JPH0738454A - Noise reduction method - Google Patents

Abstract

PURPOSE:To attain highly accurate statistic quantity estimation by detecting a silence period in an input signal and estimating a statistic quantity of noise from the input signal in the silence period. CONSTITUTION:An input signal from an input terminal 11 is divided for a predetermined time such a each several millisecond and a state discrimination section 12 discriminates the state of each frame. That is, a voice and a noise in the input signal are distinguished as follows: the states are distinguished into a state in which only a voice is in existence without a background noise, a state that neither background noise nor a voice signal is in existence in an optical signal, a completely silence state, a state that only a background noise is in existence, a state that a background noise is in existence and a speech head or a speech tail of the voice signal is set, and a state that a background noise is in existence and a voice signal is in a steady-state, a code is given to each state and the result is stored in a frame state table 13. Correction sections 16, 17 conduct adaptive correction in response to the discrimination state with respect to the statistic quantity of noise and voice to be estimated. Then each statistic quantity is used to calculate a filter coefficient and the result is set to a Kalman filter 18, in which the noise component is suppressed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is applied to a fixed or mobile voice communication system which may be used in an environment with a large amount of background noise, and determines a filter coefficient from a statistical amount of voice and a statistical amount of noise. The present invention relates to a noise reduction method for reducing the noise superimposed on the original speech by filtering the input signal with the filter coefficient.

[0002]

2. Description of the Related Art A conventional noise reduction method is based on the method described in J.
D. Gibson (“Filtering of Co
lored Noise for Speech En
hancement and Coding ”, IEEE
E Trans. SP. Vol. 39, no. 8, 19
91) et al. This method consists of Kalman filtering processing for each frame. The Kalman filtering process requires statistics of speech and noise. The method of Gibson et al. Is a method in which the statistic of voice for each frame is estimated as unknown and the statistic of noise for each frame is known, and filtering is performed. Therefore, the estimation accuracy of the noise statistics, which is assumed to be known, influences the effect of noise reduction.

[0003]

However, it is not so easy to estimate an accurate noise statistic in advance. It is necessary to record only the noise signal in advance with the influence of the voice removed, and estimate the noise statistics. There is a drawback that the time and effort for performing noise recording processing before a call increases, leading to an increase in system cost. Moreover, in an application area such as mobile radio, the noise statistic estimated in advance may fluctuate with time, and in this case, the estimation accuracy of the noise statistic estimated in advance decreases with time. As a result, it also has a drawback that noise reduction cannot be sufficiently performed.

In order to follow the temporal fluctuation, a method is conceivable in which a voice recording system and a noise recording microphone are provided at the same time as the voice recording microphone, that is, a voice recording system and a different noise recording system are provided. Using this method improves the estimation accuracy of noise statistics, but has the drawback of increasing the system cost. In addition, the conventional noise reduction method does not consider the state of the input signal, that is, the input signal is only noise, only voice, the state where noise is superimposed on the voice, the state where the voice rises, the state where the voice falls, etc. However, there was a drawback in that it might be aurally unnatural.

An object of the present invention is to provide a noise reduction method capable of sufficiently reducing noise. Another object of the present invention is to provide a noise reduction method that sufficiently reduces noise by improving the estimation accuracy of noise statistics without increasing the price. Still another object of the present invention is to provide a noise reduction method for enhancing the noise reduction effect by adaptively processing according to the state of voice and the state of noise.

[0006]

According to the first aspect of the present invention, a silent section in an input signal is detected, and a statistical amount of noise is estimated from the input signal in the detected silent section. That is, in a two-way call such as a telephone call, when one speaker is focused on, there is a certain amount of time during which no voice is uttered (the other party is talking). In this state, only noise is input, and in this state, noise statistics can be estimated with high accuracy.

According to the invention of claim 2, in the invention of claim 1, the states of voice and noise in the input signal are divided into predetermined states, and the respective states are stored as a frame state table. , The input signal is processed every fixed time (frame), the state of each frame is determined by referring to the frame state table, and the voice statistics and the noise statistics are adapted according to the determined frame state. Correct it.

