JP2003131689A - Noise removing method and device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and device for removing noise capable of obtaining dynamic voices excellent in subjective sound quality without generating noise in output signals. SOLUTION: This device has a curtain processing section 22 for subjecting the signal samples taken out of adjacent two frames of reverse Fourie transform outputs to curtain processing.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise removing method and apparatus, and more particularly to a noise removing method and apparatus for removing noise superimposed on an audio signal.

[0002]

2. Description of the Related Art Conventionally, there is a noise suppressor as a technique for eliminating noise superimposed on a voice signal. The noise suppressor estimates the power spectrum of the noise component by using the input signal transformed into the frequency domain by cos transform or Fourier transform, and subtracts the estimated power spectrum from the input signal to eliminate noise mixed in the voice signal. Operates to suppress.

The power spectrum of the noise component can be applied to suppression of non-stationary noise by detecting and updating the silent section of the voice.

Regarding the noise suppressor, for example,
"December 1984, IEE Transactions on Auctions Speech and
Signal Processing, Volume 32, Issue 6 (IEEE TRANSAC
TIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VO
L.32, NO.6 PP.1109-1121, DEC, 1984), pages 1109 to 1121 "(hereinafter referred to as" Reference 1 ").

The method described in Document 1 is known as the minimum mean square error short-time spectrum amplitude method.

FIG. 25 is a block diagram showing a schematic configuration of the noise suppressor described in Document 1 and the like.

A degraded audio signal, which is a signal in which an audio signal and noise are mixed, is supplied to the input terminal 11 as a sample value series. The deteriorated audio signal is supplied to the frame division unit 1 and divided into K / 2 sample frames. Here, K is an even number.

The deteriorated audio signal divided into frames is supplied to the windowing processing section 2 and is multiplied by the window function w (t). Input signal y _n (t) (t = 0,
The signal y _n (t) bar windowed with w (t) for 1, ..., K / 2−1) is given by Eq.

[0009]

[Equation 1] It is also widely practiced to overlap a part of two consecutive frames and to open a window. Assuming that the overlap length is 50% of the frame length, y _n (t) bar (t = 0, t = 0,
, ..., K-1) are output from the windowing processing unit 2.

[0010]

[Equation 2] A symmetric window function is used for real signals. Further, the window function is designed so that the input signal and the output signal when the suppression coefficient is set to 1 match each other except for a calculation error. This means that w (t) + w (t + K / 2) = 1.

Hereinafter, the description will be continued by taking as an example the case where a window is formed by overlapping 50% of two consecutive frames.

As w (t), for example, the Hanning window shown in Expression 3 can be used.

[0013]

[Equation 3] The windowed output y _n (t) bar is the Fourier transform unit 3
And converted into a degraded speech spectrum Y _n (k). The deteriorated speech spectrum Y _n (k) is separated into a phase and an amplitude, and the deteriorated speech phase spectrum argY _n (k) is supplied to the inverse Fourier transform unit 9 and the deteriorated speech amplitude spectrum | Y.
_n (k) | is supplied to the voice detection unit 4, the multiple multiplication unit 16, and the multiple multiplication unit 17, respectively.

The voice detection unit 4 detects the presence or absence of voice based on the deteriorated voice amplitude spectrum | Y _n (k) |, and transmits the voice detection flag determined by the result to the estimated noise calculation unit 51.

The multiplying unit 17 calculates a deteriorated voice power spectrum using the supplied deteriorated voice amplitude spectrum | Y _n (k) |, and estimates the noise calculation unit 51 and the frequency-based signal-to-noise ratio (hereinafter, " It is abbreviated as “SNR”).

The estimated noise calculation section 51 uses a voice detection flag,
The noise power spectrum is estimated using the degraded voice power spectrum and the count value supplied from the counter 13, and the SNR for each frequency is used as the estimated noise power spectrum.
It is transmitted to the calculation unit 6.

The frequency-based SNR calculation unit 6 calculates the SNR for each frequency using the input deteriorated speech power spectrum and estimated noise power spectrum, and estimates the a priori SNR calculation unit 7 and noise suppression coefficient as an acquired SNR. Supply to the part 8.

The estimated a priori SNR calculator 7 estimates the a priori SNR using the input a posteriori SNR and the suppression coefficient supplied from the noise suppression coefficient generator 8 to estimate the a priori SNR.
It is fed back to the noise suppression coefficient generation unit 8 as NR.

The noise suppression coefficient generation unit 8 generates a noise suppression coefficient using the acquired SNR and the estimated a priori SNR supplied as inputs, and returns the noise suppression coefficient to the estimated a priori SNR calculation unit 7 as a suppression coefficient and simultaneously multiplexes it. The result is transmitted to the multiplication unit 16.

The multiple multiplication section 16 uses the deteriorated speech amplitude spectrum | Y _n (k) |
Suppression coefficient G supplied from the noise suppression coefficient generation unit 8
_The weighted _n (k) bar is used to obtain the emphasized speech amplitude spectrum | X _n (k) | bar, which is transmitted to the inverse Fourier transform unit 9. The | X _n (k) | bar is given by Equation 4.

[0021]

[Equation 4] The inverse Fourier transform unit 9 receives the emphasized voice amplitude spectrum | X _n (k) | bar supplied from the multiplex multiplication unit 16 and the deteriorated voice phase spectrum arg supplied from the Fourier transform unit 3.
| Y _n (k) | is multiplied to obtain the emphasized speech X _n (k) bar. That is, the calculation shown in Formula 5 is performed.

[0022]

[Equation 5] An inverse Fourier transform is applied to the obtained emphasized speech X _n (k) bar, and a time domain sample value series X _n (t) bar (t = 0, 1, ..., K−1) in which one frame is composed of K samples. )
Is transmitted to the frame synthesis unit 10.

The frame synthesizing unit 10 extracts K / 2 samples from two adjacent frames of the X _n (t) bar and superimposes them, and obtains the emphasized speech X _n (t) hat by the formula (6). Obtained emphasized speech X _n (t) hat (t =
0, 1, ..., K−1) is transmitted to the output terminal 12 as the output of the frame synthesizing unit 10.

[0024]

[Equation 6] Document 1 does not describe the implementation method of the voice detection unit in detail. However, as an implementation example of the voice detection unit,
Since "March 2000, Acoustical Society of Japan, Proceedings, pages 321-322" (hereinafter referred to as "reference 2") is known, the method shown in reference 2 will be used as the conventional method. explain.

FIG. 26 is a block diagram showing the structure of the voice detection unit 4 of FIG. The voice detection unit 4 includes a threshold storage unit 4
01, comparator 402, multiplier 404, logarithm calculator 40
5, a power calculation unit 406, a weighted addition unit 407, a weight storage unit 408, and a logical NOT circuit 409.

The deteriorated voice amplitude spectrum supplied from the Fourier transform unit 3 in FIG. 25 is supplied to the power calculation unit 406. The power calculation unit 406 calculates the sum of the power | Y _n (k) | ² of the deteriorated speech amplitude spectrum for k = 0 to K−1, and transmits the sum to the logarithmic calculation unit 405.

The logarithmic calculation unit 405 obtains the logarithm of the input deteriorated speech spectrum power and transmits it to the multiplier 404.

The multiplier 404 multiplies the supplied logarithmic value by a constant to obtain the deteriorated voice power Q _n (901), and the comparison unit 4
02 and the weighted addition unit 407. That is,
The deteriorated voice power Q _n of the nth frame is given by Expression 7.

[0029]

[Equation 7] Incidentally, speech detection unit described in the Reference 2, with y _n (t) bar is time-domain samples, Q _n according to equation 8
Are seeking.

