WO1999050825A1

WO1999050825A1 - Noise reduction device and a noise reduction method

Info

Publication number: WO1999050825A1
Application number: PCT/JP1998/005512
Authority: WO
Inventors: Kazutaka Tomita
Original assignee: Mitsubishi Denki Kabushiki Kaisha
Priority date: 1998-03-30
Filing date: 1998-12-07
Publication date: 1999-10-07
Also published as: CA2291826A1; AU721270B2; AU1352599A; KR20000076037A; KR100314332B1; CN1258368A; EP0992978A4; EP0992978A1

Abstract

A noise reduction method which reduces unpleasant residual noise even when movement to or from noise frames is frequent and the spectrum of each frame unevenly includes certain frequencies. In a spectral subtraction noise reduction method for producing an output by subtracting by a spectral subtraction filter an estimated noise amplitude spectrum from an amplitude spectrum created by orthogonal-transforming an input signal segmented to a predetermined frame length, the subtraction rate of the spectral subtraction filter is variable according to the estimated noise amplitude spectrum and an amplitude adjusting filter circuit (6) is used. The amplitude adjust filter circuit (6) multiplies the output produced by subtraction by the spectral subtraction filter for each frame, by an amplitude adjusting coefficient to produce a desired output. The amplitude adjusting coefficient is determined by the power of the amplitude spectrum (9) and the power of the estimated noise amplitude spectrum (10).

Description

Description Noise reduction device and noise reduction method

The present invention is applied to a voice communication system or a voice recognition system used in an environment having background noise, and a noise reduction method for a voice signal that performs noise suppression by removing noise from an input voice signal and a noise section. It relates to a noise reduction method using amplitude suppression. Background art

Conventionally, a method shown in FIG. 15 below is known as a means for reducing background noise. That is, the input signal in which the background noise is superimposed on the voice is converted into an AZD (analog Z-digital) signal and divided into a fixed section (frame). (Hereinafter referred to as a frame signal) is converted to a frequency spectrum by a discrete Fourier transform, and then divided into an amplitude spectrum and a phase spectrum. By subtracting the estimated noise amplitude spectrum estimated from the silent section from the amplitude spectrum using a subtraction filter, the audio amplitude spectrum with reduced background noise (output amplitude spectrum) Is estimated as the frequency spectrum of the voice with the background noise reduced along with the previous phase spectrum, and the frequency spectrum of this voice is inversely dispersive Fourier transformed and the output signal is obtained. A scan Bae click Tsentralnyi sub tiger click Chillon becomes how obtained. This is S TEVEN F. BAL L "S upprssionofacusticnoi seinspeechusingspectr alsubtraction," I EEE T rans. A coust., Speeshand Signal Proc., vol. AS SP-29, pp. 113-120, Apr. 1979.

FIG. 15 is a block diagram showing the configuration of this noise reduction method. First, the overall operation will be described with reference to FIG.

The input signal 107 cut out to a predetermined frame length is converted into the frequency domain by the Fourier transform unit 101, and the input phase spectrum 108 and the input amplitude spectrum 109 are output. You. In addition, the input signal is determined by the noise section determination unit 102 to be a speech frame (section) if it is equal to or more than the threshold value TH, and a noise frame (section) otherwise.

When the current frame is determined to be a noise frame, the estimated noise amplitude spectrum calculation unit 103 weights and adds the input amplitude spectrum 109 at that time to the previous estimated noise amplitude spectrum. And outputs the estimated noise amplitude spectrum 110 updated.

The subtraction 'filter section 104 has a characteristic whose representative transfer function is represented by the following equation (1).

F (ω): Subtraction 'filter

S (ω): Input amplitude spectrum

Ε (Ν (ω)): Estimated noise amplitude spectrum

r: removal rate

In Equation (1), S (ω) is the amplitude spectrum term of the input signal, Ε (Ν (ω)) is the estimated noise amplitude spectrum term, and r is a constant representing the removal rate of the estimated noise amplitude spectrum. Increasing r increases the amount of noise suppression; conversely, decreasing r decreases the amount of noise suppression. The subtraction 'filter section 104 calculates the removal rate _r based on a predetermined equation.

Further, based on the subtraction rate r, the input amplitude spectrum and the estimated noise amplitude spectrum are subtracted in the same manner as in equation (1), and the difference output amplitude spectrum 111 is obtained. Output.

In other words, the subtraction / filter unit 104 reduces the removal rate for noise frames and low power frames such as consonants to reduce noise suppression.

In the first conventional spectral subtraction method, an output signal returned to the time domain by the inverse Fourier transform unit 105 is output. However, there is a disadvantage that the quality is deteriorated due to the scraping. The larger the value of r is, the more the sound is cut.

To solve this problem, a method of making the subtraction rate of the subtraction filter variable for each frame is known from Japanese Patent Application Laid-Open No. Hei 8-22092. In this method, when F (ω) in Equation 1 is used as the subtraction filter, r is increased when the input is large, and r is decreased when the input is small. Has reduced shavings.