In the invention of claim 3, the frame state table in the invention of claim 2 is prepared, noise is recorded by the noise microphone, and the statistical amount of noise is estimated from the input signal recorded in the noise microphone. On the other hand, the state of voice and noise is determined for each frame from both input signals of the microphone for voice and the microphone for noise, and the statistics of voice and the statistics of noise are referred to by referring to the frame state table according to the state. Is adaptively corrected.

According to the invention of claim 4, the frame state table according to the invention of claim 2 is prepared, the statistical amount of noise is estimated before a call, and the statistical amount of noise is estimated from the input signal. The input signal is processed at fixed time intervals (frames) by using the recorded voice statistics, the state of the frame is determined by referring to the frame state table, and the voice statistics are determined according to the determined frame state. Adaptively modify the quantity and noise statistics.

According to a fifth aspect of the invention, in the invention of any of the second to fourth aspects, the adaptive correction is performed according to the transition of the state. According to the invention of claim 6, in any one of the above inventions, a signal-to-noise ratio is obtained from each estimation statistic of speech and noise, and the adaptive correction is performed according to the signal-to-noise ratio or the noise power. Control.

According to a ninth aspect of the present invention, in any one of the first to fourth aspects of the present invention, each of the n samples at the beginning of each frame (n is an integer of 2 or more and less than the sample number N of one frame) The filter coefficient is updated for each sample, and the filter coefficient is not updated for the remaining (N−n) samples.

[0012]

FIG. 1 shows a processing procedure of an embodiment of the invention of claim 2, and FIG. 2 shows a noise reduction apparatus to which the embodiment is applied. The input signal from the input terminal 11 is divided every fixed time (frame), for example, every several tens of milliseconds (S ₁ ), and the state determination unit 12 determines the state of each frame (S ₂ ). That is, the relationship between the voice and the noise in the input signal is shown in FIG.
As shown in, there is no background noise and there is only voice, there is neither background noise nor voice, there is no silence 0, there is only background noise 20, there is background noise, and there is voice. It is divided into a head state or a tail state 21 and a stationary state 22 of voice in which background noise exists, and these states are given a code and stored in the frame state table 13 in advance.

The input signal makes such a state transition as shown in FIG. 3B. This state is determined by the voice power in the input signal and its fluctuation, the voice spectrum characteristic and its fluctuation, and the like. When the result of the state determination of each frame is the state 20 in which only noise exists (S
₃ ), the noise statistic estimating unit 14 estimates the noise statistic, for example, the AR model and power of the noise from the input signal at that time (S ₄ ). In the states other than the state 20, the speech statistic estimation unit 15
In audio statistics, for example, it estimates the AR model and power of the voice (S _5).

The estimated noise statistic and speech statistic are adaptively modified by the modifying units 16 and 17 according to the determined state of the frame (S).
₆ ). This correction is performed, for example, as shown in FIG. That is, in the state 20, noise reduction processing is strongly performed, so that the power value of the voice signal is set to a value smaller than the power of the input signal (power of noise). The coefficient a by which the power of the input signal is multiplied is a value of 1 or less, for example, 0.1, and the power of the voice with respect to the noise power is reduced so that strong noise reduction processing is performed by the Kalman filter. In this case, the AR model of voice is blunted, and the estimated AR model and power of noise are used as they are.

In the state 21, the noise power in the immediately preceding state 20 is multiplied by a coefficient of 1 or less, for example, 0.7, that is, the noise power is reduced to weaken the noise reduction processing. The AR model and power of speech used are those obtained in that frame, and the AR model of noise used is that obtained in the immediately preceding state 20. In the state 22, speech uses the AR model and power estimated in that frame as they are, and noise uses the AR model and power estimated in the immediately preceding state 20 as they are, that is, noise reduction processing of normal strength is performed.
That is, in the prior art, the estimated voice and noise statistics are used as they are, but in this embodiment, the estimated power of the voice is corrected to be small in the states 20 other than the state 22 to strengthen the noise reduction processing, and in the state 21, the noise is reduced. The estimated statistic is adaptively modified according to the state so that the estimated power of is modified to weaken the noise reduction processing.

