JPH10171497A - Background noise removing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To remove a background noise and to obtain excellent sound quality generating no sense of incongruity with hearing sensation by splitting an input signal into a plurality of small frequency areas and operating subtraction processing on the background noise. SOLUTION: During a speech period in which a background noise and a speech signal are inputted to an input terminal 18 in a form of a mixture, a power spectrum corresponding to the mixture is outputted from a calculation part 4. A subtraction is executed between this power spectrum and an estimated background noise estimated by a background noise estimation part 7. A background noise re-estimation part 12 for each band checks an S/N difference between each channel based on an S/Na of the overall frequency band and S/Nm of each channel, and re-updates an estimated background noise of a channel with a low S/N. An adder 13 executes a subtraction between a m-th channel output of an adder 5 coming from a band splitter 101 and a re-updated estimated background noise coming from a by-band background noise re- updating part 122.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for removing background noise from an input signal in which a speech signal and background noise are mixed.

[0002]

2. Description of the Related Art As one method for removing background noise from an input signal in which a voice signal and background noise are mixed, an input signal is converted into a frequency component and divided into a predetermined low frequency component and a predetermined high frequency component. It is known to reduce the amplitude by applying a gain of less than 1 to the entire component and reduce the background noise for each spectrum by applying the frequency subtraction method (spectral subtraction) to the high frequency component. ing.

[0003]

However, in the above-described background noise elimination method, (a) the frequency characteristics of the background noise are various and unknown in a normal use state. It is difficult to decide. (b) Since processing is performed using different methods in the low frequency range and the high frequency range, matching at the division point between the low frequency range and the high frequency range is difficult, and the spectrum is continuous from the low frequency range to the high frequency range. Also for the frequency characteristics, a discontinuous processing point is generated at the frequency division point due to a difference in the processing method, which causes a sense of discomfort in hearing. (c) In the low frequency range
Since the power is only reduced without improving the S / N, effective noise reduction cannot be expected, especially when the audio component is concentrated in the low frequency range. Sometimes the S / N is rather degraded and the sound quality is also degraded.
And the like.

[0004] The present invention solves such disadvantages of the prior art, and eliminates background noise in a single method in the entire frequency component band without distinguishing between a low frequency region and a high frequency region. It is an object of the present invention to provide a background noise elimination device which is excellent in sound quality and does not cause noise.

[0005]

SUMMARY OF THE INVENTION In order to solve the above-mentioned problems, the present invention provides a background noise eliminator in which background noise is input as an input signal and then the background noise and voice signal are input in a mixed form. The apparatus comprises: a signal conversion unit for sequentially converting an input signal from a time axis signal to a frequency component in frame units; a voice detection unit for detecting a voice signal included in the input signal; and a voice signal detected by the voice detection unit. During the noise period that is not performed, the background noise estimating means for generating and holding the estimated background noise by averaging the frequency component converted by the signal converting means and the estimated background noise generated one frame before, and the voice detecting means A speech period in which the speech signal is detected, a first addition unit for subtracting the estimated background noise held by the background noise estimation unit from the frequency component converted by the signal conversion unit; The frequency component obtained by the subtraction of the first adding means is used as a signal, the estimated background noise held by the background noise estimating means is used as noise, and the S / N of the entire frequency band and the small area obtained by dividing the entire frequency band into a plurality Calculate S / N for each
S / N calculation means, and for each small area in which the difference between the S / N of the small area and the S / N of the entire frequency band is equal to or less than a predetermined value, the frequency component obtained by subtraction of the first addition means and the background noise estimation means A band-based background noise re-estimating means for generating a re-updated estimated background noise including the estimated background noise held at a predetermined ratio, and a re-updated estimated background noise from a frequency component obtained by subtraction of the first adding means. A second adding unit for subtracting the signal, and a signal reproducing unit for converting a frequency component for each frame obtained by the subtraction of the second adding unit into a time axis signal and outputting the time axis signal. The magnitude of the estimated background noise to be subtracted from the estimated background noise to be subtracted by the second adding means is set so that the background noise is removed from the input signal.

In order to solve the above-mentioned problems, the present invention provides a background noise elimination device in which background noise is input as an input signal and then the background noise and the audio signal are input in a mixed form. A signal converting means for sequentially converting the input signal from the time axis signal to a frequency component in frame units, a voice detecting means for detecting a voice signal included in the input signal, and a noise period in which no voice signal is detected by the voice detecting means. Background noise estimating means for generating and holding the estimated background noise by averaging the frequency component converted by the signal converting means and the estimated background noise generated one frame before, and detecting the audio signal by the audio detecting means The audio period is
A first adding means for subtracting the estimated background noise held by the background noise estimating means from the frequency component converted by the signal converting means; and a frequency component obtained by the subtraction of the first adding means being a signal, S / N calculation means for calculating the S / N of the entire frequency band and the S / N of each small area obtained by dividing the entire frequency band into a plurality of parts, using the estimated background noise held by the estimation means as noise, For each small region where the difference between the S / N and the S / N of the entire frequency band is equal to or less than a predetermined value, the frequency component obtained by the subtraction of the first adding means and the estimated background noise held by the background noise estimating means are determined. And a noise conversion unit that generates a renewed estimated background noise by adding the estimated background noise held by the background noise estimating unit to the noise, and a conversion of the signal converting unit. Estimation background from frequency components obtained by A second adder for subtracting a sound; and a signal reproducer for converting a frequency component for each frame obtained by the subtraction of the second adder into a time axis signal and outputting the signal. The magnitude of the re-updated estimated background noise to be subtracted in is set so that the background noise is removed from the input signal.