[0030]

[Equation 8] However, as shown in, for example, "Theory of Digital Signal Processing, 1985, Corona Publishing Co., Ltd., pp. 75-76" (hereinafter referred to as "reference 3"), the equation 8 and the equation 7 are not equivalent to each other. , Known as Parseval's equation.

The threshold value TH _n is supplied from the threshold value storage unit 401 to the comparison unit 402. The comparison unit 402 compares the output of the multiplier 404 with the threshold value TH _n, and TH _n > Q _n
When it is, "1" indicating a voice is output, and when TH _n ≤Q _n , "0" indicating a silence is output as a voice detection flag.

The output of the comparison section 402 is supplied to the outside as a voice detection flag which is the output of the voice detection section 4, and at the same time, is supplied to the negative operation circuit 409. Negative operation circuit 4
The output of 09 is supplied to the weighted addition unit 407 as the weighted addition unit control signal 905.

The weighted addition unit 407 also has a threshold value from the threshold value storage unit 401 and a weight value 90 from the weight value storage unit 408.
3 and 3 are supplied. The weighted addition unit 407 selectively updates the threshold value 902 supplied from the threshold value storage unit 401 based on the weighted addition unit control signal 905, and the updated threshold value 9
It returns to 04 as the threshold value storage part 401.

The update threshold TH _n is obtained by weighted addition of the threshold TH _n-1 and the deteriorated voice power 901 using the weight 903 supplied from the weight storage unit 408.
The update threshold TH _n is calculated only when the weighted adder control signal 905 output from the logical NOT circuit 409 is equal to “1”. That is, only when there is no sound, the threshold TH
_n is updated. Update threshold 904 obtained by update
Are returned to the threshold storage unit 401.

FIG. 27 is a block diagram showing the structure of the power calculation unit 406 shown in FIG. The power calculator 406
It includes a separating unit 4061, multipliers 4062 _{0 to} 4062 _K−1 , and an adder 4063.

The deteriorated speech amplitude spectrum supplied from the Fourier transform unit 3 in FIG. 25 in the multiplexed state is separated into K samples for each frequency in the separation unit 4061, and is divided into multipliers 4062 _{0 to} 4062 _K-1 . Supplied.

Multipliers 4062 _{0 to} 4062 _K-1 square the respective input signals and transmit them to the adder 4063. The adder 4063 calculates and outputs the total sum of the input signals.

FIG. 28 shows the weighted addition unit 407 of FIG.
3 is a block diagram showing the configuration of FIG. Weighted addition unit 407
Is a multiplier 4071, 4073, a constant multiplier 4075,
It has adders 4072 and 4074.

In the weighted addition unit 407, the deteriorated voice power 901 from the multiplier 404 in FIG. 26, the threshold value storage unit 401 to the threshold value 902 in FIG. 26, and the weight storage unit 40 in FIG.
8 to the weight 903 and the weighted addition unit control signal 905 from the logical NOT circuit 409 of FIG.

The weight 903 having the value β is the constant multiplier 4
075 and the multiplier 4073. Constant multiplier 4
075 transmits -β obtained by multiplying the input signal by -1 to the adder 4074. One is supplied to the other input of the adder 4074, and the output of the adder 4074 is the sum of the two, 1-β. 1-β is a multiplier 4071
To the other input, the degraded voice power Q
_It is multiplied by _n and the product (1-β) Q _n is transmitted to the adder 4072.

On the other hand, in the multiplier 4073, β supplied as the weight 903 is multiplied by the threshold value 902, and the product βTH _n-1 is transmitted to the adder 4072. Adder 40
72 indicates the sum of βTH _n-1 and (1-β) Q _n as the update threshold value 9
Output as 04.

Update threshold TH_nIs the weighted addition part
It is performed only when the control signal 905 is equal to "1". You
That is, the function of the weighted addition unit is
TH _n-1To update TH_nIs calculated as
Therefore, it can be expressed.

[0043]

[Equation 9] Here, β is the value of the weight 903.

FIG. 29 is a block diagram showing the structure of the multiplexing and multiplying unit 17 of FIG. The multiplexing multiplication unit 17 includes multipliers 1701 _{0 to} 1701 _K−1 and separation units 1702 and 170.
3 and a multiplexing unit 1704. The deteriorated speech amplitude spectrum supplied from the Fourier transform unit 3 in FIG. 25 in the multiplexed state is separated into K samples for each frequency in the separation units 1702 and 1703, and each is multiplied by a multiplier 1701 _0.
~ 1701 _K-1 .

Multipliers 1701 _{0 to} 1701 _K-1 square the respective input signals and transmit them to multiplexing section 1704. The multiplexing unit 1704 multiplexes the input signals,
Output as degraded voice power spectrum.

FIG. 30 is a block diagram showing the structure of the estimated noise calculation section 51 of FIG. The estimated noise calculation unit 51
It includes a demultiplexing unit 502, a multiplexing unit 503, and frequency _- dependent estimated noise calculation units 514 _{0 to} 514 _K-1 .

The voice detection flag supplied from the voice detection unit 4 in FIG. 25 and the count value supplied from the counter 13 in FIG. 25 are the estimated noise calculation units by frequency 514 _{0 to} 514 _K-1.
Be transmitted to.

The deteriorated voice power spectrum supplied from the multiplex multiplier 17 of FIG. 25 is transmitted to the separator 502. The demultiplexing unit 502 demultiplexes the deteriorated voice power spectrum supplied in the multiplexed state into components corresponding to K frequencies, and the frequency-dependent estimated noise calculation units 514 _{0 to} 514.
Transmit to _K-1 .

Frequency-dependent estimated noise calculators 514 _{0 to} 514
_K-1 calculates a noise power spectrum using the deteriorated speech power spectrum supplied from demultiplexing section 502, and transfers it to multiplexing section 503. The calculation of the noise power spectrum is controlled by the count value and the value of the voice detection flag, and is executed only when a predetermined condition is satisfied.

The multiplexing unit 503 multiplexes the supplied K noise power spectrum values and outputs them as an estimated noise power spectrum.

FIG. 31 is a block diagram showing the structure of the frequency-dependent estimated noise calculation unit 514 of FIG. The noise estimation described in Reference 2 updates the noise estimation value in a silent section, and uses the instantaneous value of the estimated noise averaged by the recursive filter as the noise estimation value.

On the other hand, "May 1998, I-E-E-
E Transactions on Speech and
Audio Processing, Volume 6, Issue 3 (IEEE TRANS-
ACTIONS ON SPEECH AND AUDIO PROCESSING, VOL.6, NO.3,
PP.287-292, MAY, 1998), pp. 287-292 "(hereinafter referred to as" Reference 4 "), it is described that the instantaneous values of the estimated noise are averaged and used. .

This suggests the realization of averaging using a transversal type filter (configuration using shift registers) instead of the cyclic type. Since the functions are the same in both implementations, the method described in Reference 4 will be described here.

The frequency-dependent estimated noise calculation unit 514 includes an update determination unit 521, a register length storage unit 5041, and a switch 50.
44, shift register 5045, adder 5046, minimum value selection unit 5047, division unit 5048, counter 5049
Have.

The switch 5044 has a separating section 5 shown in FIG.
02, the deteriorated voice power spectrum for each frequency is supplied. When the switch 5044 is closed, the frequency-dependent degraded voice power spectrum is stored in the shift register 5045.
Be transmitted to.

As will be described later, the shift register 5045 shifts the storage value of the internal register to the adjacent register according to the control signal supplied from the update determining unit 521. The shift register length is the register length storage unit 5 described later.
Equal to the value stored in 941. Shift register 5
All 045 register outputs are provided to adder 5046.

The adder 5046 adds all the supplied register outputs and sends the addition result to the division unit 5048.

On the other hand, the update determination section 521 is supplied with the count value and the voice detection flag. Update determination unit 521
Is always "1" until the count value reaches a preset value, "1" when the voice detection flag is "0" (silence) after the count value, and otherwise. “0” is output and transmitted to the counter 5049, the switch 5044, and the shift register 5045.