As a second conventional spectral subtraction method, in order to enhance the noise reduction effect, a method is used in which a function called a window function is defined on the time axis of the signal to be subjected to Fourier transform and multiplied and weighted. Have been. For example, a frame signal and a continuous signal with a certain length (overlap signal) are multiplied by a window function and subjected to Fourier transform. The output frame signal of the inverse Fourier transform is added to the first weighted sample below 1 The method of adding the overlap signal of the output of the inverse Fourier transform of the frame and outputting it is: North America CDMA mobile phone system (TIA / EIA, IS-127, "Enhanced Variable Rate Code Service Option 3 for Wideband S pread Spectrum Digital Systems") I have. Figure 16 shows the window function used here.

Here, a window function is used so that the sum of the weights of the samples corresponding to the addition of the frame signal and the overlap signal becomes 1, and that both ends of the Fourier transform target signal smoothly approach 0. By performing weighting using such a window function, the effect of increasing the estimation accuracy of the noise amplitude spectrum is obtained, and as a result, the noise reduction effect is improved.

As a third conventional spectral subtraction method, there is a method of performing a Fourier transform on a frame signal and a signal of a fixed length (overlap signal) following the frame signal without weighting by a window function. Yes, it is used in the noise reduction device disclosed in Japanese Patent Application Laid-Open No. Hei 9-134449. In this case, in order to smoothly connect the output between the frames, a waveform shaping process is performed in which the overlap signal of the inverse Fourier transform output of the immediately preceding frame is superimposed on the triangular window with the frame signal of the inverse Fourier transform output. And outputs the same, and saves the inverse Fourier transform output overlap signal for overlap processing of the next frame. Equation (2) shows the waveform shaping process at this time.

(j = 0-L-1)

Oj

~ L + M— 1)

Zj = 〇 _{M + j} (j = 〇 ~ L-1)

(2)

However, 0 ”·: Output signal

D j: Frame signal

Z j: Overlap signal

M: Frame length

L: overlap signal length

It is.

The advantages of this method are described below.

In order to implement the spectral subtraction method, a 16-bit fixed-point digital signal processor (fixed-point DSP), which has a theoretically limited calculation accuracy, is often used. The operation precision of the fixed-point decimal point D SP can be increased by double-precision arithmetic operation or carry by shift. However, there is a problem that the amount of calculation for this increases.

The Fourier transform and the inverse Fourier transform basically require a large amount of computation, and there are many cases where it is desired to realize the computation with a small amount of computation at the expense of a little computation accuracy. When this method is used, the frame range to be Fourier-transformed is not weighted by the window function, so the dynamic range is not widened, and the accuracy of the output frame signal that has undergone limited Fourier transform and inverse Fourier transform is degraded. Less is. For this reason, a fixed-point DSP that requires a small amount of computation is used.

The fourth conventional noise reduction method using spectral subtraction is

In order to calculate the estimated noise amplitude spectrum, first, the power of the noise is estimated, and this is used as a noise section determination threshold. If the input power is smaller than this threshold, it is determined that the frame is a noise frame. Further, the estimated noise amplitude spectrum is generally calculated by averaging the input signals of a plurality of frames in the section determined to be noise, that is, the average of the amplitude spectrum of the noise signal. For example, the spectral sub in Japanese Patent Application Laid-Open No. This method is used in fractionation.

As a fifth conventional method of reducing background noise, there is a method of reducing noise sensation by suppressing the amplitude of a noise section. A method of determining the state of the current frame from a predetermined finite state representing the state of the signal and determining the state of the current frame when the state corresponds to the state of a noise frame is disclosed in Japanese Unexamined Patent Publication No. — Indicated by 3 8 4 5 4.

The first conventional spectral subtraction method has a drawback that a noise section is deformed and becomes unpleasant residual noise called musical noise, which is the biggest practical problem.

This is because in the noise section, the output amplitude spectrum from which the estimated noise amplitude spectrum has been subtracted has a shape in which power is concentrated in a part of the frequency, and the frequency in which the power is concentrated is not per frame. It changes by changing the rules. As a method of solving musical noise, a method of varying the removal rate of the estimated noise amplitude spectrum is used, and the removal amount may be reduced in a noise frame, but the noise suppression amount is insufficient in a noise section. There is.

In the second conventional spectral subtraction method, the noise reduction ability is improved by performing a Fourier transform on a weighted signal whose both ends approach 0 using a window function. However, there is a problem in that the accuracy of the signals at both ends is reduced when realizing with fixed-point DSP, so that both ends of the output signal of the inverse Fourier transform are degraded and abnormal noise occurs at frame boundaries.

Conversely, if the unweighted signal is Fourier-transformed as in the third conventional example, accuracy can be easily achieved when implementing it with a fixed-point DSP, but the noise amplitude spectrum included in the input amplitude spectrum can be improved. Since the variation of the torque component between frames becomes larger than when weighted, the noise reduction ability is inferior to that when weighted. In the noise section determination of the fourth conventional spectral subtraction, if the noise section determination threshold is set high so that the noise frame can be correctly determined when the power fluctuation of the noise is large, some of the speech frames may also be noise frames. There is a problem in that the speech noise leaks into the estimated noise amplitude spectrum, causing the speech to be cut off. Conversely, to prevent this, if the noise interval determination threshold is set low so that the audio frame is not determined to be a noise frame, the noise frame is determined to be a voice frame, and the estimated noise amplitude spectrum is updated. However, there is a problem that the noise reduction is not performed correctly, and as a result, the noise reduction ability is deteriorated.