In this case, it is preferable that the voice detecting means detects the voice signal included in the input signal by detecting the voice component included in the frequency component obtained by the conversion of the signal converting means.

Further, the voice detecting means divides the frequency component obtained by the conversion of the signal converting means into a plurality of small areas, detects the voice component for each of the small areas, and determines the number of the small areas in which the voice component is detected. Is larger than or equal to a predetermined value, it may be determined that the input signal includes an audio signal.

The sound detecting means divides the frequency component obtained by the conversion of the signal converting means into a plurality of small areas and detects a sound component for each of the small areas. When the ratio of the number of the small regions is equal to or more than a predetermined value, it may be determined that the input signal includes the audio signal.

[0010]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing an embodiment of a background noise elimination apparatus according to the present invention.

FIG. 1 is a block diagram showing a background noise removing apparatus according to a first embodiment of the present invention. This first embodiment is:
Generates estimated background noise by removing the steep change component from the background noise and re-updated estimated background noise including the steep change of the background noise, removes the estimated background noise from the input signal, and further divides the frequency band into multiple Among the small regions, the small noise region having a particularly small S / N is used to remove the background noise re-updated, thereby removing the background noise accurately without deteriorating the sound quality.

In FIG. 1, a background noise N1 is first input to an input terminal 18, and then a speech signal S is applied to the background noise N1.
It is assumed to be input in a mixed form. For example, a signal output from a microphone of a mobile phone device, a car phone device, an audio device, or the like corresponds to the voice signal S and the background noise N1. The input signal (background noise N1,
Alternatively, the audio signal S and the background noise N1) are input to an analog-digital converter (A / D converter) 1 connected to the input terminal 18. The A / D converter 1 samples a signal from the input terminal 18 at a predetermined frequency, for example, 8 KHz in the case of a telephone device, converts an analog signal into a digital signal, and calculates a window function connected to the output side. To the detector 2 and the time axis sound detector 6.

A window function calculator 2 combines the digital signals from the A / D converter 1 into one frame for each predetermined number of samples, performs a window function calculation on each frame, and is connected to the output side. Output to the fast Fourier transform calculator (FFT calculator) 3. It is well known that a window function such as a Hanning window or a Hamming window is used in the window function calculation in order to prevent generation of high-frequency components based on data jumps between frames. The FFT operation unit 3 performs a fast Fourier transform operation on the signal from the window function operation unit 2 for each frame, and converts a time axis signal represented as a function of time t into a frequency component (spectral signal). Then, the spectrum signal is output to the power spectrum calculator 4 and the phase calculator 8 connected to the output side.

Here, the spectrum signal output from the FFT operator 3 when the background noise N1 is input to the input terminal 18 is
N1 (f, k) (f is frequency, k is frame number), audio signal
Assuming that the spectrum signal output from the FFT operator 3 when only S is input is S (f, k), the spectrum signal X (f, k) output from the FFT operator 3 is input to the input terminal 18 at the background. noise
In the noise period in which only N1 is input, it can be expressed by equation (1), and in the audio period in which background noise N1 and audio signal S are input to input terminal 18 in a mixed form, it can be expressed by equation (2) .

X (f, k) = N1 (f, k) (1) X (f, k) = N1 (f, k) + S (f, k) (2) The power spectrum calculator 4 performs an FFT operation. spectrum signal X (f, k) from the vessel 3 the power spectrum X _f of _{the k,} calculated for each frame. And the calculated power spectrum X
_{f and k} are output to the adder 5 and the background noise estimator 7 connected to the output side. The phase calculator 8 calculates the phase Φ (f, k) of each frequency component from the in-phase component and the quadrature component of each frequency component of the spectrum signal X (f, k) from the FFT calculator 3,
The phase Φ (f, k) is output to the phase holder 9 connected to the output side. The phase holder 9 receives the phase Φ from the phase calculator 8.
(f, k) is held for one frame period.

On the other hand, the time axis sound detector 6 detects whether or not sound is included in the digital signal from the A / D converter 1, and outputs the detection result to a background noise estimator 7 connected to the output side. Output to In the present embodiment, voice detection is performed on the time axis, and the difference between the average power of the input signal in a long time, for example, 500 ms, and the average power of the input signal in a short time, for example, 50 ms, is calculated. Is determined to include a voice in the input signal when exceeds a preset threshold value. It should be noted that the sound detection method is not limited to the present embodiment, and any method may be used as long as the sound can be detected on the time axis.

