JP3454402B2 - Band division type 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 noise reduction method for outputting a voice signal with reduced noise from an input signal in which a target voice signal and signals such as unnecessary noise are mixed in a voice / sound device such as a TV conference device.

[0002]

2. Description of the Related Art In a voice communication system such as a voice conference and a video conference, when a microphone receives a sound and a surrounding noise other than the intended voice is mixed in a transmission signal sent to the other party, the clarity of the voice becomes clear. And the call quality is significantly deteriorated. Therefore, it is strongly required to reduce ambient noise other than the target voice included in the transmitted signal.

The noise reduction method is a technique of outputting a noise-reduced signal from an input signal in which a target voice signal and signals such as unnecessary ambient noise are mixed. FIG. 2A shows a sound collection system, and a conventional noise reduction method will be described using this. In this specification, the time representation of a signal is represented by, for example, X (n) using an integer value n representing discrete time. The target audio signal 1 that the speaker 12 has just uttered
3 is S (n), and unnecessary ambient noise 14 such as air conditioning is N
(N), X (n) is the input signal 15 that is received by the microphone 11 of the voice signal 13 and the noise 14 and is input to the noise reduction device 16, and the output signal 17 of the noise reduction device 16.
Be Y (n). Input signal X to the noise reduction device 16
Ambient noise N (n) is mixed in (n) in addition to the target audio signal S (n). That is, X (n) = S (n) + N (n) (1) At this time, noise N in the input signal X (n)
A device that reduces (n) and extracts a signal close to the target audio signal S (n) as the output signal Y (n) is called a noise reduction device.

FIG. 3 shows Spectrum Subtraction (SFBol
l, IEEE Trans. on ASSP, vol.27, no.2, pp.113-120,
Apr (1979)). Wiener Filer (JSLim. & AVOppenheim,
in Proc. IEEE, vol.67, no.12, pp.1586-1604, Dec (1
979)). Maximum Likelihood Envelop (RJMcAulay &
MLMalpass, IEEE Trans.on ASSP, vol.28, no.2, p
p.137-145, Apr (1980)). minimum mean squared error
method (MMSE) (Y.Ephraim & D.Malah, IEEE Trans. on AS
SP, vol.32, no.6, pp.1109-1121, Dec (1984)). Short time spectral amplitude (STSA) evaluation (Short Time S
This is a noise reduction method based on pectral Amplitude (STSA) Estimation) and shows a functional configuration of a method that has been conventionally used. A conventional noise reduction method will be described using this. Common symbols are used for the same elements as in FIG. 2A.

First, the A / D converter 21 digitizes the input signal 15 in which the target signal and unwanted noise are received by the microphone 11, and transfers it to the frequency band divider 22 and the noise discriminator 23. . The frequency band division unit 22 divides the transferred signal into a plurality of frequency bands. The division into frequency bands is performed using, for example, discrete Fourier transform. Here, the band-divided signal is generally a complex number, but it may be a real number depending on the division method. Generally, the discussion will be made assuming a complex number, but the same discussion can be applied to the case of a real number. The signal of the k-th frequency band divided into the frequency bands is represented by X _k (n) = X _{k, r} (n) + jX _{k, i} (n) (2) (X _{k, r} , X _{k, i} are each X When _k (n) the real and imaginary parts of), X _k (n) is the input signal power calculation section 2
4, transferred to the input signal phase calculator 25 and the noise power calculator 26. In the input signal power calculation unit 24, the power level of the input signal for each band P _{X, k} (n) = X _{k, r} (n) ² + X _{k, i} (n) ² (3) is calculated as the input signal phase. In the unit 25, the phase Φ _k (n) = tan ⁻¹ [X _{k, i} (n) / X _{k, r} (n)] (4) for each band is calculated. After that, P _{X, k} (n) is transferred to the S / N ratio estimation unit 27 and the gain factor insertion unit 28, and Φ _k (n) is transferred to the time domain conversion unit 29. On the other hand, in the noise discriminating unit 23, for X (n) transferred from the A / D converting unit 21, first, the power level P _X (n) = Σ {X (nk)} ² (5) Σ is k = 0 to L-1 are calculated. Here, L represents the integration time. Next, for example, a determination of P _X (n) <P _th (6) is made with _respect to a predetermined threshold value P _th , and when this conditional expression is satisfied, it is determined to be noise. In the noise power calculation unit 26, the power level for each band of noise is P _{N, k} (n) = X _{k, r} only when the noise determination unit 23 determines that the input signal X (n) is noise. (n) ² + X _{k, i} (n) ² (7) is calculated, and the time average Pav _{N, k} (n) is transferred to the S / N ratio estimation unit 27. The time average is calculated as, for example, Pav _{N, k} (n) = (1 / A) Σ _m γ _m P _{N, k} (nm) (8). Here, γ _m is, for example, an exponentially weighted coefficient represented by γ _m = (γ) ^m (9) (γ <1),
A is a constant for normalization such that (1 / A) Σ _m γ _m = 1 (10).