The switch 5044 closes the circuit when the signal supplied from the update judging section 521 is "1", and is "0".
Open at.

The counter 5049 increments the count value when the signal supplied from the update judging section 521 is "1", and does not change it when it is "0".

The shift register 5045 takes in one sample of the signal sample supplied from the switch 5044 when the signal supplied from the update determination section is "1", and at the same time shifts the stored value of the internal register to the adjacent register.

The minimum value selector 5047 has a counter 50.
The output of 49 and the output of the register length storage unit 5941 are supplied. The minimum value selection unit 5047 selects the smaller one of the supplied count value and the register length and transmits it to the division unit 5048.

The dividing unit 5048 divides the added value of the frequency-dependent deteriorated speech power spectrum supplied from the adder 5046 by the smaller value of the count value and the register length, and the quotient is the frequency-dependent estimated noise power spectrum λ _n. Output as (k). Letting B _n (k) (n = 0, 1, ..., N−1) be a sample value of the deteriorated speech power spectrum stored in the shift register 5045, λ _n (k) is represented by Formula 1
It can be indicated by 0.

[0064]

[Equation 10] However, N is the smaller value of the count value and the register length. Since the count value starts from zero and increases monotonically, division is first performed by the count value and later division is performed by the register length. The division by the register length means obtaining the average value of the values stored in the shift register.

Initially, since not enough values are stored in the shift register 5045, division is performed by the number of registers in which values are actually stored. The number of registers in which values are actually stored is equal to the count value when the count value is smaller than the register length, and is equal to the register length when the count value is larger than the register length.

FIG. 32 is a block diagram showing the structure of the update determination unit 521 shown in FIG. The update determination unit 521 includes a logical NOT circuit 5202, a comparison unit 5203, and a threshold value storage unit 520.
4. It has a logical sum calculation unit 5211.

The count value supplied from the counter 13 in FIG. 25 is transmitted to the comparison section 5203. Threshold storage unit 5
The threshold value that is the output of 204 is also transmitted to the comparison unit 5203. The comparing unit 5203 compares the supplied count value with the threshold value, and when the count value is smaller than the threshold value, “1”
Is transmitted to the logical sum calculation unit 5211 when the count value is larger than the threshold value.

On the other hand, the supplied voice detection flag is transmitted to the logical NOT circuit 5202. Logical NOT circuit 5202
Calculates the logical negation value of the input signal and transmits it to the logical sum calculation unit 5211. That is, the logical sum calculation unit 5211 outputs “0” in the voiced part whose voice detection flag is “1” and “1” in the silent part whose voice detection flag is “0”.
Will be transmitted to.

As a result, the output of the logical sum calculation unit 5211 becomes "1" when the voice detection flag is the silent part where the voice detection flag is "0" or when the count value is smaller than the threshold value, and the switch of FIG. Close and count up the counter 5049.

FIG. 33 is a block diagram showing the configuration of the frequency-specific SNR calculation unit 6 of FIG. The frequency-based SNR calculation unit 6 includes division units 601 _{0 to} 601 _K−1 and separation units 602 and 6.
03, and a multiplexing unit 604.

The deteriorated voice power spectrum supplied from the multiplex multiplier 17 of FIG. 25 is transmitted to the separator 602. The estimated noise power spectrum supplied from the estimated noise calculation unit 51 in FIG. 25 is transmitted to the separation unit 603. In noisy speech power spectrum separation unit 602, the estimated noise power spectrum in the separation unit 603 is separated into K samples corresponding to the frequency components, respectively, supplied to each divider 601 ₀ ~601 _K-1.

In the division units 601 _{0 to} 601 _K-1 ,
According to the above, the supplied degraded speech power spectrum is divided by the estimated noise power spectrum to obtain SNR γ for each frequency.
_n (k) is obtained and transmitted to the multiplexing unit 604.

[0073]

[Equation 11] Here, λ _n (k) is the estimated noise power spectrum. The multiplexing unit 604 receives the transmitted K SNs for each frequency.
R is multiplexed and output as an acquired SNR.

FIG. 34 is a block diagram showing the configuration of the estimated a priori SNR calculation unit 7 of FIG. Estimated innate SNR
The calculation unit 7 includes a multiple range limitation processing unit 701 and an acquired SNR.
Storage unit 702, suppression coefficient storage unit 703, multiplex multiplication unit 70
4, 705, weight storage unit 706, multiple weighted addition unit 7
07 and an adder 708.

The acquired SNR γ _n (k) (k = 0, 1, ..., K−) supplied from the frequency-based SNR calculation unit 6 of FIG.
1) is transmitted to the acquired SNR storage unit 702 and the adder 708.

The acquired SNR storage unit 702 stores the acquired SNR γ _n (k) in the nth frame, and
The acquired SNR γ _n-1 (k) in the ( _n-1 ) _th frame is transmitted to the multiplex multiplication unit 705.

The suppression coefficient G _n (k) bar (k = 0, 1, ..., K-1) supplied from the noise suppression coefficient generator 8 of FIG.
Is transmitted to the suppression coefficient storage unit 703. The suppression coefficient storage unit 703 stores the suppression coefficient G _n (k) in the nth frame.
The bar is stored, and the suppression coefficient G _n−1 (k) bar in the ( _n−1 ) _th frame is transmitted to the multiplex multiplication unit 704.

The multiplying unit 704 receives the supplied G
_{The n-1} (k) bar is squared to obtain G ² _n-1 (k) bar, which is transmitted to the multiplex multiplication unit 705.

The multiplying unit 705 multiplies G ² _n-1 (k) bar and γ _n-1 (k) by k = 0, 1, ..., K-1 to obtain G ² _n-1. (K) bar γ _n-1 (k) is obtained, and the result is transmitted to the multiple weighted addition unit 707 as the past estimated SNR 922.

The configurations of the multiple multiplication units 704 and 705 are the same as those of the multiple multiplication unit 17 described with reference to FIG.

-1 is supplied to the other terminal of the adder 708, and the addition result γ _n (k) -1 is transmitted to the multiple range limiting processing unit 701.

The multi-value range limiting processing section 701 includes an adder 70.
The addition result γ _n (k) -1 supplied from 8 is subjected to calculation by the range limiting operator P [·], and the result P [γ
_[n (k) -1] is transmitted to the multiple weighted addition unit 707 as the instantaneous estimated SNR 921. However, P [x] is determined by Expression 12.

[0083]

[Equation 12] The multiple weighted addition unit 707 also includes a weight storage unit 70.
The weight 923 is supplied from 6. The multi-weighted adder 707 receives these supplied instantaneous estimated SNRs 92
1, the estimated a priori SNR 924 is obtained using the past estimated SNR 922 and weight 923. The weight 923 is α, and ξ
_{If n} (k) hat is the estimated innate SNR, then ξ _n (k)
The hat is calculated using Equation 13.

[0084]

[Equation 13] Here, it is assumed that G ² ₋₁ (k) γ ₋₁ (k) bar = 1.

FIG. 35 is a block diagram of the multiple range limitation processing unit 7 of FIG.
It is a block diagram which shows the structure of 01. The multiple value range limitation processing unit 701 includes a constant storage unit 7011 and a maximum value selection unit 701.
2 ₀ ~7012 _K-1, the separation section 7013, multiplexing section 7014
Have.

The separating section 7013 includes an adder 70 shown in FIG.
8 supplies γ _n (k) -1. Separation unit 7013
Separates the supplied γ _n (k) -1 into K frequency components and supplies them to the maximum value selection units 7012 _{0 to} 7012 _K-1 .