In the method of reducing noise by suppressing the amplitude of the fifth noise section, the frame state of the current frame is determined from a predetermined finite state representing the state of the signal, and this frame state is a noise frame state. In this case, if a method is adopted in which the amplitude is suppressed to a certain level, the frame state frequently moves between the voice frame state and the noise frame state in a short period of time depending on the type of input noise. The intensity of the power changes frequently, resulting in unstable output power and poor hearing.

The present invention has been made in order to solve the above-mentioned problem, and unpleasant residuals are obtained even when noise frames frequently move to or from noise frames and the spectrum of each frame is biased to some frequencies. The aim is to obtain a noise reduction method with less noise. Disclosure of the invention

The noise reduction apparatus according to the present invention subtracts an estimated noise amplitude spectrum from an amplitude spectrum obtained by orthogonally transforming an input signal cut out to a predetermined frame length by a spectral subtraction filter. Noise reduction device to obtain output For each frame, the required power is calculated by multiplying the reduced power by the spectral subtraction filter by an amplitude adjustment coefficient determined by the power of the amplitude spectrum and the power of the estimated noise amplitude spectrum. An amplitude adjustment filter unit for obtaining an output is provided.

The spectral subtraction 'filter makes the subtraction rate of the subtraction of the estimated noise amplitude spectrum variable according to the estimated noise amplitude spectrum.

The amplitude adjustment filter unit is characterized in that the amplitude adjustment coefficient is made variable in accordance with the withdrawal rate.

When the subtraction rate is high, the amplitude adjustment filter unit increases the amplitude adjustment coefficient to enhance noise suppression in the voice section and increases the output value of the output signal. When the rate is small, the amplitude adjustment coefficient is reduced, thereby reducing noise suppression in a noise section and reducing the output value of the output signal. It is characterized by multiplying the subtraction output by the central subtraction filter with the time-domain amplitude spectrum subjected to the inverse orthogonal transform.

The amplitude adjustment filter section is characterized in that the amplitude spectrum is multiplied by the amplitude adjustment coefficient in the frequency domain for each frame, and an inverse orthogonal transform is performed to obtain an output.

The amplitude adjustment coefficient is obtained by adding the amplitude adjustment coefficient of the current frame obtained based on the difference between the power of the amplitude spectrum of the input signal of the current frame and the power of the estimated noise amplitude spectrum of the current frame to the previous frame. It is characterized in that the value of the amplitude adjustment coefficient obtained in (1) is weighted and added.

A noise reduction apparatus according to the present invention provides an estimated noise amplitude spectrum from an amplitude spectrum obtained by orthogonally transforming an input signal cut out to a predetermined frame length. In a noise reduction device that obtains an output by subtracting with a spectral subtraction 'filter,

An input signal generation unit that cuts out a set section before and after the current frame and multiplies by a weighting function of 1 or less and adds the input signal before and after the current frame so that the end is close to 0, and outputs the output of the input signal generation unit Is used as an input signal to obtain an amplitude spectrum.

A noise reduction apparatus according to the present invention includes a waveform shaping processing unit that shapes a waveform of a current frame using a set section after a previous frame multiplied by one weighting function.

The noise reduction apparatus according to the present invention subtracts an estimated noise amplitude spectrum from an amplitude spectrum obtained by orthogonally transforming an input signal cut out to a predetermined frame length by a spectral subtraction filter. In a noise reduction device that obtains output,

A subtraction rate calculation unit that obtains an average noise power from a plurality of frames of the estimated noise amplitude spectrum and compares the average noise power with a determination threshold of the noise frame to obtain a removal rate;

The subtraction rate used for the subtraction of the spectral subtraction filter is set to the above-described subtraction rate.

The noise reduction apparatus according to the present invention includes an average noise power calculation unit that calculates an average noise power from a plurality of frames of the estimated noise amplitude spectrum, and the amplitude adjustment filter unit includes a spectral sub-frame for each frame. The required output is obtained by multiplying the subtraction output by the task filter by an amplitude adjustment coefficient determined by the power of the amplitude spectrum and the average noise amplitude power obtained above.

In the noise reduction method according to the present invention, an estimated noise amplitude spectrum is obtained from an amplitude spectrum obtained by orthogonally transforming an input signal cut out to a predetermined frame length. In the noise reduction method of obtaining an output by subtracting with a spectral subtraction filter,

Varying the removal rate of the spectral subtraction filter according to the estimated noise amplitude spectrum;

The required power is calculated by multiplying the reduced power by the above-mentioned spectral subtraction filter for each frame by an amplitude adjustment coefficient determined by the power of the above-mentioned amplitude spectrum and the power of the above-mentioned estimated noise amplitude spectrum. It is characterized by a process for obtaining output.