The S / N ratio estimation unit 27 calculates P _{X, k} (n) calculated by the input signal power calculation unit 24 and the noise power Pav calculated by the noise power calculation unit 26 for each band.
_{The N / k} (n) is used to estimate the S / N ratio which is the ratio of the target speech signal to the noise signal. Regarding the S / N ratio estimation unit 27, a flow chart of processing is shown in FIG. 4A. Detailed processing will be described with reference to this figure.

[0007] First, in step 31, calculates the _{SNR k (n) '= P} X, k (n) / P N, k (n) SNR k defined in (11) (n)'. Next, in step 32, for the determined SNR _k (n) ′, (1
SNR _{k obtained} by averaging using the estimated value (noise-reduced power) P _{Y, k} (n-1) one hour before, which is represented by the equation 3).
(n). That is, SNR _k (n) = (1-β) P [SNR _k (n) '-1] + β [P _{Y, k} (n-1) / _{PN, k} (n-1)] (12) . P [*] takes * if * is positive and 0 if * is negative. This SNR _k (n), if necessary, SNR _k (n) ′
Is transferred to the gain factor calculation unit 30.

In the gain factor calculation unit 30, the S / N
Using the SNR _k (n) transferred from the ratio estimation unit 27, and possibly this and the SNR _k (n) ′, the gain factor G (SNR _k (SNR _k ( n)) is calculated. The gain factor G (SNR _k (n)) is transferred to the gain factor insertion unit 28. FIG. 5 shows the gain factor by each method. That is, although the gain factors are obtained using only the SNR _k (n) in the upper three methods in FIG. 5, the lowest MM
When by SE method, in addition to SNR _k of _{SNR k (n) (n)} '
To use.

The gain factor inserting section 28 performs noise reduction for each band using the gain factor calculated by the gain factor calculating section 30. That is, the band signal P transferred from the input signal power calculator 24
_{For X, k} (n), P _{Y, k} (n) = G (SNR _k (n)) × P _{X, k} (n) (13) and noise-reduced band output P _{Y, k} ( output n).
P _{Y, k} (n) is transferred to the time domain transforming unit 29, and using Φ _k (n) sent from the input signal phase calculating unit 25, Y _k (n) = Y _{k, r} (n) + JY _{k, i} (n) where Y _{k, r} (n) = P _{Y, k} (n) cos [Φ _k (n)] Y _{k, i} (n) = P _{Y, k} (n) sin [Φ _k (n)] (14), combined into a full band signal, and further converted into a time domain signal by, for example, an inverse discrete Fourier transform. The result is converted into an analog signal in the D / A converter 34, and the noise-reduced signal 17, Y (n), is output. Spectr
In al subtraction, the noise power P _{N, k} (n) is subtracted from the input signal power P _{X, k} (n) by the calculation of the equation (13).

In these conventional systems, 256 to 102
Processing is often performed by dividing into 4 frequency bands,
If this is used as it is in the conference system, a large delay occurs. If a signal is delayed when a conference is held, there is a problem that the communication performance is deteriorated. On the other hand, if the number of frequency divisions is reduced to reduce the delay, the S / N ratio defined by the following equation (15), S / N = 10 log ₁₀ (P _S / P _N ) where P _S : Average power of target signal P _N : Average power of noise signal (15) When the noise is about 15 dB or more, the noise quality improves the voice quality, but when it is about 10 dB or less,
Although the noise is reduced, it has been found that the audio signal is distorted, and the noise left unerased changes temporally, resulting in poorer hearing and lower quality. This is because, in the conventional noise reduction method, when processing is performed with a signal divided into a small number of bands, distortion due to nonlinear processing becomes large.