Zeros are supplied from the constant storage unit 7011 to the other inputs of the maximum value selection units 7012 _{0 to} 7012 _K-1 . Maximum value selector 7012 ₀ ~7012 _K-1 are, gamma _n
(K) -1 is compared with zero, and the larger value is multiplexed by the multiplexing unit 7
It is transmitted to 014. This maximum value selection operation is equivalent to executing Expression 12. The multiplexing unit 7014 multiplexes these values and outputs them.

FIG. 36 shows the multiple weighted addition unit 7 of FIG.
It is a block diagram which shows the structure of 07. The multiple weighted addition unit 7071 is configured to add weighted addition units 7071 _{0 to} 707.
1 _K-1 , demultiplexing units 7072 and 7074, multiplexing unit 7075
Have.

In the demultiplexing unit 7072, P [γ _n (k) -1] is instantaneously estimated S from the multiple value range limiting processing unit 701 of FIG.
Supplied as NR921. The separation unit 7072 is P
The [γ _n (k) -1] is separated into K frequency-components, as frequency-instantaneous estimation SNR921 ₀ ~921 _K-1, and transmits the weighted adder 7071 ₀ ~7071 _K-1.

The demultiplexing unit 7074 is supplied with the G ² _n-1 (k) bar γ _n-1 (k) as the past estimated SNR 922 from the multiplex multiplication unit 705 of FIG. Separation unit 7074 is G
² _n-1 (k) bar gamma _n-1 (k) of separating into K frequency-components, as the past frequency domain estimated SNR922 ₀ ~922 _K-1, weighted adder 7071 ₀ ~7071 _K- Propagate to ₁ .

On the other hand, weighted addition units 7071 _{0 to} 707
A weight 923 is also supplied to 1 _K-1 . The weighted addition units 7071 _{0 to} 7071 _K-1 execute the weighted addition represented by Expression 13 and perform the frequency-specific estimation a priori SNR92.
The 4 ₀ ~924 _K-1 is transmitted to the multiplexing unit 7075.

The multiplexing unit 7075 multiplexes the estimated a priori SNRs 924 _{0 to} 924 _K-1 for each frequency to obtain the estimated a priori SN.
Output as R924.

The weighted addition units 7071 _{0 to} 707 are added.
The operation and configuration of 1 _K-1 are the same as those of the weighted addition unit 407 described with reference to FIG. 28, but the weighted addition calculation is always performed.

FIG. 37 shows the noise suppression coefficient generator 8 of FIG.
3 is a block diagram showing the configuration of FIG. Noise suppression coefficient generator 8
Is a suppression coefficient search unit 801 _{0 to} 801 _K-1 and a separation unit 80.
2, 803 and a multiplexing unit 804.

The separating unit 802 stores the SN for each frequency in FIG.
The acquired SNR is supplied from the R calculation unit 6. Separation part 80
2 separates the supplied acquired SNR into K frequency components and transmits the _K frequency components to the suppression coefficient search units 801 _{0 to} 801 _K-1 .

The separating unit 803 stores the estimated a priori S
The estimated a priori SNR is supplied from the NR calculator 7. The separation unit 803 separates the supplied estimated a priori SNR into K frequency components, and the suppression coefficient search units 801 _{0 to} 801 _K-1.
Communicate to.

The suppression coefficient search units 801 _{0 to} 801 _K-1 search the supplied acquired SNR and the suppression coefficient corresponding to the estimated a priori SNR, and transmit the search result to the multiplexing unit 804.

The multiplexing unit 804 multiplexes the supplied suppression coefficient and outputs it.

FIG. 38 is a block diagram of the suppression coefficient search unit 801 of FIG.
It is a block diagram showing a _{0 ~801 K-1} configuration. The suppression coefficient search unit 801 includes a suppression coefficient table 8011 and address conversion units 8012 and 8013.

The address conversion unit 8012 is supplied with the acquired SNR for each frequency from the separation unit 802 of FIG. The address conversion unit 8012 uses the acquired frequency-dependent acquired S
The NR is converted into a corresponding address, and the suppression coefficient table 8
011 is transmitted.

The address conversion unit 8013 is supplied with the estimated a priori SNR for each frequency from the separation unit 803 in FIG. The address conversion unit 8013 converts the supplied frequency-specific estimated a priori SNR into a corresponding address, and transfers it to the suppression coefficient table 8011.

The suppression coefficient table 8011 outputs the suppression coefficient stored in the area corresponding to the address supplied from the address conversion unit 8012 and the address conversion unit 8013 as the suppression coefficient for each frequency.

[0103]

However, the conventional technique obtains emphasized speech by superposing and adding signal samples taken from two adjacent frames of the time domain signal obtained by the inverse Fourier transform. On the other hand,
The window function applied to the time domain signal before the Fourier transform was designed so that the input is reproduced at the output when no noise suppression processing is performed.

Therefore, when the samples to be superimposed and added are suppressed with different suppression coefficient values in adjacent frames, discontinuity occurs in the signal samples at the frame boundaries, and noise generated in the output signal is generated. There was a problem that subjective sound quality deteriorates.

Therefore, it is an object of the present invention to provide a noise removing method and apparatus capable of obtaining an emphasized voice excellent in subjective sound quality without generating noise in an output signal.

[0106]

In order to solve the above problems, a noise removing method of the present invention converts an input signal into a frequency domain signal, obtains a signal-to-noise ratio based on the frequency domain signal, and Determine the suppression coefficient based on the signal-to-noise ratio,
When the noise-removed output signal is obtained by weighting the frequency domain signal based on the suppression coefficient and converting the weighted frequency domain signal into a time domain signal, the first signal is added to the input signal. After windowing, the signal is transformed into a frequency domain signal, and the time domain signal is subjected to a second windowing to obtain the output signal.

Further, the noise removing method of the present invention converts an input signal into a frequency domain signal, obtains estimated noise based on the frequency domain signal, and subtracts a value corresponding to the estimated noise from the frequency domain signal. When obtaining the emphasized speech in the frequency domain and obtaining the output signal from which noise is removed by converting the emphasized speech into the time domain signal, the input signal is first windowed and then converted into the frequency domain signal. , A second window is applied to the time domain signal to obtain an output signal.

Further, the noise removing apparatus of the present invention frequency-converts the first windowing processing section for subjecting the input signal to the windowing processing, and the input signal subjected to the windowing processing by the first windowing processing section. A conversion unit for converting into a domain signal, a signal-to-noise ratio calculation unit for obtaining a signal-to-noise ratio based on the amplitude component of the frequency-domain signal converted by the conversion unit, and a signal obtained by the signal-to-noise ratio calculation unit A suppression coefficient generation unit that generates a suppression coefficient based on a noise ratio; a multiplication unit that weights the amplitude component of the frequency domain signal based on the suppression coefficient generated by the suppression coefficient generation unit; An inverse transform unit for transforming the amplitude component of the frequency domain signal weighted by the unit and the phase component of the frequency domain signal transformed by the transform unit into a time domain signal, and a time domain signal transformed by the inverse transform unit. First window processing And a windowing processing unit.

Further, the present invention is characterized in that emphasized speech is obtained by subjecting the inverse Fourier transform outputs extracted from two adjacent frames to windowing processing and then superposing and adding. More specifically, the windowing processing unit is provided for windowing the signal samples extracted from two adjacent frames of the inverse Fourier transform output.

That is, the present invention improves the continuity of signal samples at frame boundaries and prevents noise by windowing the inverse Fourier transform outputs that form two adjacent frames and then performing superposition addition. I am trying.

[0111]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings.

(Embodiment 1) FIG. 1 is a block diagram showing a schematic configuration of a noise removing apparatus according to Embodiment 1 of the present invention. The noise eliminator shown in FIG. 1 is different from the noise eliminator shown in FIG. 25 in that it has a windowing processing unit 22. In FIG. 1, the same parts as those shown in FIG. 25 are designated by the same reference numerals. The operation will be described below focusing on this difference.