The noise reduction method according to the present invention is characterized in that an estimated noise amplitude spectrum is subtracted from an amplitude spectrum obtained by orthogonally transforming an input signal cut out to a predetermined frame length by a spectral subtraction 'filter. In the noise reduction method to obtain output,

A subtraction rate calculation step of obtaining an average noise power from a plurality of frames of the estimated noise amplitude spectrum and comparing the average noise power with a noise frame determination threshold to obtain a removal rate;

A step of setting the removal rate of the spectral subtraction filter to the removal rate. BRIEF DESCRIPTION OF THE FIGURES

FIG. 1 is a block diagram showing a configuration of a spectral subtraction method according to Embodiment 1 of the present invention.

FIG. 2 is a flowchart showing the operation performed by the amplitude adjustment filter in the noise reduction method according to the first embodiment.

FIG. 3 is a diagram illustrating an operation performed by the noise section determination unit according to the first embodiment.

FIG. 4 is a diagram for explaining the operation performed by the amplitude adjustment filter according to the first embodiment. is there.

FIG. 5 is a diagram illustrating an output signal of the spectral subtraction method according to the first embodiment.

FIG. 6 is a block diagram showing a configuration of a spectral subtraction method according to Embodiment 2 of the present invention.

FIG. 7 is a diagram showing a configuration of a spectral subtraction method according to Embodiment 3 of the present invention.

FIG. 8 is a diagram illustrating an example of an input signal in the noise reduction method according to the third embodiment.

FIG. 9 is a diagram illustrating an example of a weighting function multiplied by the input signal generation unit in the noise reduction method according to the third embodiment.

FIG. 10 is a diagram illustrating an example of an output signal of the inverse Fourier transform unit in the noise reduction method according to the third embodiment.

FIG. 11 is a block diagram showing a configuration of a spectral subtraction method according to Embodiment 4 of the present invention.

FIG. 12 is a diagram illustrating an operation in the noise reduction method according to the fourth embodiment.

FIG. 13 is a diagram showing a configuration of a spectral subtraction method according to the fifth embodiment of the present invention.

FIG. 14 is a diagram illustrating an operation in the noise reduction method according to the fifth embodiment.

FIG. 15 is a diagram showing a configuration of a conventional noise reduction method.

FIG. 16 is a diagram showing a window function used in the conventional noise reduction method. FIG. 17 is a comparison diagram of the operation of the first embodiment.

FIG. 18 is a diagram illustrating a waveform shaping process according to the third embodiment. BEST MODE FOR CARRYING OUT THE INVENTION

Embodiment 1.

A noise reduction method according to the first embodiment of the present invention in which the subtraction rate of the spectral subtraction filter is made variable and an amplitude adjustment filter section is provided to further subtract an amplitude from a normal output depending on the extraction rate will be described. You.

FIG. 1 is a block diagram illustrating a configuration of a noise reduction method according to the present embodiment. FIG. 2 is a flowchart illustrating an operation of the amplitude adjustment filter unit in FIG. FIG. 17 is a diagram showing a comparison of operations in a speech section and a noise section. First, the overall operation will be described with reference to FIG.

The input signal 7 cut out to a predetermined frame length is orthogonally transformed by the Fourier transform unit 1 and transformed into the frequency domain, and the input phase spectrum 8 and the input amplitude spectrum 9 are output. In addition, the input signal is judged as a speech frame (segment) if it is equal to or greater than the threshold value TH, and a noise frame (segment) if not more than the threshold value TH. Here, if the following equation (3) holds, replace it with a new threshold ΤΗ.

P _in <2.0 TH (3)

TH _new = 0.9 TH + 0. l P _in (4)

Here, P _in is the input signal power, TH is before updating the noise section determination threshold value, TH _{new new} is the noise section determination threshold updated.

Thus, the relationship shown in FIG. 3 is obtained. When the current frame is determined to be a noise frame, the estimated noise amplitude spectrum calculation unit 3 replaces the input amplitude spectrum 9 at that time (with the estimated noise amplitude spectrum up to that point) by, for example, the following equation (5). Then, the estimated noise amplitude spectrum 10 updated by weighting and adding is output.

E (N (ω)) = E _old (Ν (ω)) α + S (ω) α) (5)

Ε (Ν (ω)) is the updated estimated noise amplitude spectrum,

E _old (N (ω)) is the estimated noise amplitude spectrum before updating,

S (ω) is the amplitude spectrum of the input signal,

0 <α <1

It is.

The subtraction 'filter unit 4 calculates the removal rate _r based on the following equation (6), and outputs it as the noise removal strength 13.

r = min {1.0, r _TH · P OWs / P OW _N } (6)

POWs is the percentage of the amplitude spectrum of the input signal at all frequencies in the current frame.

POW _N is the power of the estimated noise amplitude spectrum for all frequencies of the current frame 0 <r _TH <1

Further, based on the subtraction rate r, the input amplitude spectrum and the estimated noise amplitude spectrum are subtracted in the same manner as in equation (1), and the difference output amplitude spectrum 11 is output. I do.

In other words, in the subtraction 'filter unit 4 (smooth subtraction' filter), the noise suppression is reduced by reducing the extraction rate for noise frames and low power frames such as consonants.

The output amplitude 12 returned to the time domain by the inverse Fourier transform unit 5 has a smaller amplitude adjustment coefficient in the next amplitude adjustment filter unit 6 in a frame having a small power, so that the amplitude suppression is enhanced.