The background noise estimating unit 7 includes a switch 71, a background noise updating unit 72, and a background noise holding unit 73. During a noise period in which the time axis speech detector 6 does not detect speech, the background noise estimating unit 7 receives a signal from the power spectrum calculation unit 4. The background noise is estimated using the power spectrum _{Xf, k} . Specifically, the switch 71 is
The switch is closed only during the noise period based on the determination result from the time axis speech detector 6, and the output side of the power spectrum calculation unit 4 is connected to the input side of the background noise update unit 72. Therefore, the power spectrum from the power spectrum calculator 4
X _{f, k} is input to the background noise updating unit 72 only during the noise period.

The background noise updating section 72 includes a power spectrum X _{f, k} from the power spectrum calculating section 4 and a background noise holding section.
Using the estimated background noise N2 _{f, k-1} estimated one frame before held in 73, the estimated background noise N2
_{In addition} to calculating _{f, k} , the estimated background noise N2 _{f, k−1} one frame before, which is held in the background noise holding unit 73, is updated to the estimated background noise N2 _{f, k} . The calculation of equation (3) is executed for each frame and each frequency component.

N2f _{, k} = α · N2f _{, k−1} + (1−α) · _{Xf, k} (3) Here, when k = 0, N2f _{, 0} = 0. Also, α
Is a coefficient that determines the speed of the noise estimation, and 1> α
It is set within the range of> 0.

By the way, the power spectrum of the background noise fluctuates temporally with a certain variance around its average value. Therefore, if the power spectrum of the background noise is used as it is to remove the background noise, the device will respond excessively to the background noise for each frame, and the operation will be unstable. Therefore, in the present embodiment, the average of the power spectrum _{Xf, k} from the power spectrum calculation unit 4 and the estimated background noise N2f _{, k-1} estimated one frame before is calculated by using the equation (3). By using the averaged power spectrum as the estimated background noise, the apparatus is prevented from responding sensitively to the background noise for each frame, and the apparatus is stabilized.

As described above, the background noise estimating unit 7 estimates the background noise by the equation (3) for each frame during the noise period. Then, the estimated background noise N2 _{f, k-1} held by the background noise holding unit 73 approaches the power spectrum X _{f, k} output from the power spectrum calculation unit 4, and N2 _{f, k-1} ≒
When X _{f, k} , X _{f, k} ≒ N1 _{f, k} (N1
_{Since f, k} is the power spectrum of the background noise), the background noise N1 _{f, k} can be removed by using the estimated background noise N2 _{f, k,} and the background noise estimation algorithm of the background noise estimation unit 7 is as follows. Converge. The background noise updating unit 72 outputs the estimated background noise N2 _{f, k} after the convergence of the estimation algorithm to the adder 5 and the power operation unit 10 connected to the output side, and the background noise holding unit 73 outputs the estimated background noise N2 Hold _{f and k} .

When the noise period ends, and then the operation proceeds to the audio period in which the background noise N1 and the audio signal S are input to the input terminal 18 in a mixed manner, the background noise N1 and the audio signal S are output from the power spectrum calculation unit 4. A power spectrum _{Xf, k} corresponding to the mixed signal is output. The adder 5 calculates the power spectrum X _{f, k} and the estimated background noise N 2 from the background noise estimating unit 7.
_{Using f and k} , subtraction is performed by equation (4), and as a result E1
Power operation unit 10 and adder that connect _{f and k} to the output side
Output to 13. In Equation (4), α _c is a coefficient for adjusting the amount of subtraction of the estimated background noise N2 _{f, k} , and 1 _c
It is determined in advance within the range of ≧ α _c > 0 in consideration of the use environment and the like.

E1 _{f, k} = X _{f, k} −α _c · N2 _{f, k} (4) As described above, during the voice period, the output X (f,
k) is given by X (f, k) = N1 (f, k) + S (f, k). Therefore, assuming that the power spectrum of the sound output from the power spectrum calculation unit 4 is _{Sf, k} , the adder The output E1 _{f, k} of 5 is as shown in equation (5).

E1 _{f, k} ≒ S _{f, k} (5) As described above, the adder 5 outputs the estimated background noise N2 from the power spectrum X _{f, k} corresponding to the audio signal including the background noise.
Power spectrum S corresponding to the audio signal from which _{f and k} have been removed
_{f and k} will be output. Where the estimated background noise N2
_{Since f, k} does not completely match the actual background noise, it is difficult to obtain only the power spectrum S _{f, k} corresponding to the audio signal, and E1 _{f, k} contains some background noise. Becomes In the noise period, the output X of the FFT operator 3
Since (f, k) becomes X (f, k) = N (f, k), the output E1
_{f, k} becomes E1 _{f, k} ≒ 0.

The power operation unit 10 includes a band splitter 101, a power calculation unit 102, a logarithmic calculation unit 103, a switch 104, and a power holding unit 105, and a plurality of power spectra from the adder 5 and the background noise estimation unit 7. Is divided into small frequency bands (channels), and the power is calculated for each channel. Specifically, the power spectrum E1 _{f, k} from the adder 5 and the power spectrum N2 _{f, k} from the background noise estimator 7 are input to both the band divider 101 and the power calculator 102, respectively.