The windowing processing section 22 includes an inverse Fourier transform section 9
The window function h (t) is multiplied by the X _n (t) bar supplied from the above, and the product h (t) X _n (t) bar is transmitted to the frame synthesis unit 10.

The frame synthesizing unit 10 uses h (t) X
K / 2 samples are taken out from two adjacent frames of the _n (t) bar and superposed on each other, and the emphasized speech X _n (t) hat is obtained by Expression 14.

[0115]

[Equation 14] Obtained emphasized speech X _n (t) hat (t = 0, 1, ..., K
/ 2-1) is transmitted to the output terminal 12 as the output of the frame synthesizing unit 10. When the overlap is M samples instead of 50% and the frame length is L samples (M <L), Expression 15 obtains the emphasized speech X _n (t) hat.

[0116]

[Equation 15] In accordance with this, the frame division unit 1 is also modified.

As described above, the symmetric window function is used for real signals. Further, the window function is designed so that the input signal and the output signal when the suppression coefficient is set to 1 match each other except for a calculation error.

Any window function satisfying these conditions can be used as h (t). As an example, a function (root Hanning window) in which the Hanning window is square rooted can be cited. Since there are other window functions that satisfy these conditions, they may be used.

(Embodiment 2) FIG. 2 is a block diagram showing a schematic configuration of a noise removing apparatus according to Embodiment 2 of the present invention. 2 and 1 are the same except for the estimated noise calculation unit 5, the weighted deteriorated speech calculation unit 14, and the suppression coefficient correction unit 15, and the same reference numerals are given to the same portions.

The structure shown in FIG. 2 is "April 2000, IEICE Technical Research Report, DSP, pages 53-60" (hereinafter,
It is referred to as "reference 5". ), The windowing processing unit 22 is provided.

Unlike the method described in Document 1, the method described in Document 5 can obtain an accurate estimated noise by estimating the power spectrum of noise using the weighted deteriorated speech spectrum. it can. The operation will be described below focusing on these differences.

FIG. 3 shows the weighted deteriorated speech calculation unit 1 of FIG.
4 is a block diagram showing the configuration of FIG. The weighted deteriorated speech calculation unit 14 includes an estimated noise storage unit 1401, SNR for each frequency.
The calculation unit 1402, the multiplex nonlinear processing unit 1405, and the multiplex multiplication unit 1404 are included.

The estimated noise storage section 1401 stores the estimated noise power spectrum supplied from the estimated noise calculation section 5 of FIG. 2 and outputs the estimated noise power spectrum stored one frame before to the frequency-based SNR calculation section 1402. To do.

The frequency-based SNR calculation section 1402 uses the estimated noise power spectrum supplied from the estimated noise storage section 1401 and the degraded speech power spectrum supplied from the multiplex multiplication section 17 of FIG. 2 to calculate the SNR for each frequency. Seeking,
It outputs to the multiplex nonlinear processing unit 1405.

The multiple non-linear processing unit 1405 determines S for each frequency.
The weighting coefficient vector is obtained using the SNR supplied from the NR calculating section 1402, and the weighting coefficient vector is multiplied by the multiplying section 1.
Output to 404.

The multiplex multiplication unit 1404 calculates, for each frequency, the product of the deteriorated speech power spectrum supplied from the multiplex multiplication unit 17 of FIG. 2 and the weighting coefficient vector supplied from the multiplex nonlinear processing unit 1405, and the weighting is performed. The degraded voice power spectrum is output to the estimated noise storage unit 5 in FIG.

The configuration of the frequency-dependent SNR calculation section 1402 is as follows.
This is the same as the frequency-based SNR calculation unit 6 described using FIG. 33. The configuration of the multiplex multiplication unit 1404 is similar to that of the multiplex multiplication unit 17 described with reference to FIG.

FIG. 4 is a block diagram of the multiple nonlinear processor 1405 of FIG.
3 is a block diagram showing the configuration of FIG. Multiple Nonlinear Processing Unit 14
Reference numeral 05 denotes a separation unit 1495 and nonlinear processing units 1485 ₀ to 1 485.
It has 485 _K−1 and a multiplexing unit 1475.

The separating unit 1495 is provided with the SNR for each frequency in FIG.
The SNR supplied from the calculation unit 1402 is set to the SN for each frequency.
It is separated into R and output to the non-linear processing units 1485 _{0 to} 1485 _K-1 .

The non-linear processing units 1485 _{0 to} 1485 _K-1 are
Each has a non-linear function that outputs a real value according to an input value.

FIG. 5 shows the non-linear processing units 1485 _{0 to} 148.
It is a figure which shows the example of the nonlinear function output from _5K-1 . When f ₁ is used as an input value, the output value f ₂ of the nonlinear function shown in FIG.

[0132]

[Equation 16] The non-linear processing units 1485 _{0 to} 1485 _K-1 are separated by the separation unit 149.
The frequency-dependent SNR supplied from the signal No. 5 is processed by a non-linear function to obtain a weighting coefficient, which is output to the multiplexing unit 1475. That is, the non-linear processing units 1485 _{0 to} 1485 _K-1 output weighting factors from 1 to 0 according to the SNR. S
When NR is small, 1 is output, and when it is large, 0 is output.

The multiplexing unit 1475 is a nonlinear processing unit 148.
5 _0-1485 weighting coefficient outputted from the _K-1 multiplexing,
The weighting coefficient vector is output to the multiple multiplication unit 1404.

The weighting coefficient to be multiplied by the deteriorated voice power spectrum in the multiplex multiplication unit 1404 in FIG. 3 has a value corresponding to the SNR. The larger the SNR, that is, the larger the voice component included in the deteriorated voice, The value of the weighting factor becomes smaller.

Generally, the deteriorated speech power spectrum is used for updating the estimated noise, but the deteriorated speech power spectrum used for updating the estimated noise is weighted according to the SNR to be included in the deteriorated speech power spectrum. The influence of the voice component generated can be reduced, and more accurate noise estimation can be performed.

Although an example of using a non-linear function for the calculation of the weighting coefficient is shown, it is also possible to use a SNR function represented in another form such as a linear function or a high-order polynomial in addition to the non-linear function. is there.

FIG. 6 is a block diagram showing the configuration of the estimated noise calculation section 5 of FIG. The estimated noise calculation unit 5 of FIG. 2 is different from the estimated noise calculation unit 51 shown in FIG.
5 and the estimated noise calculation unit for each frequency 514 ₀ to 514 ₀ to
By frequency 514 _K-1 estimated noise calculator 504 _0-504
The difference is that it is replaced with _K-1 . In FIG. 6, the same parts as those shown in FIG. 30 are designated by the same reference numerals. The operation will be described below focusing on these differences.

The separation unit 505 separates the weighted deteriorated speech power spectrum supplied from the weighted deteriorated speech calculation unit of FIG. 2 into frequency-dependent weighted deteriorated sound power spectra, and the frequency-dependent estimated noise calculation unit 504 _0. Output to ~ 504 _K-1 .

Frequency-dependent estimated noise calculators 504 _{0 to} 504
_K-1 is the frequency-dependent deteriorated voice power spectrum supplied from the separation unit 502, the frequency-weighted deteriorated voice power spectrum supplied from the separation unit 505, the voice detection flag supplied from the voice detection unit 4 in FIG. The estimated noise power spectrum for each frequency is calculated from the count value supplied from the counter 13 in FIG. 2 and output to the multiplexing unit 503.

The multiplexing unit 503 multiplexes the frequency _- dependent estimated noise power spectrum supplied from the frequency _- dependent estimated noise calculation units 504 _{0 to} 504 _K-1 , and the estimated noise power spectrum is calculated by the frequency-based SNR calculation unit 6 in FIG. To the weighted deteriorated speech calculation unit 14.