That is, in the conventional method, the delay is kept small,
There is a problem that noise cannot be effectively reduced.
For this reason, this method cannot be applied as it is to a voice conference device, a TV conference device, or the like, in which sound quality is important for the purpose of listening and a real-time property is required.

[0012]

In the conventional method of subtracting the amplitude of noise from the speech signal as a method of reducing the noise of the input signal mixed with ambient noise, the delay increases as the number of frequency band divisions increases. When the number of frequency band divisions is reduced, there is a problem in that distortion occurs in the audio signal and the noise left behind changes with time, making a sound that is unfavorable to the sense of hearing. An object of the present invention is to provide a noise reduction processing method with less processing delay and less deterioration of sound quality in hearing.

[0013]

SUMMARY OF THE INVENTION According to the present invention, a sound signal received by a microphone in which a target voice signal and ambient noise are mixed is divided into a plurality of bands, and noise power is supplied to each band-specific signal. Estimate and compare the estimated noise power with the input signal power actually input to estimate the ratio of the voice signal and the noise signal, and also estimate the amplitude of the voice signal as a noise-reduced signal. The gain factor for noise suppression for each band calculated based on the ratio of the voice signal and the noise signal is obtained by multiplying the input signal of the corresponding band. The estimation of the noise-reduced signal masks the distortion generated at that time by adding a small amount of signal for each band according to the ratio of the voice signal and the noise signal by adding more, resulting in less distortion. Enables effective noise reduction.

The present invention has the following features. First,
Degradation of sound quality is prevented by adding a small amount of signal for each band and masking the distortion caused by noise reduction. In addition, by changing the addition rate of signals for each band depending on the ratio of the voice signal and the noise signal, when the noise is small, that is, the processing distortion due to the noise reduction is small, the addition amount of the signals for each band is reduced. However, it is possible to maximize the noise reduction effect.

[0015]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows a functional configuration of a noise reduction device to which an embodiment of the present invention is applied, and common symbols are used for those corresponding to FIG. For noise power estimation,
3 may be used, but here, Japanese Patent Application No. 8-68.
548, Sasaki, Haneda "On band-division noise reduction method using loss control" Spring Acoustics Society of Japan Proceedings 515-5
16 (1996), the following explanation will be made by using the method proposed. According to this method, the noise power can be detected if there is a time-free period even in a period considered as a state in which voice and noise are mixed.

First, the input signal 15 received by the microphone 11 and mixed with the target signal and unnecessary noise is digitized by the A / D converter 21 and transferred to the frequency band divider 22. The frequency band division unit 22 divides the transferred signal into frequency bands. The divided band signals are transferred to the input signal power calculation unit 24 and the gain factor insertion unit 28. Hereinafter, the k-th band signal of the input signal as X _k (n), for explaining the flow of processing for X _k (n).

In the input signal power calculator 24,
Came X_kCalculate the power level of (n) by the above formula (7)
And transfers it to the S / N ratio estimation unit 27 and the noise power estimation unit 51.
To be done. The noise power estimator 51 transfers the transferred P
_{X, k}noise power Pav using (n) _{N, k}(n) is estimated
Be done. FIG. 2A is a flowchart of the process of the noise power estimation unit 51.
Show.

First, in step 61, the time average power level P _{X, k} (n) of the transferred P _{X, k} (n) is calculated using the above equations (8), (9) and (10). calculate. In this case, the average time is, for example, 5 to 6 msec. Next, in step 62, Pav at a fixed time
Take a histogram of the level distribution of _{X, k} (n). That is, P
The number of power intervals to which av _{X, k} (n) belongs is incremented by 1, that is, h _k (int Pav _{X, k} (n)) = h _k (int Pav _{X, k} (n)) + 1 (16) . int (*) indicates that the number after the decimal point is truncated to be an integer. Further, in step 63, the histogram h _k (i)
Is detected and stored. That is, i is _{obtained such} that h _k (i ′) ≦ h _k (i) (17) with respect to the preceding and following values. The smallest i among the i is the noise power Pav _{N, x} (n). That is, when a plurality of peak values are obtained, the smallest peak value is used as the noise power Pav _{N, k} (n)
This is transferred to the S / N ratio estimation unit 27.