The band divider 101 receives the power spectrum N2 _{f, k} from the background noise estimator 7 during the noise period, and the power spectrum E1 _{f, k} from the adder 5 input for each frame during the voice period. Is divided into a predetermined number of channels, and a power calculation unit 102 connected to the output side
And, it outputs to the background noise re-estimation unit 12 for each band described later.
The power calculator 102 uses the power spectrum of each channel from the band divider 101 to calculate the power sum E1 _{m, k of} each channel corresponding to the power spectrum E1 _{f, k according} to the equations (6) and (7). , The power sum N2 _{m, k} of each channel corresponding to the power spectrum N2 _{f , k} is calculated.

[0027]

(Equation 1) Here, m represents the m-th frequency band (m-th channel) of the divided small region frequency bands, and fs is the m-th frequency band.
Represents the start frequency in the channel, and fe represents the end frequency in the m-th channel.

The power calculator 102 uses the power spectrum E1 _{f, k} from the adder 5 and the power spectrum N2 _{f, k} from the background noise estimator 7 according to equations (8) and (9). , The power sum E1 _ALL in the entire frequency band corresponding to the power spectrum E1 _{f, k} and the power sum N2 _ALL in the entire frequency band corresponding to the power spectrum N2 _{f, k} .

E1 _ALL = ΣE1 _{f, k} (8) N2 _ALL = ΣN2 _{f, k} (9) The power calculator 102 calculates the calculated power sum E1 _{m, k} , N2
_{m, k} and the power sum E1 _ALL and N2 _ALL are output to the logarithmic calculation unit 103 connected to the output side. The logarithmic calculation unit 103 calculates (10) to
According to the equation (13), logarithmic conversion is performed on each power sum and total power from the power calculation unit 102, and the result is output to the switch 104 connected to the output side.

E1 _mlog = log (E1 _{m, k} ) (10) N2 _mlog = log (N2 _{m, k} ) (11) E1 _alog = log (E1 _ALL ) (12) N2 _alog = log (N2 _ALL ) (13) ) The switch 104 is controlled by the time axis sound detector 6,
During the noise period, the switch is tilted to the terminal a side and the logarithmic calculation unit 10
3 is connected to the power retainer 105. As a result, during the noise period, the output of the logarithmic calculation unit 103 is input to the power holding unit 105, and the power holding unit 105
Holds the power sum N2 _alog power sum N2 _mlog and the entire frequency band of each channel from 3, and outputs to the S / N calculator 11 connected to the output side. The switch 104 connects the output side of the logarithmic calculation unit 103 to the S / N calculation unit 11 by turning the switch to the terminal b side during the voice period. Thus, during the voice period, the power sum E1 _mlog of each channel from the logarithmic calculation unit 103
And power sum E1 _alog of all frequency bands are S / N for each frame
It is input to the calculation unit 11.

The S / N calculation unit 11 includes a full-band S / N calculation unit 111 and a band-specific S / N calculation unit 112. The power sum E1 _mlog and N2 _mlog of each channel from the power operation unit 10 The S / N is calculated based on the power sum E1 _alog and N2 _alog of the band. More specifically, the full-band S / N calculator 111 includes a power sum E1 _alog of the entire frequency band transmitted from the logarithmic calculator 103 via the switch 104 during the voice period, and a noise held in the power holder 105. Using the power sum N2 _alog of all frequency bands during the period, S / N _a
And the background noise re-estimator for each band connected to the output side
Output to 12.

On the other hand _{S / N a = E1 alog -N2} alog (14), the band-by-band S / N calculation unit 112, the switch 10 to the speech periods
(15) using the power sum E1 _{mlog of} each channel output from the logarithmic calculation unit 103 through 4 and the power sum N2 _mlog of each channel during the noise period held in the power holder 105. To calculate the S / N _m for each small area, and output the calculated S / N _m to the background noise re-estimation unit 12 for each band connected to the output side.

S / N _m = E1 _mlog −N2 _mlog (15) Here, m is the frequency band of the divided small regions.
Represents the m-th frequency band (m-th channel).

In this embodiment, after the output of the power calculator 102 is logarithmically converted by the logarithmic calculator 103, the output of the S / N calculator 11 is calculated.
Is used to calculate S / N, but S / N may be calculated without logarithmic conversion. In this case, the logarithmic calculation unit 103 becomes unnecessary,
S / N calculation is a little complicated.

The background noise re-estimating unit 12 for each band performs the S / N
A S / N difference for each channel based on the S / N _a of the entire frequency band and the S / N _m of each channel from the S / N calculation unit 11 including a difference determination unit 121 and a background noise re-update unit 122 for each band. , And the estimated background noise is re-updated for the channel having a low S / N. In particular, the per-band S / N difference determination unit 121, the S / N calculation section S / N _a and the band-by-band S / N calculation unit 112 of the entire frequency band from the entire band S / N calculation unit 111 of the 11 Using the S / N _m of each channel of the following formulas (16) and (17),
Determining the difference between the S / N _m and S / N _a for each channel, and outputs the determination result to the band-by-band background noise re-update unit 122 connected to the output side.