FIG. 7 shows the frequency-dependent estimated noise calculator 5 of FIG.
04₀~ 504_K-13 is a block diagram showing the configuration of FIG. Figure 6
Frequency-dependent estimated noise calculation unit 504₀~ 504_K-1Is shown in FIG.
The frequency-dependent estimated noise calculation unit 514 shown in FIG.
Estimated noise calculation unit 504 by wave number ₀~ 504_K-1Is estimated noise
In addition, the update determination unit 521 determines that the update determination unit 521 has the storage unit 5942.
To the switch 5044 and the point that the constant part 520 is replaced.
Input the frequency from the degraded voice power spectrum by frequency
It is replaced with another weighted speech power spectrum
The point is different. Note that the section shown in FIG. 31 in FIG.
The same reference numerals are given to the same parts as the minutes.

Incidentally, the frequency _- dependent estimated noise calculation units 504 _{0 to} 504 _K-1 shown in FIG. 6 may be configured as will be described later with reference to FIG.

Frequency-dependent estimated noise calculators 504 _{0 to} 504
_K-1 uses the weighted deteriorated speech power spectrum instead of the deteriorated speech power spectrum to calculate the estimated noise, and uses the estimated noise and the deteriorated speech power spectrum to update the estimated noise. Difference occurs.

The estimated noise storage unit 5942 has a division unit 504.
The estimated noise power spectrum for each frequency supplied from 8 is stored, and the estimated noise power spectrum for each frequency stored one frame before is output to the update determination unit 520.

FIG. 8 is a block diagram showing the structure of the update determination unit 520 shown in FIG. The update determination unit 520 of FIG. 7 is different from the update determination unit 521 shown in FIG. 31 in that the logical sum calculation unit 5211 is replaced by the logical sum calculation unit 5201, and the update determination unit 520 has a comparison unit 5205. Threshold storage unit 5
The difference is that 206 and a threshold calculation unit 5207 are included. In FIG. 8, the same parts as those shown in FIG. 32 are designated by the same reference numerals. The operation will be described below focusing on these differences.

The threshold calculation unit 5207 calculates a value according to the frequency-dependent estimated noise power spectrum supplied from the estimated noise storage unit 5942 of FIG. 7, and outputs the calculation result to the threshold storage unit 5206 as a threshold value. The simplest is to set the threshold value to a constant multiple of the frequency-dependent estimated noise power spectrum. Besides, it is also possible to calculate the threshold value using a high-order polynomial or a non-linear function.

The threshold storage unit 5206 has a threshold calculation unit 520.
The threshold value output from No. 7 is stored, and the threshold value stored one frame before is output to the comparison unit 5205.

The comparing unit 5205 compares the threshold value supplied from the threshold value storage unit 5206 with the frequency-dependent deteriorated voice power spectrum supplied from the separating unit 502 of FIG. 6, and if the frequency-dependent deteriorated voice power spectrum is smaller than the threshold value. If it is larger, “1” is output to the logical sum calculation unit 5201. That is, whether or not the deteriorated voice signal is noise is determined based on the size of the estimated noise power spectrum.

The logical sum calculation unit 5201 includes a comparison unit 5203.
Of the output value of, the output value of the logical NOT circuit 5202, and the output value of the comparison unit 5205 are calculated, and the calculation result is output to the switch 5044, the shift register 5045, and the counter 5049 in FIG. 7.

As described above, when the deteriorated voice power is small not only in the initial state or in the silent section but also in the voiced section, the update determining section 520 outputs "1". That is, the estimated noise is updated. Since the threshold value is calculated for each frequency, the estimated noise can be updated for each frequency.

In FIG. 7, a CNT counter 5049
, And N is the register length of the shift register 5045. Then, B _n (k) (n = 0, 1, ..., N−
Let 1) be the frequency-dependent weighted deteriorated speech power spectrum accumulated in the shift register 5045. At this time, the frequency-dependent estimated noise power spectrum λ _n (k) output from the division unit 5048 is as shown in Expression 17.

[0152]

[Equation 17] That is, λ _n (k) is an average value of the weighted deteriorated speech power spectrum for each frequency stored in the shift register 5045. The average value may be calculated using a weighted addition unit (cyclic filter).

FIG. 9 shows the frequency-dependent estimated noise calculation unit 5 of FIG.
04 is a block diagram showing a configuration example of a _{0 ~504 K-1.} The frequency _- dependent estimated noise calculators 504 _{0 to} 504 _K-1 in FIG. 6 are the shift register 5045, the adder 5046, the minimum value selector 5047, the divider 5048, the counter 5049, and the shift register 5045 in FIG.
Instead of the register length storage unit 5941, the weighted addition unit 5
071 and a weight storage unit 5072 are different. In FIG. 9, the same parts as those shown in FIG. 6 are designated by the same reference numerals.

The weighted addition unit 5071 outputs the estimated noise power spectrum for each frequency before one frame supplied from the estimated noise storage unit 5942, the weighted deteriorated speech power spectrum for each frequency supplied from the switch 5044, and the weight storage unit 5072. Frequency-based estimation noise is calculated using the output weights and output to the multiplexing unit 503.

That is, when the weight stored in the weight storage unit 5072 is δ and the frequency-dependent weighted deteriorated speech power spectrum is | Y _n (k) | ² bar, the weighted addition unit 50
The estimated noise power spectrum λ _n (k) for each frequency output from 71 can be expressed by Equation 18.

[0156]

[Equation 18] The configuration of the weighted addition unit 5071 is the same as that of the weighted addition unit 407 described with reference to FIG. 28, but the weighted addition calculation is always performed.

FIG. 10 is a block diagram showing the configuration of the suppression coefficient correction unit 15 of FIG. In order to prevent residual noise generated due to insufficient suppression when the SNR is low and sound quality deterioration due to audio distortion that occurs due to excessive suppression when the SNR is high, the suppression coefficient correction unit 15 corrects the suppression coefficient according to the SNR. I do.

As an example of correction, when the SNR is low, a correction value is added to the suppression coefficient to suppress residual noise, and when the SNR is high, a lower limit value is set to the suppression coefficient to prevent voice distortion. The suppression coefficient correction unit 15 includes frequency-based suppression coefficient correction units 1501 _{0 to} 1501 _K−1 and separation units 1502 and 1503.
And a multiplexing unit 1504.

Separation unit 1502 uses estimated congenital SN of FIG.
The estimated a priori SNR supplied from the R calculation unit 7 is separated into frequency components, and each frequency suppression coefficient correction unit 150
And outputs it to the _{1 0} ~1501 _K-1.

The separation unit 1503 separates the suppression coefficient supplied from the suppression coefficient generation unit 8 of FIG. 2 into frequency components, and the frequency suppression coefficient correction units 1501 _{0 to} 1501 _{K-1 respectively.}
Output to.

Frequency-dependent suppression coefficient correction units 1501 ₀ to 150
01 _K−1 calculates a frequency-specific correction suppression coefficient from the frequency-specific estimated a priori SNR supplied from the separation section 1502 and the frequency-specific suppression coefficient supplied from the separation section 1503, and outputs it to the multiplexing section 1504. To do.

The multiplexing unit 1504 multiplexes the frequency _- dependent correction suppression coefficients supplied from the frequency-specific suppression coefficient correction units 1501 _{0 to} 1501 _K-1, and multiplexes them as the correction suppression coefficient and the estimated a priori SNR calculation unit. And output to 7.

FIG. 11 shows frequency _- dependent suppression coefficient correction units 1501 _{0 to} 1501 _K-1 included in the suppression coefficient correction unit 15 shown in FIG.
3 is a block diagram showing the configuration of FIG. The frequency-dependent suppression coefficient correction unit 1501 includes a maximum value selection unit 1591, a suppression coefficient lower limit storage unit 1592, a threshold storage unit 1593, and a comparison unit 159.
4, a switch 1595, a correction value storage unit 1596, and a multiplier 1597.