The filter coefficient is calculated by using the statistics of the voice and noise adaptively corrected in this way (S
₇ ), the filter coefficient is set in the Kalman filter 18, and the input signal from the terminal 11 is filtered by the filter 18 for each corresponding frame to suppress the noise component in the input signal (S ₈ ). The signal filtered by the filter 18 is voice-encoded as necessary and transmitted or stored.

Since the noise statistic is estimated in the silent section, that is, the state 20 as described above, highly accurate estimation can be realized, there is no need to preliminarily estimate the noise statistic, and only one recording system is required. When the noise state changes, the noise statistic corresponding to this is estimated, and the noise can be reduced satisfactorily. In addition, the noise reduction degree is strengthened, weakened, or adaptively changed according to the state, and the noise reduction is effectively performed.

The number of states and the definition of the states are not limited to the above example. For example, in the above-described embodiment, the head and tail are combined and defined as the state 21, but the head is set to the state 21 and the tail is set to the state 23. In the state 23, as shown in FIG. The coefficient by which the power is multiplied is gradually reduced, for example, 0.9, 0.89, 0.8
Multiply by gradually smaller values such as 8, 0.87, ..., 0.7. By this operation, the sound, that is, the state 21 or 22
If there is no sound, that is, if there is a transition to state 20, the same state 20
Even if it is determined that the noise is gradually reduced in this example by controlling the multiplier a in accordance with the state transition, a natural sense of hearing can be obtained.

FIG. 7 shows an embodiment when the state transition is more complicated. The number of states is increased and the state transitions are diversified as compared with the embodiment shown in FIG. In this embodiment, in order to prevent erroneous noise reduction processing when noise is not superimposed, the configuration shown in FIG. 7 is adopted. Figure 7
8 and 9 show the conditions for state transition in FIG. 8 and the explanation of each variable in FIGS. 8 and 9 is given below.

P: Average power per sample of the input signal. Pratio: If the average power of the current frame is P and the average power of the previous frame is P ', then Pratio = P / P'st-cnt: auxiliary variable for transition control. The initial value is 8. It is decremented by 1 under the transition conditions (a1, b1, d1, e1, f1, g1, j1). Also, the number (e1) is initialized to 8.

Ps1, Pn, Psp: constants. Threshold for average power of input voice. Ps1 = 3.0, Pn = 640.0, Psp = 2500.
0 spg: Variable. Dimensionless number. The initial value is 1.0. Transition condition (i2) and'P ≤ Psp or Ratio ≤ 0.3 or P
When ratio ≧ 1.0 / 0.3 'is satisfied 10 times in a row,
Let psg = 3.0 × Pmin / Psp. It is set to 1.0 in the transition condition (c1).

Pmin: Pmin = MIN (Pmin, P) (number (i2) and'P ≦ Psp or Pratio ≦ 0.3 or Pratio ≧ 1.
0 / 0.3 'is satisfied) Pmin = maximum positive value (other cases) Pow10: a variable having a dimension of electric power. The initial value is 0.0.
In the number (e2), Pow10 = 0.9 × Pow10 + 0.1 ×
Process P is performed.

Pow10cnt: Variable. The initial value is 0. In the number (e2), Pow10cnt = Pow10cnt + 1 (Pow10> 2.0 ×
In the case of Pn), the processing of Pow10cnt = 0 (in the case of Pow10 ≦ 2.0 × Pn) is performed. Pr: Pr = P / Pow20 Pow20: Pow20 = P (when condition d1), Pow20 = 0.9 x
Pow20 = 0.1 × P (when condition l1), Pow20 = Min
(P) (when st22 continues 50 times) K: K = (1.0-ref1 x ref1) x (1.0-ref)
2 x ref2) required. However ref1 and ref2
Is the k parameter when the input is passed through a quadratic linear prediction filter.