That is, the amplitude adjustment filter unit 6 is configured to perform steps 3 to 3 in FIG. Perform two actions. In the first step S1, the noise removal strength 13 (removal rate r) is input. Then, the difference between the power of the amplitude spectrum of the speech section of the input signal of the current frame and the power of the estimated noise amplitude spectrum of the current frame is obtained.

(POW _s -POW _N ≥ 0 and P〇W _S — r-POW _N > 0)

G = 0

From (P〇W _S — POW _N 0 0 or P〇W _S — r · POW _N ≤ 0), the amplitude suppression coefficient G for each frame is obtained by the following equation (7).

(7)

Here, POWs is the power of the amplitude spectrum of all the frequencies of the input signal of the current frame, POWn is the power of the estimated noise amplitude spectrum of all the frequencies of the current frame, and G is the amplitude suppression coefficient of the current frame. , R is the withdrawal rate. As shown in Fig. 17, when the withdrawal rate r is large, the amplitude suppression coefficient G increases. When the withdrawal rate r is small, the amplitude suppression coefficient G becomes small.

Further, in S2, the amplitude adjustment coefficient g (n

).

g (n) = g (n-1) AR + G (1— AR) (8)

However, it is 0 and AR is 1.

G (n) is the amplitude adjustment coefficient of the n-th sample of the current frame, and n-1 is the previous sample. As shown in Fig. 17, when the withdrawal rate r is large, the amplitude suppression coefficient G increases, and the amplitude adjustment coefficient g (n) increases. When the removal ratio r is small, the amplitude suppression coefficient G becomes small, and the amplitude adjustment coefficient g (n) becomes small.

Further, in S 3, the output amplitude 1 2 from the inverse Fourier transform unit 5 and Table Wa at S _in, by the following equation (9), the output signal is multiplied by the amplitude adjustment coefficient g (n) Get 1 4

S _out (n) = S _in (n) X g (n) (9)

FIG. 4 is a diagram showing the input / output characteristics of the amplitude adjustment filter unit 6.In the noise section, the extraction rate r is small, so that the amplitude adjustment coefficient g (n) is small, and the suppression on the output side is large, and the output value is large. On the other hand, the amplitude adjustment coefficient _g ( _n ) is large because the withdrawal rate r is large in the voice section. On the output side, the suppression is small and the output value is relatively large.

According to this embodiment, the amount of noise removal in the frequency domain is reduced in a noise frame or a consonant voice frame having a small power, so that the generation of musical noise in the noise frame can be suppressed and the voice section can be prevented from being cut. . Figure 5 shows the amplitude spectrum of the noise frame before and after the subtraction.

When the removal rate _{r is set} to 1.0 and the noise is largely removed, only the frequency component of the strong amplitude spectrum of the removed noise remains and becomes a source of musical noise. When the withdrawal rate r is set to 0.5 and the withdrawal is weakened, the deviation of the amplitude to a part of the frequency in the amplitude spectrum of the noise after the withdrawal is reduced, and no musical noise is generated. In the frame in which the removal amount is reduced, the amplitude is suppressed accordingly, so that it is possible to prevent the noise suppression amount from becoming insufficient. In addition, in the amplitude adjustment filter unit 6 according to the present embodiment, the amplitude adjustment coefficient is gradually changed in the time axis direction for each sample even in one frame by equation (8), so that Even if the amplitude is suddenly adjusted such as at the rising edge, a natural output can be obtained.

Embodiment 2.

The amplitude adjustment may be performed in the frequency domain. That is, an amplitude adjustment filter section is provided after the Fourier transform, and the signal after the amplitude adjustment is subjected to the inverse Fourier transform. Configuration.

The method according to the second embodiment of the present invention will be described with reference to the drawings.

FIG. 6 is a block diagram showing a configuration of a spectral subtraction method according to the embodiment.

The operation of the method will be described with reference to FIG.

The configuration in FIG. 6 shows a configuration in which the amplitude adjustment filter in FIG. 1 is moved immediately after the subtraction filter 4 and the amplitude adjustment is performed in the frequency domain. The operation is the same as that of mode 1. In the figure, the subtraction ′ filter unit 4 calculates the subtraction rate by the equation (6), and subtracts the estimated noise amplitude spectrum 10 from the input amplitude spectrum 9 based on the subtraction rate, Outputs the output amplitude spectrum. The amplitude adjustment filter section 15 multiplies the amplitude adjustment coefficient calculated based on the withdrawal rate r by the output amplitude spectrum, and outputs a final output amplitude; spectrum.

According to this embodiment, since the withdrawal strength is reduced in a noise frame or a voice frame having a low power such as a consonant, the generation of musical noise can be suppressed in the noise frame, and the deformation or disappearance of the consonant section can be prevented. In frames with a reduced amount of subtraction, amplitude suppression is performed accordingly, so that it is possible to prevent shortage of noise suppression.

Also, the amplitude adjustment coefficient need only be calculated once per frame, and it is not necessary to calculate the amplitude adjustment coefficient for each sample of the signal as in S3 of FIG. 2 of the first embodiment. .

Embodiment 3.