S / N _m ≧ S / N _a + δ (16) S / N _m <S / N _a + δ (17) Here, δ is a constant that is determined in advance in consideration of the apparatus and the use environment. is there. FIG. 2 is a diagram illustrating an example of S / N _a of the entire frequency band and S / N _m of the channel. Where α =
0.5, α _c = 0.7, β = 0.8, δ = −6 (dB), and m = 5, and the unit of the vertical axis is dB.

The band-specific background noise re-updating unit 122 calculates the estimated background noise N2 _{mf, of the} channel satisfying the expression (17) among the estimated background noise of each channel output from the band divider 101 of the power operation unit 10 _{. k} and the output E1 _{mf, k of} the channel that satisfies the expression (17) among the outputs obtained by dividing the output of the adder 5 output from the band divider 101 for each channel. The estimated background noise N2 _{mf, k} is re-updated by the equation to generate a re-updated estimated background noise N3 _{mf, k,} which is output to the adder 13 connected to the output side. Note that for channels that satisfy Equation (16), re-update is not performed.

N3 _{mf, k} = (1−β) · N2 _{mf, k} + β · E1 _{mf, k} (18) where m is the m-th frequency band (the m-th channel) of the plurality of divided frequency bands. ). Β is
This is a coefficient representing the speed of re-update, and is used to reflect the output E1 _{mf, k} of the m-th channel of the adder 5 on the estimated re-update background noise N3 _{mf, k} .

In equation (18), when β is increased _, the effect of E1 _{mf, k on} the re-updated estimated background noise N3 _{mf, k} increases, and when β is reduced, the effect decreases. here,
The output E1 _{mf, k} of the m-th channel of the adder 5 changes abruptly in response to a steep change of the background noise. Therefore, if β is large, the re-updated estimated background noise N3 _{mf, k} quickly follows the steeply changing background noise component, and conversely, if β is small, the re-updated estimated background noise N3 _{mf, k} is
This follows the steeply changing background noise component slowly. Therefore, β can be said to be a coefficient representing the re-update speed, and is determined according to the use environment and the like.

The adder 13 includes a background noise re-update unit 122 for each band.
, The output E1 _{mf, k} of the m-th channel of the adder 5 from the band divider 101 and the re-updated estimated background noise from the band-based background noise re-update unit 122 Using N3 _{mf, k} , subtraction is performed according to equation (19), and the result is output to the smoothing operation unit 14 connected to the output side.

E2 _{mf, k} = E1 _{mf, k} −α _k · N3 _{mf, k} = X _{mf, k} −α _c · N2 _{mf, k} −α _k · N3 _{mf, k} (19) where α _k is , 1 ≧ α _k > 0, and is a coefficient for adjusting the amount of subtraction of the renewed estimated background noise N3 _{mf, k} .

The adder 13 connects the output E1 _{mf, k} of the adder 5 to the output side as it is, as shown in the equation (20), for the channel in which the estimated background noise is not updated again. Output to the smoothing operation unit 15.

E2 _{mf, k} = E1 _{mf, k} (20) FIG. 3 is a diagram for explaining the operation of the background noise re-update unit for each band. In FIG. 3, the vertical axis represents the power level of each component, and the horizontal axis represents the channel (frequency). Further, dashed line a, the background noise is estimated by the updater 72 were estimated noise N2 _f, spectrum of _k, solid lines b spectrum X _f where the audio signal and background noise are _{mixed, k,} dotted c the spectrum of background noise, Dotted line d is a background noise re-update unit 122 for each band.
7 _shows the spectrum of the re-updated estimated background noise N3 _{mf, k} re-updated in FIG. Channels with diagonal lines (channels 4, 6, and 9) are channels on which the estimated background noise has been re-updated in the background noise re-update unit 122 for each band.

The estimated background noise N2 _{f, k} indicated by the dashed line a is noise that has been smoothed with respect to the frame as shown by the equation (3), and quickly corresponds to the background noise component that changes sharply for each frame. It does not do. Therefore, even if the estimated background noise N2 _{f, k} indicated by the dashed line a is removed from the spectrum in which the audio signal and the background noise indicated by the dotted line b are mixed in the adder 5, the background noise component that changes sharply remains.

Therefore, for a channel having a low S / N, which corresponds to the expression (17), in other words, a channel having less audio components than the other channels (channels 4, 6, and 9 in FIG. 3), The renewed estimated background noise N3 indicated by the dotted line d corresponding to the background noise component that changes rapidly using the component E1 _{mf, k} corresponding to the channel in the output
_{mf, k} is generated by equation (18). Then, the adder 13 removes the re-updated estimated background noise N3 _{mf, k} from the output E1 _{f, k} of the adder 5 according to equation (19). Thereby, the adder 13
The S / N of the channels having a small audio component at the output of (4, 6, and 9 in FIG. 3) will be greatly improved.