The suppression coefficient lower limit storage unit 1592 supplies the stored lower limit value of the suppression coefficient to the maximum value selection unit 1591.

The maximum value selection unit 1591 compares the suppression coefficient for each frequency supplied from the separation unit 1503 of FIG. 10 with the suppression coefficient lower limit value supplied from the suppression coefficient lower limit value storage unit 1592, and determines the larger value. Output to the switch 1595.
That is, the suppression coefficient has a large value equal to or larger than the lower limit value stored in the suppression coefficient lower limit value storage unit 1592. Therefore, it is possible to prevent the distortion of the voice generated by the excessive suppression.

The comparing unit 1594 compares the threshold value supplied from the threshold value storage unit 1593 with the frequency-specific estimated a priori SNR supplied from the separating unit 1502 in FIG. 10, and the frequency-specific estimated a priori SNR is higher than the threshold value. If it is larger, “0” is supplied to the switch 1595, and if it is smaller, “1” is supplied to the switch 1595.

The switch 1595 has a maximum value selection section 159.
The signal supplied from 1 is output to the multiplier 1597 when the output value of the comparison unit 1594 is “1”, and the comparison unit 1594
When the output value of is 0, it is output to multiplexing section 1504 in FIG. That is, the suppression coefficient is corrected when the estimated a priori SNR for each frequency is smaller than the threshold value. In this way, the residual noise amount can be reduced without excessively suppressing the voice component.

The multiplier 1579 calculates the product of the output value of the switch 1595 and the output value of the correction value storage unit 1596,
The calculation result is output to the multiplexing unit 1504 in FIG. The correction value is usually smaller than 1 in order to reduce the suppression coefficient value, but it is not limited to this depending on the purpose.

In this embodiment, the correction suppression coefficient is not supplied to the multiplex multiplier 16 and the estimated a priori SNR calculator 7, but the corrected suppression coefficient is supplied to the multiplex multiplier 16 and the estimated a priori SN.
It is supplied to the R calculation unit 7.

FIG. 12 is a block diagram showing the configuration of a noise suppression coefficient generator 81 which is a modification of the noise suppression coefficient generator 8 of FIG. The noise suppression coefficient generation unit 81 includes a gain function value calculation unit 811, a generalized likelihood ratio calculation unit 812, a voice existence probability storage unit 813, and a suppression coefficient calculation unit 814.

The noise suppression coefficient generator 81 shown in FIG.
The difference from the noise suppression coefficient generation unit 8 of FIG. 2 that is obtained by search is that the suppression coefficient is calculated from the supplied estimated innate SNR and acquired SNR. Hereinafter, the calculation method of the suppression coefficient will be described based on the calculation formula described in Document 1.

The frame number is n and the frequency number is k,
γ _n (k) is the frequency-dependent a posteriori SNR supplied from the frequency-based SNR calculating unit 6 in FIG. 2, and ξ _n (k) hat is the frequency-based estimation a priori supplied from the estimated a priori SNR calculating unit 7 in FIG. Target SNR. _{Moreover, η n (k) = ξ} n (k) hat _{/ q v n (k) =} (η n (k) γ n (k)) / (1 + η
_n (k)).

The gain function value calculation unit 811 receives the acquired SNR γ supplied from the frequency-based SNR calculation unit 6 of FIG.
_n (k), the estimated a priori SNR ξ _n (k) hat supplied from the estimated a priori SNR calculation unit 7 and the voice existence probability q supplied from the voice existence probability storage unit 813. For example, the MMSE STSA gain function value as described in Reference 1 is calculated for each frequency and output to the suppression coefficient calculation unit 814.

The MMSE STSA gain function value G _n (k) for each frequency can be expressed by Equation 19.

[0175]

[Formula 19] Where I ₀ (z) is the 0th-order modified Bessel function, and I ₁ (z)
Is a first-order modified Bessel function. The modified Bessel function is described in "1985, Mathematical Dictionary, Iwanami Shoten, page 374.G" (Reference 6).

The generalized likelihood ratio calculation unit 812 receives the acquired SNR γ supplied from the frequency-based SNR calculation unit 6 of FIG.
_n (k), the estimated a priori SNR Λ _n (k) hat supplied from the estimated a priori SNR calculation unit 7 in FIG. 2 and the voice existence probability q supplied from the voice existence probability storage unit 813. Then, the generalized likelihood ratio is calculated and output to the suppression coefficient calculation unit 814.

The generalized likelihood ratio Λ _n (k) for each frequency can be expressed by Equation 20.

[0178]

[Equation 20] The suppression coefficient calculation unit 814 supplies the MMSESTSA gain function value G _n (k) supplied from the gain function value calculation unit 811.
2 and the generalized likelihood ratio Λ _n (k) supplied from the generalized likelihood ratio calculation unit 812, the suppression coefficient is calculated for each frequency, and FIG.
To the suppression coefficient correction unit 15.

The suppression coefficient G _n (k) bar for each frequency can be expressed by Equation 21.

[0180]

[Equation 21] Instead of calculating the SNR for each frequency, it is also possible to obtain an SNR common to a band composed of a plurality of frequencies and use this.

FIG. 13 is a frequency-specific SNR calculation unit 6 of FIG.
9 is a block diagram showing a configuration example of a frequency-based SNR calculation unit 61 that is a modification example of FIG. Frequency-dependent SNR calculator 6 in FIG.
1 corresponds to the frequency-based SNR calculation unit 6 shown in FIG.
The band-based SNR calculation unit 61 is a band-based power calculation unit 611.
612 is different. In FIG. 13, the same parts as those shown in FIG. 33 are designated by the same reference numerals.

The power calculation section 611 for each band is composed of the separation section 60.
Frequency noisy speech power spectrum supplied from 2 to calculate the per-band power on the basis of the division unit 601 _0-60
Output to 1 _K-1 .

The power calculation section 612 for each band is divided by the separation section 60.
The power for each band is calculated based on the estimated noise power spectrum for each frequency supplied from No. 3, and the division units 601 _{0 to} 60 ₀
Output to 1 _K-1 .

FIG. 14 is a band-specific power calculation unit 6 of FIG.
It is a block diagram which shows the structure of 11. Here, an example of equally dividing into M bands having a bandwidth L will be described. Here, it is assumed that L and M are natural numbers that satisfy the relationship of K = LM.

The band-by-band SNR calculation unit 61 uses the adder 611.
With a ₀ 0 ~6110 _M-1. Frequency-dependent degraded voice power spectrum 910 ₀ supplied from the separation unit 602 of FIG.
.About.910 _K-1 (910 _{0 to} 910 _ML-1 ) are transmitted to the adders 6110 _{0 to} 6110 _M-1 corresponding to the respective frequencies.

[0186] For example, since L-1 frequency number from 0 corresponding to the band number 0, frequency noisy speech power spectrum 910 ₀ ～910 _L-1 are transmitted to adder 6110 _0. The frequency numbers corresponding to band number 1 are from L to 2
Since it is L-1, the degraded voice power spectrum 91 for each frequency
0 _{L to} 910 _2L-1 is transmitted to the adder 6110 ₁ .

The adders 6110 _{0 to} 6110 _M-1 respectively calculate the sums of the supplied frequency _- dependent degraded voice power spectra, and the band-specific degraded voice power spectra 911 ₀ to 910.
911 _ML-1 (911 _{0 to} 911 _K-1 ) is divided by the division unit 6 in FIG.
Output to 01 _{0 to} 601 _K-1 . Each adder 6110 _{0 to} 6
The calculation result of 110 _M-1 is output as a band-specific deteriorated speech power spectrum for each frequency corresponding to each band number.