In the above embodiment, since the input signal extracts the voice for each section divided by the predetermined frame, the input signal becomes discontinuous at the change point of the frame (section), which may cause abnormal noise. . In the present embodiment, this is Better to smooth signal changes between frames.

Hereinafter, a method according to an embodiment of the present invention will be described with reference to the drawings.

FIG. 7 is a block diagram showing a configuration of a spectral subtraction method according to Embodiment 3 of the present invention. The input signal generation unit 19 and the waveform shaping processing unit 20 are new elements, and the other elements are the same as those in FIG. 1. FIG. 8 shows an example of the input signal 7 in the noise reduction method according to the third embodiment. FIG. 9 is a diagram illustrating a weighting function for multiplying the input signal performed by the input signal generation unit in the noise reduction method according to the third embodiment. FIG. 10 is a diagram illustrating an example of an output signal 5 a of the inverse Fourier transform unit 5 in the noise reduction method according to the third embodiment.

The operation of the noise reduction method according to the present embodiment will be described based on the configuration diagram of FIG.

FIG. 8 shows the time-series amplitude of the input signal 7 after the AD conversion, and this is input to the input signal generation unit 19. When the frame signal to be subjected to noise suppression processing in the current frame is s (i) (i = n, n + 1,..., Η + fr), the input signal Cut out the added s (i) (i = n−a, n−a + 1,..., N + fr + sm + a). A in Fig. 8 is the set section before the frame. B and C in Fig. 8 are the set sections after the frame. Next, the weighting function of FIG. 9 is multiplied to obtain a signal s ′ (i) (i = n−a, n−a + 1,..., N + fr + s m + a) is calculated. In the setting section of A and C in Fig. 8, the weight is set so that the end approaches 0. In the setting section B in Fig. 8, the same 1 weight as the frame signal is applied.

s' (i) = S (i)-W (i -n + a)

(l = n—a, n—a + 1,..., n + fr + sm + a) ( Ten)

The signal s ′ (i) (i = n−a, n−a + 1,..., N + fr + sm + a) is subjected to Fourier transform by the Fourier transform unit 1. Assuming that the output signal of the inverse Fourier transform unit 5 is u (i) (i = 0, 1,..., 2a + fr + sm), FIG. 10 shows u (i), and U (i) (i = 0, 1, ..., 2a + fr + sm) is s (n-a + i) (i = 0, 1, ..., 2a + fr + sm).

As shown in FIG. 18, the waveform shaping processing unit 20 converts the inverse Fourier-transformed signal corresponding to the frame signal into u (i) (i = a, a + 1,..., A + fr). Waveform shaping processing to prevent discontinuity between the signals is output. The following equation (11) gives u (i) (i = a + fr, a + fr + 1, ..., a + fr + sm) of the inverse Fourier transform output of the set section B after the previous frame. This is the expression representing the output signal after waveform shaping processing when is set to up (i) (i = 0, 1,..., Sm).

u '(i)-(u (i)-i + up (i-a)

• s m— 1 no) / s m

(i = a, a + 1,..., a + s m) (1 1)

u '(i): Signal after waveform shaping processing

According to this embodiment, the weighted signal whose end approaches 0 is subjected to Fourier transform, so that the noise reduction capability is improved. Also, when this system was implemented with a fixed-point arithmetic digital signal processor (fixed-point DSP (digital signal processor)), the signals at both ends (A and C in Fig. 18), whose accuracy deteriorates due to the weighting, are reduced. Since it is not used for waveform shaping, it is easy to ensure output accuracy and to prevent abnormal noise at frame boundaries.

Using DSP, fixed-point arithmetic with a word length as short as 16 bits is performed. Therefore, when processing signals with a large dynamic range, accuracy degradation occurs. Weighting the edges closer to zero will increase the dynamic range of the signal, and the use of DSPs will degrade accuracy. In this embodiment, the accuracy of the waveform shaping process is improved because the signal at the end where the accuracy degradation has occurred is not used for the waveform shaping process.

Embodiment 4.

As a method to enhance the noise reduction effect of spectral subtraction, we consider a method that makes it easy to identify noise sections and that reduces speech shaving due to overdrawing.

Hereinafter, the method according to the fourth embodiment of the present invention will be described with reference to the drawings.

FIG. 11 is a block diagram showing a configuration of a spectral subtraction method according to Embodiment 4 of the present invention. In the figure, there is a withdrawal rate calculator 16 as a new element. The other elements are the same as those with the same numbers in FIG.

FIG. 12 is a diagram illustrating an operation in the noise reduction method according to the present embodiment. In the figure, the solid line represents the input power P〇Ws in the noise section, the dotted line represents the average noise power POW _AVE , and the chain line represents the noise section determination threshold POW _TH . Note that POW _TH is the threshold TH in the pre-noise frame described in equation (3).

The operation of the noise reduction method according to the present embodiment will be described based on the overall configuration diagram of FIG.

This configuration has a configuration in which a removal rate calculation unit 16 is added to the overall configuration of the spectral subtraction of the conventional method. In FIG. 11, the filter of the subtraction filter unit 4 performs the operation of equation (1). The noise interval determination unit 2 calculates a noise interval determination threshold value POW _TH , and determines that the noise frame is a noise frame when the input level falls below this threshold value. The subtraction rate calculation unit 16 calculates the average power of a plurality of noise frames close to the current based on the determination result and sets the average power as the average noise power. Based on the following equation (12), the noise section determination threshold is determined. P〇W _TH , average noise power P

_{R 1} (! ₂ )

Calculate the ratio rl between the noise section judgment threshold value 17 and the average noise power when OW _AVE is set, and output this as the noise fluctuation magnitude 18.