The smoothing operation unit 14 performs a process according to the equations (19) and (20) on the signal from the adder 13 to smooth the frequency characteristics of the signal. Since the discontinuity of each frame is reduced by the smoothing of the frequency characteristics, a natural sound feeling can be obtained. The smoothing operation unit 14 outputs the output E3 _{f, k obtained} by smoothing the frequency characteristic to an inverse fast Fourier transform operation unit (inverse FFT operation unit) 15 connected to the output side.

E3 _{f, k} = (1−γ) · E2 _{f, k−1} + γ · E2 _{f, k} (21) The inverse FFT calculator 15 outputs the phase Φ (f) from the phase holder 9 every frame. , k) by performing an inverse fast Fourier transform on the signal E3 _{f, k} from the smoothing operation unit 14 to convert the signal E3 _{f, k} represented by the frequency component into a time axis signal E4 ( k). Then, it outputs to the window function overlap processing unit 16 connected to the output side. It should be noted that a method of converting a signal represented by this frequency component into a time axis signal by an inverse FFT operation is already known.

The window function overlap processing unit 16 matches the time axis signal E4 (k) from the inverse FFT calculator 15 by overlapping at a predetermined ratio so as not to be discontinuous between frames. To a digital-analog converter (D / A converter) 17 connected to the output side.
The D / A converter 17 converts the signal from the window function overlap processing unit 16 from a digital signal to an analog signal, and outputs it to an output terminal 19 connected to the output side. In this way,
The audio signal input to the input terminal 18 is output from the output terminal 17 after removing background noise.

As described above, according to the first embodiment, the input signal is divided into a plurality of small frequency bands (channels) and the background noise is subtracted.
Even if the characteristics of the input signal are unknown, there is no need to change the processing contents, and the background noise is removed by a single processing method without discrimination between the low frequency range and the high frequency range. It is possible to realize a background noise elimination device with excellent sound quality that does not cause discomfort.

Further, according to the first embodiment, the input signal is divided into a plurality of channels, and a noise subtraction process is performed for each channel using estimated background noise indicating an average characteristic of the background noise. Since the S / N is smaller than that of the channel, that is, the channel with less voice component, the noise is further subtracted using the re-updated estimated background noise corresponding to the steep component of the background noise. Even when the audio components are concentrated in the low frequency range, the background noise can be accurately removed from the channel having a small audio component.

FIG. 4 is a block diagram showing a background noise removing apparatus according to a second embodiment of the present invention. This second embodiment is:
The time axis speech detector 6 in the background noise elimination device of the first embodiment shown in FIG. 1 is changed to a frequency axis speech detector 20, and the output side of the power spectrum calculator 4 and the band divider 101 of the power operation unit 10 are changed. Are connected to the input side of the frequency axis sound detector 20, respectively.

In FIG. 4, the frequency axis sound detector 20 is
The output _{Xf, k} of the power spectrum calculator 4 and the band splitter 101
Using the output _{N2mf, k} of (1), the output N4c _{, k} is calculated by the equation (22) for each frequency band (channel) of the small region divided by the band divider 101.

_{N4cf, k} = ε · _{N2cf, k} + (1−ε) · _{Xcf, k} (22) where _{N4cf, 0} = 0. Here, c represents a c-th frequency band (c-th channel) of a plurality of divided frequency bands, and ε determines a predetermined noise averaging speed within a range of 1>ε> 0. Is a constant.

Next, the frequency axis sound detector 20 adds a preset threshold value κ to the output _{N4cf, k} calculated by the equation (22) according to the equation (23). And the addition result J
_{If cf, k} satisfies Equation (24), it is determined that the audio component is included in the c-th channel, and the addition result J _{cf, k}
Satisfies Expression (25), it is determined that the c-th channel does not include a sound component.

J _{cf, k} = κ + N4 _{cf, k} (23) X _{cf, k} > J _{cf, k} (24) X _{cf, k} ≦ J _{cf, k} (25) Next, the frequency axis sound detector 20 If the number V of channels determined to contain the component satisfies Equation (26) with respect to the predetermined number of channels M, the k-th
The k-th frame is determined to be a frame that does not include an audio signal, and if the expression (27) is satisfied, the k-th frame is determined to be a frame that does not include an audio signal. Note that the number of channels M is not limited to a specific value.

Number of Channels with Audio Component V ≧ M (26) Number of Channels with Audio Component V <M (27) FIG. 5 is a diagram for explaining the operation of the above-described frequency axis audio detector 20. In FIG. 5, the vertical axis is the power level,
The horizontal axis is the channel number (frequency). Also, the solid line a
Is the power spectrum output X _{f, k of} the k-th frame output from the power spectrum calculator 4, the solid line b is the spectrum N 2 _{f, k} of the estimated background noise estimated by the background noise estimator 7, and the dotted line c represents the addition result J _{f, k} calculated by equation (23).