The S / N ratio estimating unit 27 uses P _{X, k} (n) calculated by the input signal power calculating unit 24 and noise power Pav _{N, k} (n) estimated by the noise power estimating unit 51. hand,
SNR _k (n) 'and SNR _k (n) by the method shown in FIG. 4A.
Is estimated. SNR estimated by the S / N ratio estimation unit 27
_k (n) is transferred to the gain factor calculation unit 30 and the input signal addition rate determination unit 52. Gain factor calculator 3
Depending on the calculation method used for 0, SNR _k (n) ′ is also input.

In the gain factor calculation unit 30, the S / N
SNR _k which has been transferred from the ratio estimating unit 27 (n), using the SNR _k (n) 'optionally, a gain factor G
(SNR _k (n)) is determined. Here, for the specific calculation of the gain factor, the method shown in FIG. 5 in the section of the prior art is used. The gain factor G (SNR _k (n)) estimated by the gain factor calculation unit 30 is transferred to the gain factor insertion unit 28. The method of the S / N ratio estimation unit 27 and the method of the gain factor calculation unit 30 may be methods other than those described in this specification.

The gain factor insertion unit 28 performs noise reduction using the gain factor G (SNR _k (n)) calculated by the gain factor calculation unit 30. That is, the band signal X _k transferred from the frequency band division unit 22
Equation (18) is calculated for (n). Y ′ _k (n) = G (SNR _k (n)) × X _k (n) (18) This noise-reduced signal Y ′ _k (n) is input signal adder 5
Transfer to 3.

The input signal addition rate determining unit 52 determines the addition rate α'of the input signal based on the S / N ratio using the transferred SNR _k (n). It is desirable that α ′ has a small value when SNR _k (n) is large. For example, 15 ≦ 10 log ₁₀ [SNR _k (n)] ⇒ α ′ = 0 5 <10 log ₁₀ [SNR _k (n)] <15 ⇒ α ′ = 0.1 10 log ₁₀ [SNR _k (n)] <5 ⇒ Determine as α '= 0.3 (19). Further, in this example, the frequency band division unit 2
2 is the case where the input signal addition rate α is obtained by multiplying α ′ by the band division number factor β determined according to the band division number M. For example, α = β × α ′ 1024 ≦ M ⇒ β = 0.01 M <1024 ⇒ β = 1 (20). That is, when the number of band divisions M is large, the addition rate α is reduced. The input signal addition rate α is calculated by the input signal addition unit 53.
Transferred to.

In the input signal addition section 53, the band signal X _k (n) transferred from the frequency band division section 22, the gain factor insertion section 28, and the input signal addition rate determination section 52 and the noise reduced signal Y ′ _k. The band output signal Y _k (n) is output using (n) and the input signal addition rate α. That is, _Yk (n) = [alpha] * _Xk (n) + (1- [alpha]) * _Y'k (n) (21). Y _k (n) is transferred to the time domain conversion unit 29, and the whole area is combined and then converted into the time domain. The result is converted into an analog signal by the D / A converter 34, and the noise-reduced signal 17, Y (n), is output.

The effectiveness of the present invention was evaluated using an opinion evaluation method assuming the situation of an actual TV conference system. The evaluation experiment was performed with a volume of 87 m ³ and a reverberation time of 300 ms.
ec, the background noise level was 46 dB (A) in a variable reverberation room. The number of band divisions M is set to 64 (this is the number of divisions that is preferable for the subband echo canceller). The evaluation was carried out with 30 subjects. The created stimulus sound was sent from the speaker, and the subject evaluated the quality as the received voice at the loud conference. 8d
Noise-added voices with three types of S / N ratios of B, 6 dB, and 1 dB were processed by the four types of noise reduction methods shown in FIG. 5, and it was examined how the evaluation changes before and after the noise reduction process.

FIG. 2B (a) shows changes in the MOS value in each noise reduction method. Comparing before and after noise reduction processing, when the S / N ratio is good,
By the processing, the MOS value is higher than that of the noise-added voice (original voice), but it is lowered when the S / N ratio is bad. This is considered to be because when the number of band divisions is small, the distortion of the voice due to the processing and the deterioration due to the unerased noise are more severe than the quality improvement due to the noise reduction, and the MOS value is lowered. From this fact, when the S / N ratio is poor, not only does the conventional method not improve the sound quality by noise reduction,
On the contrary, you can see that it is decreasing.