Var: The power of the residual when the input is passed through the second-order linear prediction filter. sstt: Takes any value of 0, 2, 4, 6, 8, and 10. Transition conditions (k1, k2, l1, m1, o1, o2
In p1, q1, s1, t1, t2), the value is calculated based on the following calculation formula. However, if Var ′ is the value of Var of the previous frame, Ratio = Var ′ / Var
Is defined by

When sstt = 0 or sstt = 2: 0.6 <Ratio <1.4: sstt = sstt + 2 0.5 <Ratio ≦ 0.6 or 1.4 ≦ Ratio <1.5: sstt =
2 Ratio ≦ 0.5 or 1.5 ≦ Ratio: When sstt = 2 sstt = 4: 0.7 <Ratio <1.3: sstt = 6 0.6 <Ratio ≦ 0.7 or 1.3 ≦ Ratio <1.4: sstt =
4 0.5 <Ratio ≦ 0.6 or 1.4 ≦ Ratio <1.5: sstt =
2 Ratio ≦ 0.5 or 1.5 ≦ Ratio: When sstt = 0 and sstt = 6: 0.8 <Ratio <1.2: sstt = 8 0.7 <Ratio ≦ 0.7 or 1.2 ≦ Ratio <1.3: sstt =
6 0.6 <Ratio ≦ 0.7 or 1.3 ≦ Ratio <1.4: sstt =
4 0.5 <Ratio ≦ 0.6 or 1.4 ≦ Ratio <1.5: sstt =
2 Ratio ≦ 0.5 or 1.5 ≦ Ratio: When sstt = 0 or sstt = 8 or sstt = 10: 0.9 <Ratio <1.1: sstt = 10 0.8 <Ratio ≦ 0.9 or 1 .1 ≦ Ratio <1.2: sstt =
8 0.7 <Ratio ≦ 0.7 or 1.2 ≦ Ratio <1.3: sstt =
6 0.6 <Ratio ≦ 0.7 or 1.3 ≦ Ratio <1.4: sstt =
4 0.5 <Ratio ≦ 0.6 or 1.4 ≦ Ratio <1.5: sstt =
2 Ratio ≦ 0.5 or 1.5 ≦ Ratio: sstt = 0 States 10 and 11 in FIG. 7 are states in which the background noise is estimated to be sufficiently small. By making the transition conditions from these states to other states (20 or more) sufficiently strict, it is possible to avoid erroneous filter operation for noise reduction despite the absence of background noise.

Normally, the coefficient of the Kalman filter 18 is calculated for each sample. The amount of processing required to calculate this coefficient is larger than the filtering operation itself. on the other hand,
Since the AR model coefficient and the power value used for the calculation of the filtering coefficient take a constant value for each frame (for example, the number of samples N in one frame is 160), the filtering coefficient converges to a constant value in the latter half of the frame. Therefore, instead of updating the filter coefficient for every sample, the filter coefficient is updated for the first n samples of the frame, for example, 3 samples, and the filtering process is performed using the obtained coefficient to (N-n)
For the sample, the filter coefficient is not updated, and filtering is performed using the coefficient obtained at the nth sample.
In this way, the calculation amount and processing time can be reduced.

In the above description, the coefficient a by which the input signal power is multiplied
Performs adaptive control according to the estimated noise power level and voice power level. That is, when the noise power is small or the input SRN (signal-to-noise ratio) is high, the direction in which the Kalman filter 18 strongly operates as described above is set,
When the noise power is high or the input SRN is low, the Kalman filter 18 is set to operate weakly. If strong noise reduction is performed when the input SNR is low, unnaturalness of the audio signal increases, and the coefficient a for avoiding this is adaptively controlled. Input SNR
10A, the coefficient a is linearly changed between a large value at a small SNR and a small value at a large SNR, as shown in FIG. 10A.