FIG. 5 shows that, for the second, fourth, and sixth channels, equation (24) holds, and it is determined that a speech component is included, and the remaining first, third, fifth, seventh through seventh channels are determined. This is an example of a case where Expression (25) is established for the tenth channel and it is determined that no audio component is included. In this case, since the number of channels V determined to include an audio component is 3 channels, for example, if the number of channels M in equations (26) and (27) is 3, the K-th frame Is determined to include an audio signal. In the present embodiment, whether or not the audio signal is included in the frame is determined based on the number of channels V including the audio component. However, the audio is included in the frame based on the ratio of the number of channels V including the audio component to the total number of channels. It may be determined whether or not a signal is included.

The judgment result by the frequency axis sound detector 20 is as follows.
As in the case of the first embodiment, the signals are sent to the switch 71 of the background noise estimating unit 7 and the switch 104 of the power operating unit 10, and
Control each switch. The components other than the frequency axis sound detector 20 have the same configuration and operation as the corresponding components of the first embodiment shown in FIG.

As described above, according to the second embodiment, since the detection of the audio signal is performed on the frequency axis, it is possible to set the threshold value which can be detected with emphasis on the frequency component of the audio. In addition, since the audio signal is detected on a frame basis, data in many frames is collectively used as determination data, and the audio signal can be detected reliably.
Therefore, it is possible to provide a background noise elimination apparatus which can remove background noise with high accuracy and has excellent sound quality. According to the second embodiment, the same effects as those of the first embodiment can be obtained.

FIG. 6 is a block diagram showing a background noise removing apparatus according to a third embodiment of the present invention. In the background noise elimination device of the first embodiment shown in FIG.
The estimated background noise N2f _{, k} generated by the background noise estimator 7 is subtracted by the adder 5 from the power spectrum _{Xf, k} output from the power spectrum calculator 4, and then the background noise re-estimator 12 for each band is subtracted. The background noise is removed by subtracting the re-updated estimated background noise N3 _{mf, k} generated in step (3) by the adder 13.

On the other hand, in the third embodiment, the power spectrum X _{f, k} output from the power spectrum calculator 4
From the estimated background noise N2 _{f, k} and the re-updated estimated background noise N3 _{mf, k}
Are collectively subtracted by the adder 13. Therefore, in FIG. 6, the output side of the power spectrum calculator 4 is directly connected to the input side of the adder 13 without passing through the adder 5. Then, the band-based background noise re-estimating unit 12 _{adds a} certain amount of estimated background noise N2 _{f, k to the} re-updated estimated background noise N3 _{mf , k.}
To generate an estimated background noise (N3 _{mf, k} + α _c · N2 _{f, k} ), which is output to the adder 13 connected to the output side. In the third embodiment, the same effect as in the first embodiment can be obtained.

FIG. 7 is a block diagram showing a background noise removing apparatus according to a fourth embodiment of the present invention. In the background noise elimination device of the second embodiment shown in FIG.
The estimated background noise N2f _{, k} generated by the background noise estimator 7 is subtracted by the adder 5 from the power spectrum _{Xf, k} output from the power spectrum calculator 4, and then the background noise re-estimator 12 for each band is subtracted. The background noise is removed by subtracting the re-updated estimated background noise N3 _{mf, k} generated in step (3) by the adder 13.

On the other hand, in the fourth embodiment, the power spectrum X _{f, k} output from the power spectrum calculator 4
From the estimated background noise N2 _{f, k} and the re-updated estimated background noise N3 _{mf, k}
Are collectively subtracted by the adder 13. Therefore, in FIG. 7, the output side of the power spectrum calculation unit 4 is directly connected to the input side of the adder 13 without passing through the adder 5. Then, the band-based background noise re-estimating unit 12 _{adds a} certain amount of estimated background noise N2 _{f, k to the} re-updated estimated background noise N3 _{mf , k.}
To generate an estimated background noise (N3 _{mf, k} + α _c · N2 _{f, k} ), which is output to the adder 13 connected to the output side. In the fourth embodiment, the same effects as in the second embodiment can be obtained.

It should be noted that the present invention can be widely applied not only to portable telephone devices, car telephone devices, and audio devices, but also to voice recognition devices, video conference devices, wireless communication devices, and the like.

[0065]

As described above, according to the present invention, since the frequency band of the input signal is divided into a plurality of small areas and the background noise is subtracted for each small area, the characteristics of the input signal are unknown. Even if there is no need to change the processing content,
Since the background noise removal processing is performed by one processing method without distinction between the low frequency range and the high frequency range, an effect that a sense of incongruity does not occur in the auditory sense can be obtained.

Further, according to the present invention, a noise subtraction process is performed for each small area using an estimated background noise having an average characteristic of the background noise, and further, the S / N is reduced as compared with other channels. For small channels, noise subtraction processing is performed using renewed estimated background noise corresponding to a sharp change in background noise, so background noise in channels with small S / N can be accurately removed. Thus, it is possible to realize a background noise elimination device having excellent sound quality.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is a diagram illustrating an example of S / N of a channel and S / N of an entire frequency band.