The magnitude r1 of this noise fluctuation is defined as the noise vector removal rate r in equation (1). The noise section determination threshold POW _TH value uses a high value that can be correctly determined as a noise section in the case of noise with large power fluctuations. Figure 12 shows the changes in the average noise power and the noise section determination threshold with respect to the change in the input power in the noise section at this time.

According to this embodiment, even when the power fluctuation of the noise is large, as shown in FIG. 12, the average noise power POW _AVE of the noise frame takes a value smaller than the noise section determination threshold value POW _TH . Therefore, the withdrawal rate r 1 is smaller than 1. As a result, the effect of suppressing the removal of the estimated noise amplitude spectrum can be obtained, the voice shaving can be reduced, and the estimated noise amplitude spectrum is updated normally.

Embodiment 5

A method for reducing the unpleasant residual noise by suppressing the amplitude of the input signal regarded as noise in the noise section will be described.

Hereinafter, the method according to the fifth embodiment of the present invention will be described with reference to the drawings.

FIG. 13 is a block diagram showing main components of a noise reduction method according to Embodiment 5 of the present invention.

FIG. 14 is a diagram illustrating an operation in the noise reduction method according to the present embodiment. In this configuration, the noise is reduced by suppressing the amplitude in the noise section. The amplitude adjustment coefficient for performing the amplitude suppression is determined depending on the current input power and the average noise power.

The operation will be described with reference to FIGS.

The noise interval determination unit 301 calculates the input power POWs 305 indicated by the solid line in FIG. 12 from the input signal 304, and performs the noise interval determination using the noise interval determination threshold POW _TH similar to the first embodiment. Do. The average noise power calculation section 302 calculates the average power of a plurality of past noise frames close to the current frame based on the result of the noise section determination, and calculates the average noise power P OW _AVE 3 shown by the dotted line in FIG. 06.

The amplitude adjustment filter section 303 calculates the amplitude suppression coefficient G from the input power P〇Ws 305 of the input signal and the average noise power P OW _AVE 306 according to equation (13), and then directly adjusts the amplitude. The output signal 307 is obtained by multiplying the input signal 304 by the obtained amplitude adjustment coefficient.

_G : POWs POWAVE

POWs

(POW _s -POW _AVE ≥ 0)

G = 0

(POW _s -POW _AVE <0)

( 13)

Also, in order to make the amplitude adjustment coefficient change smoothly between frames using the amplitude suppression coefficient G calculated by equation (13), as shown in equation (8), G is smoothed for each sample. FIG. 14 shows the relationship between the input power P OWs 305 of the input signal 304 and the output power of the output signal 307 after the amplitude adjustment. The dotted line in Fig. 14 shows the case where the input signal was output as it was without using the amplitude adjustment filter. That is, the dotted line indicates a case where the input power of the input signal is equal to the output power of the output signal. The solid line shows the relationship between the input power POWs and the output power when the amplitude adjustment filter is used. This indicates that a value smaller than the dotted line is output by amplitude suppression. Also, when the input power is the average noise power P_〇_W _AVE smaller than the output signal Pawa one indicates that zero.

According to this embodiment, since the withdrawal rate is made variable and the amplitude adjustment filter section for changing the amplitude suppression degree based on the variable is provided, the amplitude adjustment coefficient is determined from the result of the voice state determination. Frequently, as in the case of direct determination, the amplitude suppression does not change significantly, and the audio output stabilizes. Industrial applicability

As described above, according to the present invention, since the withdrawal rate is made variable and the amplitude adjustment filter section that changes the degree of amplitude suppression based on this is provided, it is possible to prevent the voice section from being cut off. Thus, even if the subtraction strength in a noise frame or the like is reduced, noise is suppressed by performing amplitude suppression corresponding to the strength, so that a balanced and easy-to-hear output can be obtained as a whole.

Furthermore, since the amplitude adjustment is performed in the frequency domain, there is no need to calculate the amplitude adjustment coefficient for each signal sample, and the amount of calculation can be reduced.

Further, since the amplitude adjustment coefficient gradually changes in the time axis direction even within a frame, a natural output can be obtained even when a sharp amplitude adjustment such as the rising of a voice section is performed.

Further, the signals in the preceding and succeeding sections of the current frame are also multiplied by a weight function and added. Since the input signal generation unit includes the input signal generation unit, the target signal of the Fourier transform is weighted, the estimation accuracy of the noise spectrum is improved and the noise reduction effect is enhanced, and the unweighted signal of the inverse Fourier transform output is obtained. Is output as a frame signal, so that even with a fixed-point arithmetic digital signal processor, there is an effect that the arithmetic operation is small and high signal accuracy can be obtained.

Alternatively, the strength of noise removal is adjusted according to the variability of the section determined to be a noise section, so that even when the variability of noise is large, a correct noise determination threshold can be set. It also has the effect of preventing mixed speech components from squeezing the speech.