Further, as shown in FIG. 10B, when the amplitude of the input signal is a predetermined value or less, that is, 1 / α or less of the maximum amplitude (for example, when the maximum amplitude is 1, α = 1000), After setting to zero, the process shown in FIG. 2 may be performed, that is, when the input signal is at a low level, it may be regarded as noise and suppressed. Figure 10C
, The power of the input signal is a predetermined value β (eg β /
Below the maximum input power = 30 dB), the input signal may be attenuated according to the power. That is, the processing shown in FIG. 2 may be performed after the processing similar to that of a so-called expander. Time-based (using a time constant) control may be applied to the attenuation by the expander.
This expander process is described in, for example, Research and Practical Report Vol. As with the panda, a time constant may be provided, the attack time may be, for example, about 3.0 ± 0.1 mS, and the recovery time may be about 1.35 ± 0.6 mS.

The various signal processes shown in FIG. 10 are not limited to the case where the level (that is, the amplitude or the power) of the input signal itself is used, but the power of the estimated noise statistic or the estimated voice statistic is used. Good. Further, these signal processes are shown in
11B as well as the case where the signal processing means 31 performs the processing as shown in A and then supplies the noise reduction means 32 shown in FIG.
The various processes shown in FIG. 10 may be performed on the output of the noise reduction means 32 as shown in FIG.

The signal processed by the noise reducing means 32 shown in FIG. 2 may be encoded by the voice encoding means 33 and output in FIG. 11C. As the encoding means 33, for example, V
SELP, PSI-CELP (Technical Report RC593-7
8-November 1993 "Pitch Syncron
ous Innovation CELP (PSI-C
ELP) ”) and the like are used.

As shown in FIG. 11D, the processing means 31 shown in FIG. 10 is provided between the noise reduction means 32 and the coding means 33.
May be inserted. As shown in FIG. 11E, the noise reducing means 32 may be provided on the output side of the audio decoding means 34 corresponding to the audio encoding means 33. Although not shown in the figure, in FIG. 11E, the signal processing means 31 may be provided before or after the noise reduction means 32.

In FIG. 1, the modification of the statistic in step S ₆ may be omitted. This is an embodiment of the invention of claim 1. Also, in FIG. 1, the check of whether or not there is a silent section in step S ₃ , and the estimation of the noise statistic in step S ₄ are omitted, the noise statistic is estimated prior to the call, and this estimated statistic is used. The statistics may be adaptively modified according to the state of each frame. This is an embodiment of the invention of claim 4 and is particularly effective in the case where the noise state does not fluctuate significantly like in fixed communication. In addition to the voice microphone, a noise microphone is provided, the voice statistic is estimated from the input signal from the voice microphone, the noise statistic is estimated from the input signal of the noise microphone, and various types of noise are estimated from both input signals. A state or a state transition may be detected, and the statistic may be adaptively corrected accordingly. This is the invention of claim 3.

10 and 11, noise reducing means 32
This applies not only to the invention of claims 1 and 2, but also to the invention of claim 3 or 4. The filtering process is not limited to the Kalman filter process.

[0034]

As described above, according to the first aspect of the invention, it is possible to avoid the trouble of estimating the statistical amount of noise each time before making a call, and only one recording system is required, and the system cost can be reduced. In addition, when the noise state fluctuates with time as in mobile communication, noise statistic estimation is performed following this fluctuation, and noise reduction is performed accordingly.

According to the third aspect of the invention, noise reduction is strengthened or weakened according to the state of the input signal for each short time, and the audibility is improved. According to the invention of claim 2, the effects of the inventions of claims 1 and 2 are combined. According to the invention of claim 9, the amount of calculation is reduced and the processing time is shortened. SN ratio of input signal (before noise reduction) and output signal (after noise reduction) when the invention of claim 2 is applied
The relationship with the SN ratio of is shown in FIG. From this figure, the SN ratio of the output signal is always larger than the SN ratio of the input signal,
It can be seen that the application effect is particularly large when there is no input signal, that is, when there is only background noise.