FIG. 3 is a diagram illustrating an operation example of a background noise re-update unit for each band.

FIG. 4 is a block diagram showing a second embodiment of the present invention.

FIG. 5 is a diagram illustrating an operation example of a frequency axis sound detector.

FIG. 6 is a block diagram showing a third embodiment of the present invention.

FIG. 7 is a block diagram showing a fourth embodiment of the present invention.

[Explanation of symbols]

 Reference Signs List 1 A / D converter 2 Window function calculator 3 FFT calculator 4 Power spectrum calculator 5, 13 adder 6 Time axis voice detector 7 Background noise estimator 8 Phase calculator 9 Phase holder 10 Power operation unit 11 S / N calculation unit 12 Band-wise background noise re-estimation unit 14 Smoothing operation unit 15 Inverse FFT operation unit 16 Window function overlap processing unit 17 D / A converter 20 Frequency axis speech detector

Claims (5)

[Claims] 1. A background noise elimination device in which background noise is input as an input signal and then the background noise and the audio signal are input in a mixed form. A signal conversion unit that converts a signal into a frequency component, a voice detection unit that detects a voice signal included in the input signal, and a noise period in which a voice signal is not detected by the voice detection unit. Background noise estimating means for generating and holding the estimated background noise by averaging the component and the estimated background noise generated one frame before; and a voice period in which a voice signal is detected by the voice detection means,
A first method for subtracting the estimated background noise held by the background noise estimating means from the frequency component converted by the signal converting means;
An S / N of the entire frequency band and a plurality of S / Ns of the entire frequency band, with the frequency component obtained by the subtraction of the first adding unit as a signal and the estimated background noise held by the background noise estimating unit as noise. S / N calculating means for calculating the S / N for each of the divided small regions, and for each small region in which the difference between the S / N of the small region and the S / N of the entire frequency band is equal to or smaller than a predetermined value, Band-based background noise re-estimating means for generating a re-estimated estimated background noise including a frequency component obtained by the subtraction of the adding means of No. 1 and the estimated background noise held by the background noise estimating means at a predetermined ratio; A second addition means for subtracting the re-updated estimated background noise from a frequency component obtained by subtraction of the first addition means, and a frequency component for each frame obtained by the subtraction of the second addition means is converted into a time axis signal. Signal reproducing means for outputting the first addition Means for subtracting the estimated background noise and the second
The magnitude of the renewed estimated background noise to be subtracted by the adding means of
A background noise elimination device, which is set so that background noise is eliminated from the input signal. 2. A background noise elimination apparatus in which background noise is input as an input signal and then the background noise and the audio signal are input in a mixed form, the apparatus comprising: A signal conversion unit that converts a signal into a frequency component, a voice detection unit that detects a voice signal included in the input signal, and a noise period in which a voice signal is not detected by the voice detection unit. Background noise estimating means for generating and holding the estimated background noise by averaging the component and the estimated background noise generated one frame before; and a voice period in which a voice signal is detected by the voice detection means,
A first method for subtracting the estimated background noise held by the background noise estimating means from the frequency component converted by the signal converting means;
An S / N of the entire frequency band and a plurality of S / Ns of the entire frequency band, with the frequency component obtained by the subtraction of the first adding unit as a signal and the estimated background noise held by the background noise estimating unit as noise. S / N calculating means for calculating the S / N for each of the divided small regions, and for each small region in which the difference between the S / N of the small region and the S / N of the entire frequency band is equal to or smaller than a predetermined value, 1 generates noise including a predetermined ratio of a frequency component obtained by subtraction of the adding means and the estimated background noise held by the background noise estimating means, and the estimated background held by the background noise estimating means is included in the noise. Band-based background noise re-estimating means for generating re-updated estimated background noise by adding noise, and second adding means for subtracting the re-updated estimated background noise from frequency components obtained by conversion of the signal converting means. The frame obtained by the subtraction of the second adding means. Signal reproduction means for converting a frequency component for each system into a time axis signal and outputting the converted signal. The magnitude of the renewed estimated background noise to be subtracted by the second addition means is such that background noise is removed from the input signal. A background noise elimination device, wherein 3. The background noise elimination apparatus according to claim 1, wherein the voice detection unit detects the voice component included in a frequency component obtained by the conversion by the signal conversion unit, thereby detecting the input. A background noise elimination device for detecting an audio signal included in a signal. 4. The background noise elimination device according to claim 3, wherein said voice detection means detects a voice component for each of a plurality of small regions obtained by dividing a frequency component obtained by the conversion by said signal conversion means, and A background noise eliminator characterized in that it is determined that an audio signal is included in the input signal when the number of small regions in which components are detected is equal to or greater than a predetermined value. 5. The background noise elimination apparatus according to claim 3, wherein said audio detecting means detects an audio component for each of a plurality of small regions obtained by dividing a frequency component obtained by the conversion of said signal conversion means, and A background noise eliminator characterized in that it is determined that an audio signal is included in the input signal when a ratio of the number of small areas in which audio components are detected to the number of small areas is equal to or greater than a predetermined value.