the plurality of frequency bands divided in the process of step S46, to obtain a band output signal in which noise for each band is reduced. Then, step S
In the process of 47, the noise-reduced band output signal is converted into the time domain to be a full-band signal, and an analog signal with reduced noise is obtained.

The effectiveness of the present invention was verified by computer simulation on a female voice (16 kHz sampling) collected under the conditions of an actual TV conference system. FIG. 5 shows a spectrogram of the noise-added voice, the spectrum subtraction shown in the prior art, and the data subjected to the noise reduction processing according to the present invention. The spectrum subtraction and the present invention are both 250 Hz width 3
The processing was performed after the signal was divided into two band signals.

FIG. 5A shows a voice with noise.
The black band below 00 Hz and the gray part which spreads out as a whole are the ambient noise mainly composed of the air-conditioning sound. It can be seen that the voice is erased by the noise.

FIG. 5B shows the result of the processing performed by the spectrum subtraction. The noise is removed as a whole, and the difference from the voice is clearly shown. However, the noise that has not been completely eliminated appears as a thin stripe. This frequency component changes with time, which causes a curly and unpleasant sound. In the audio part, a part having a small amplitude disappears due to excessive subtraction, and deterioration is recognized.

FIG. 5C shows the result of loss control according to the present invention. Compared with the spectrum subtraction, the contrast between the voice and the noise part is not clear, but it can be seen that the noise part is thin overall and the voice is emphasized. By increasing the clarity without significantly changing the original sound, a sound that is easy to hear is created.

When the S / N ratio was calculated, it was 18 dB and 30 dB before and after the loss control processing, respectively, and it was confirmed that the noise control had a 12 dB noise reduction effect.

[0055]

According to the present invention, there is provided a noise reduction method and apparatus for listening, which has the following features. First, by dividing the band into bands and processing, even if a loss insertion error occurs in some bands, the overall effect is small. In addition, since a linear process of inserting a loss is provided, distortion to a signal is small. By averaging the loss values, the ratio between the audio signal and the noise signal greatly changes, such as the rising portion and the falling portion of the voice, and the discontinuity is eliminated even in the portion where the loss value changes drastically with it. . Also, by using a histogram for noise power estimation, noise power estimation can be used even in a section where speech and noise are mixed.

[0056] The noise reduction for the purpose of listening by way of the present invention, it is possible to obtain a likely target signal heard. As a result, in the hands-free communication systems such as audio conferencing · TV conference, and the sound receiving by a microphone, the transmission signal sent to the other party, even if the ambient noise other than the voice to be objective is mixed, in the how Due to the reduction of noise, clarity of voice can be maintained, and communication quality is improved.

[Brief description of the drawings]

1 is a block diagram showing a configuration of an embodiment of the present invention.

FIG. 2 is a flowchart of a process performed by a noise power estimation unit in the embodiment of FIG.

FIG. 3 is a flowchart of a process performed by a loss value calculator in the embodiment of FIG. 1;

Is a flow diagram showing such a noise reduction methods from the present invention; FIG.

FIG. 5 is a spectrogram for explaining the effect of the present invention.

FIG. 6 is a diagram for explaining the principle of a conventional noise reduction device.

FIG. 7 is a block diagram illustrating an example of a conventional noise reduction device.

[Description of Signs] 11 Microphone 15 Input signal 16 Noise reduction device 17 Output signal 21 A / D conversion unit 22 Frequency band division unit 28 Time domain conversion unit 29 D / A conversion unit 31 Noise power estimation unit 32 Loss value calculation unit 33 Loss insertion part

Continuing on the front page (72) Inventor Yutaka Kaneda 3-19-2 Nishi-Shinjuku, Shinjuku-ku, Tokyo Nippon Telegraph and Telephone Corporation (72) Inventor Junji Kojima 3-192-2, Nishi-Shinjuku, Shinjuku-ku, Tokyo Date (56) References JP-A-3-266899 (JP, A) JP-A-59-67732 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 21/02

Claims (3)

(57) [Claims] 1. A frequency band dividing step of dividing an audio / noise mixed signal into a plurality of bands, and an input signal power for calculating an input signal power for each band with respect to each band signal of the plurality of divided frequency bands. A calculating process, a noise power estimating process of estimating a noise power of each of the frequency bands by using each band signal of the plurality of divided frequency bands, and an input signal power of each of the bands and each of the bands. A prediction process of comparing the noise power of each band with the noise power of each band, and a loss value for determining a loss value of each band based on the predicted ratio. A calculation step of: inserting the loss value determined for each band into each band signal of the plurality of divided frequency bands, and outputting a band output signal in which noise of each band is reduced. Excessive insertion When the audio signal and the unwanted noise signal to be objective by having a time-domain transform process of synthesizing the entire band signal by converting the band output signal into the time domain
Noise signal is removed from mixed voice noise signal
A noise reduction processing method for outputting a signal, wherein the noise power estimating step includes the steps of:
Histogram of input signal power level distribution for each band signal
And the peak section of the histogram is detected.
Noise characterized by using the smallest value as the noise power
Reduction method. 2. A loss value calculating step, wherein a predetermined value is added to the estimated noise power signal for each band, and this is set as a threshold value, and the determined threshold value for each band is set. The value is compared with the magnitude of the input signal power for each band, and a predetermined loss value is given for a band where the input signal power for each band is larger than a threshold, and the noise reduction method according to claim 1, wherein a is not inserted loss. 3. The loss value calculating step is characterized in that the loss value determined for each band is averaged with a loss value of an adjacent band to newly determine a loss value for each band. The noise reduction method according to claim 1, wherein