As described above, according to the present invention, the noise reduction effect is large, and therefore, when various kinds of encoding of the voice signal are performed, the noise reduction processing according to the present invention is performed prior to the encoding to obtain a good quality. A coded voice signal can be obtained.

[Brief description of drawings]

FIG. 1 is a flow chart showing an embodiment of the invention of claim 2;

FIG. 2 is a block diagram showing a noise reduction device to which the invention of claim 2 is applied.

3A is a diagram showing a definition example of a state of an input signal, and FIG. 3B is a diagram showing a state transition of an input signal.

FIG. 4 is a diagram showing a modified example of the estimated statistic in each state in the invention of claim 2;

FIG. 5 is a diagram showing another example.

FIG. 6 is an S of an input signal when the invention of claim 2 is applied.
9 is a diagram showing an example of the relationship between / N and S / N of an output signal. FIG.

FIG. 7A is a diagram showing another example of the definition of the state of the input signal;
B is a diagram showing another example of the state transition of the input signal.

FIG. 8 is a diagram showing conditions of state transition in FIG. 7.

FIG. 9 is a view showing a continuation of FIG. 8;

FIG. 10 is a diagram showing a characteristic example of various signal processes used before and after the noise reduction process.

11A and 11B are block diagrams showing a combination of noise reduction means and signal processing means, noise reduction means and speech coding,
It is a block diagram which shows the combination with a decoding means.

Claims (9)

[Claims] 1. A voice statistic is estimated from an input signal, a filter coefficient is calculated using the voice statistic and a noise statistic in the input signal, and the input signal is calculated by the filter coefficient. A noise reduction method for filtering, comprising detecting a silent section in the input signal and estimating a statistical amount of the noise from the input signal in the detected silent section. 2. The state of voice and noise in an input signal is divided into predetermined states, each state is stored as a frame state table, and the input signal is processed at regular time intervals (frames). And determining the state of each frame by referring to the frame state table, and adaptively correcting the voice statistic and the noise statistic according to the determined frame state. The noise reduction method according to Item 1. 3. A voice statistical amount is estimated from an input signal recorded by a voice microphone, a noise statistical amount is estimated from an input signal recorded by a noise microphone, and the voice estimated statistics amount and the noise estimation are estimated. In the noise reduction method that calculates the filter coefficient using the statistic and the input signal is filtered by the filter coefficient, the state of voice and noise in the input signal is divided into predetermined states, and each state is divided. Is stored as a frame state table, the both input signals are processed at fixed time intervals (frames), the state of each frame is determined by referring to the frame state table, and the state of each frame is determined according to the determined frame state. A noise reduction method comprising adaptively correcting the voice statistic and the noise statistic. 4. A noise reduction method for estimating a voice statistic from an input signal, calculating a filter coefficient using the voice statistic and noise statistic, and filtering the input signal with the filter coefficient. In the above, the statistics of the noise are estimated from the input signal in the silent section before the call, the states of the voice and noise in the input signal are divided into predetermined states, and each state is stored as a frame state table. The input signal is processed at regular intervals (frames), the state of each frame is determined by referring to the frame state table, and the statistics of the voice and the noise are determined according to the determined frame state. A noise reduction method characterized by adaptively modifying the statistic of and. 5. The noise reduction method according to claim 2, wherein the adaptive correction is performed according to the transition of the state. 6. The noise reduction method according to claim 2, wherein the correction amount of the adaptive correction is adaptively controlled according to a signal-to-noise ratio or noise power. 7. The noise reduction method according to claim 1, wherein the input signal or the output signal is equal to or less than a predetermined value, and the input signal or the output signal is attenuated according to the level thereof. . 8. The noise reduction method according to claim 7, wherein the attenuation is temporally controlled. 9. In the first n samples of each frame (n is an integer of 2 or more and smaller than the number N of samples of one frame), the filter coefficient is updated for each sample, and the remaining (N−n 9. The noise reduction method according to claim 1, wherein the filter coefficient is not updated in the sample.