Furthermore, since the amplitude suppression amount is determined using the average power of the noise section that changes gradually without directly determining the amplitude suppression amount from the determination result of the noise section, the output is caused by frequent changes in the noise section determination. This has the effect of avoiding adverse effects.

Claims

The scope of the claims

1. A noise reduction apparatus that obtains an output by subtracting an estimated noise amplitude spectrum from an amplitude spectrum obtained by orthogonally transforming an input signal cut out to a predetermined frame length by a spectral sub-fraction filter. For each frame, the required power is calculated by multiplying the reduced power by the spectral subtraction filter by an amplitude adjustment coefficient determined by the power of the amplitude spectrum and the power of the estimated noise amplitude spectrum. A noise reduction device comprising an amplitude adjustment filter for obtaining an output.

2. The spectral subtraction filter makes the subtraction rate of the subtraction of the estimated noise amplitude spectrum variable according to the above-mentioned estimated noise amplitude spectrum, and the amplitude adjustment filter section changes the subtraction rate according to the subtraction rate. 2. The noise reduction device according to claim 1, wherein the amplitude adjustment coefficient is variable.

3. When the withdrawal rate is large, the amplitude adjustment filter unit increases the amplitude adjustment coefficient to enhance noise suppression in a voice section, increases the output value of an output signal, and increases the output value of the output signal. 3. The noise reduction apparatus according to claim 2, wherein when the withdrawal rate is small, the noise suppression is reduced in a noise section and the output value of the output signal is reduced by reducing the amplitude adjustment coefficient.

4. The noise reduction according to claim 1, wherein the amplitude adjustment filter section multiplies an amplitude spectrum in a time domain obtained by performing an inverse orthogonal transform on an output obtained by subtracting the amplitude adjustment coefficient from the spectral subtralation filter. Equipment.

5. The amplitude adjustment filter section multiplies the amplitude spectrum by the amplitude adjustment coefficient in the frequency domain for each frame, and performs inverse orthogonal transform to obtain an output. The noise reduction device according to claim 1, wherein

6. The amplitude adjustment coefficient is the amplitude adjustment coefficient of the current frame obtained based on the difference between the power of the amplitude signal of the input signal of the current frame and the power of the estimated noise amplitude spectrum of the current frame. 2. The noise reduction apparatus according to claim 1, wherein the value of the amplitude adjustment coefficient obtained in the previous frame is a value obtained by weighting and adding.

7. A noise reduction device that obtains an output by subtracting an estimated noise amplitude spectrum from an amplitude spectrum obtained by orthogonally transforming an input signal cut out to a predetermined frame length using a spectral subtralation filter. An input signal generation unit that cuts out a set section before and after the current frame and multiplies by a weighting function of 1 or less and adds the input signal before and after the current frame so that the end is close to 0, and outputs the output signal of the input signal generation unit A noise reduction apparatus characterized in that an amplitude spectrum is obtained by using the input signal as an input signal.

8. The noise reduction apparatus according to claim 7, further comprising a waveform shaping processing unit that shapes a waveform of the current frame using a set section after the previous frame multiplied by the weight function of 8.1.

9. A noise reduction device that obtains an output by subtracting an estimated noise amplitude spectrum from an amplitude spectrum obtained by orthogonally transforming an input signal cut out to a predetermined frame length by a spectral subtralation filter. A subtraction rate calculation unit that obtains an average noise power from a plurality of frames of the estimated noise amplitude spectrum and compares the average noise power with a determination threshold of the noise frame to obtain a removal rate;

A noise reduction apparatus characterized in that a subtraction rate used for subtraction of the spectral subtraction filter is defined as the subtraction rate.

10. An average noise power calculator that calculates the average noise power from a plurality of frames of the estimated noise amplitude spectrum is provided.

The amplitude adjustment filter section multiplies the subtraction output by the spectral subtraction filter for each frame by an amplitude adjustment coefficient determined by the power of the amplitude spectrum and the average noise amplitude power obtained as described above. The noise reduction device according to claim 1, wherein a required output is obtained.

1 1. Noise reduction to obtain the output by subtracting the estimated noise amplitude spectrum from the amplitude spectrum obtained by orthogonally transforming the input signal cut out to a predetermined frame length by using a spectral subtrac tion filter Varying the removal rate of the spectral subtraction 'filter in accordance with the estimated noise amplitude spectrum.

For each frame, the required power is calculated by multiplying the reduced power by the spectral subtraction filter by an amplitude adjustment coefficient determined by the power of the amplitude spectrum and the power of the estimated noise amplitude spectrum. A noise reduction method comprising a step of obtaining an output.

1 2. A noise reduction method in which an estimated noise amplitude spectrum is subtracted from an amplitude spectrum obtained by orthogonally transforming an input signal cut out to a predetermined frame length by a spectral subtraction filter to obtain an output. A subtraction rate calculating step of obtaining an average noise power from a plurality of frames of the estimated noise amplitude spectrum and comparing the average noise power with a noise frame determination threshold to obtain a removal rate;

A noise reduction method, comprising a step of setting the removal rate of the spectral subtraction filter to the removal rate.