FIG. 2B (b) shows the result of adding the original sound to the noise-reduced sound according to the present invention. here,
Α was set to 0.3 regardless of the S / N ratio. When the original voice is added when the S / N ratio is poor, the MOS value is increased by about 0.5 to 1 as compared with before the addition, and as a result, the sound quality is improved by the noise reduction process. Further, it can be seen that a more effective sound quality improvement can be expected by changing the addition rate α according to the S / N ratio and reducing the addition rate α when the S / N ratio is good.

The gain factor insertion may be performed as shown in FIG. 3, that is, by the equation (13), and from this, Y _k (n) 'may be obtained by the equation (14). As can be understood from the above embodiment, the ratio of the noise to the target voice signal of the band is estimated from the input signal power of each band and the noise power of each corresponding band, and the target voice of the band is estimated. The amplitude of the signal may be estimated as a noise-reduced signal.

[0028]

As described above, according to the present invention, a signal in which target speech and noise are mixed is divided into frequency bands, the ratio of the noise signal to the target signal for each band is estimated, and the band is calculated based on this estimation result. The noise is reduced for each, but in order to reduce the distortion of the target voice that occurs when the number of frequency band divisions is small, the original signal is added more to the noise reduced voice according to the ratio of the noise to the voice as the noise is higher. As a result, noise-reduced speech with less distortion can be obtained. At this time, the smaller the number of band divisions, the larger the ratio of addition.

By the noise reduction for the purpose of listening according to the method of the present invention, it becomes possible to obtain a target signal which has less delay and is easy to hear. As a result, in a voice call system such as a voice conference and a video conference, even when ambient noise other than the intended voice is mixed in the transmission signal received by the microphone and transmitted to the other party, the method according to the present invention is used. By reducing noise, it is possible to maintain the clarity of voice and improve communication quality.

[Brief description of drawings]

FIG. 1 is a block diagram showing a functional configuration of a noise reduction device to which an embodiment of the present invention is applied.

2A is a process flow chart in a noise power estimation unit 51 in FIG. 1, B (a) is a diagram showing a result of subjective evaluation of a noise reduction effect by a conventional method, and FIG. 2 (b) is a noise reduction effect by the present invention. It is a figure which shows the result of subjective evaluation.

FIG. 3 is a block diagram showing a functional configuration of a device to which a conventional noise reduction method is applied.

FIG. 4A is a flow chart of processing in a gain factor calculation unit, and B is an explanatory diagram of the principle of the noise reduction device.

FIG. 5 is a diagram showing a typical formula for calculating a gain factor.

Continuation of the front page (56) Reference JP-A-8-221092 (JP, A) JP-A-3-247011 (JP, A) JP-A-59-67732 (JP, A) JP-A-3-266899 (JP , A) JP-A-9-258792 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 11/00 G10L 21/02 G10L 15/20

Claims (3)

(57) [Claims] 1. A noise reduction processing method for outputting a speech signal from which a noise signal has been removed from an input signal, comprising: a band division process of dividing the input signal into a plurality of frequency bands; and a step of dividing the input signal into the respective frequency bands. Input signal power calculation step for calculating power, noise power calculation step for obtaining noise power in the input signal divided into each band, input signal power for each band, and noise power for each band together to estimate the ratio of the noise with respect to the target speech signal for each band from the the estimation process for estimating the said target speech signal noise low reduced signal amplitude of each band, the estimated the target speech Determining a ratio of addition of the noise-reduced signal for each band and the input signal divided into each band based on a signal-to-noise ratio; A process of adding the noise-reduced signal for each band and the input signal divided into each band based on an addition ratio and outputting a band output signal, and converting the band output signal into a time domain A band-division type noise reduction method comprising: a process of combining into a band signal. 2. The noise reduction method according to claim 1, wherein the estimation process uses the input signal power for each band and the noise power for each band, and the target speech signal pair for each band. A step of estimating a noise ratio, a gain factor calculation step of determining a gain factor for each band based on the target speech signal-to-noise ratio, and a band for each input signal divided into each band. A band-division-type noise reduction method comprising a step of inserting the gain factor determined for each band to obtain the noise-reduced signal for each band. 3. The noise reduction method according to claim 1, wherein, in accordance with the number of frequency bands divided in the band division process,
A band-division type noise reduction method, wherein an addition ratio of the noise-reduced signal for each band and the input signal divided into each band is determined.