For example, the calculation result of the adder 6110 ₀ is
The deteriorated voice power spectrum for each band is output as 911 _{0 to} 911 _L-1 . The calculation result of the adder 6110 _1, the bandwidth noisy speech power spectrum 911 _L ~911
It is output as _2L-1 . The configuration and operation are similar to those of the band-specific power calculation unit 611.

Although an example of equally dividing into a plurality of bands has been shown here, the critical band described in "Hearing and Speech, The Institute of Electronics, Information and Communication Engineers, Pages 115 to 118, 1980" (Reference 7) is used. Dividing method, 1983, Multirate Digital Signal Processing
gnal Processing), 1983, Prentice-Hall Inc., USA "(hereinafter referred to as" Reference 8 "), and other band division methods can also be used. .

(Embodiment 3) FIG. 15 is a block diagram showing a schematic configuration of a noise removing device according to Embodiment 3 of the present invention. The noise removal apparatus shown in FIG. 15 differs from FIG. 2 in that the estimated noise calculation unit 5 is replaced by the estimated noise calculation unit 52 and that the weighted deteriorated speech calculation unit 14 does not exist. In FIG. 15, the same parts as those shown in FIG. 2 are designated by the same reference numerals. The operation will be described below focusing on these differences.

FIG. 16 is a block diagram showing the structure of the estimated noise calculation section 52 shown in FIG. In the estimated noise calculation unit 52 of FIG. 15, in contrast to the estimated noise calculation unit 5 of FIG. 2, the frequency _- dependent estimated noise calculation units 504 _{0 to} 504 _K-1 are changed to frequency _- dependent estimated noise calculation units 506 _{0 to} 506 _K-1 . The difference lies in that the estimated noise calculation unit 52 does not have a weighted deteriorated speech power spectrum in the input signal. 16, the same parts as those shown in FIG. 6 are designated by the same reference numerals.

Since the frequency _- dependent estimated noise calculation units 504 _{0 to} 504 _K-1 require the frequency _- dependent weighted deteriorated speech power spectrum for the input signal, the estimated noise calculation does not have the deteriorated speech power spectrum. Parts 506 _{0 to} 506
_{This is because K-1} does not require a weighted degraded speech power spectrum for each input signal.

FIG. 17 is a block diagram showing the structure of the frequency _- dependent estimated noise calculation units 506 _{0 to} 506 _K-1 shown in FIG.
Frequency-dependent estimated noise calculation units 506 _{0 to} 506 in FIG.
_K-1 is the frequency-dependent estimated noise calculation unit 504 _{0 to} 504 in FIG.
_{For K-1} , the frequency-dependent estimated noise calculation unit 506 does not have a frequency-dependent weighted deteriorated speech power spectrum in the input signal, the division unit 5041, and the non-linear processing unit 5042.
And a multiplier 5043 is provided. 17, the same parts as those shown in FIG. 7 are designated by the same reference numerals. The operation will be described below focusing on these differences.

The division unit 5041 is the separation unit 502 of FIG.
Deteriorated voice power spectrum for each frequency supplied from
It divides by the estimated noise power spectrum one frame before supplied from the estimated noise storage unit 5942, and outputs the division result to the non-linear processing unit 5042. A non-linear processing unit 5 having the same configuration and function as the non-linear processing unit 1485 shown in FIG.
042 calculates a weighting coefficient according to the output value of the division unit 5041, and outputs it to the multiplier 5043.

The multiplier 5043 is the separation unit 502 of FIG.
Calculates the product of the frequency-dependent deteriorated speech power spectrum supplied from the product and the weighting coefficient supplied from the non-linear processing unit 5042, and outputs the product to the switch 5044. The output signal of the multiplier 5043 is the same as the weighted deteriorated speech power spectrum by frequency in the estimated noise by frequency calculation unit 504 in FIG. That is, the weighted deteriorated speech power spectrum for each frequency can be calculated in the estimated noise for each frequency calculation unit 506. Therefore, the weighted deteriorated speech calculation unit 14 can be omitted.

(Embodiment 4) FIG. 18 is a block diagram showing a schematic configuration of a noise removing device according to Embodiment 4 of the present invention. The noise removing apparatus shown in FIG. 18 is different from the noise removing apparatus shown in FIG. 2 in the estimated a priori SNR calculation unit 71.
The difference is that it has. 18, the same parts as those shown in FIG. 2 are designated by the same reference numerals. The operation will be described below focusing on this difference.

FIG. 19 is a block diagram showing the configuration of the estimated a priori SNR calculation unit 71 of FIG. The estimated a priori SNR calculation unit 71 of FIG. 18 includes the acquired a priori SNR storage unit 702 and the suppression coefficient storage unit 7 of the estimated a priori SNR calculation unit 7 of FIG.
03, instead of the multiple multiplication units 705 and 704, an estimated noise storage unit 712, an emphasized speech power spectrum storage unit 713,
It has a frequency-specific SNR calculation unit 715 and a multiple multiplication unit 716. In FIG. 19, the same parts as those shown in FIG. 34 are designated by the same reference numerals.

Here, the estimated a priori SNR calculation unit 71 is
Instead of the suppression coefficient, the emphasized speech amplitude spectrum and the estimated noise power spectrum are used as input signals.

The multiplying unit 716 squares the emphasized voice amplitude spectrum supplied from the multiplex multiplier 16 of FIG. 18 for each frequency to obtain an emphasized voice power spectrum, and outputs it to the emphasized voice power spectrum storage unit 713. The configuration of the multiple multiplication unit 716 is the multiple multiplication unit 1 described with reference to FIG.
Similar to 7.

Enhanced voice power spectrum storage unit 713
Stores the emphasized voice power spectrum supplied from the multiplex multiplication unit 716, and outputs the emphasized voice power spectrum supplied one frame before to the frequency-based SNR calculation unit 715. The configuration of the frequency-based SNR calculation unit 715 is the same as that of the frequency-based SNR calculation unit 6 described with reference to FIG.

The estimated noise storage unit 712 stores the estimated noise power spectrum supplied from the estimated noise calculation unit 5 of FIG. 18, and outputs the estimated speech power spectrum supplied one frame before to the frequency-based SNR calculation unit 715. To do.

The frequency-based SNR calculation unit 715 calculates the SNR of the emphasized speech power spectrum supplied from the emphasized speech power spectrum storage unit 713 and the estimated noise power spectrum supplied from the estimated noise storage unit 712 for each frequency. , To the multiple weighted addition unit 707.

The output signal of the frequency-specific SNR calculation unit 715 and the output signal of the multiplex multiplication unit 705 of FIG. 34 are the same. Therefore, in the present embodiment, it is possible to replace the estimated a priori SNR calculator 7 with the estimated a priori SNR calculator 17.

(Fifth Embodiment) FIG. 20 is a block diagram showing a schematic configuration of a noise removing apparatus according to a fifth embodiment of the present invention. The noise removing apparatus shown in FIG. 20 is different from the noise removing apparatus shown in FIG. 2 in that the estimated noise calculation unit 5 is the estimated noise unit 52 and the estimated a priori SNR calculation unit 7 is the estimated a priori SN.
The difference is that the R calculation unit 71 is replaced and the weighted deteriorated speech calculation unit 14 does not exist.

The structure and operation of the estimated noise unit 52 are the same as those of the estimated noise unit described in the third embodiment. The estimated a priori SNR calculation unit 71 is the same as the estimated a priori SNR calculation unit described in the fourth embodiment.

(Sixth Embodiment) FIG. 21 is a block diagram showing a schematic configuration of a noise removing device according to a sixth embodiment of the present invention. The noise removal apparatus of FIG. 21 differs from the noise removal apparatus shown in FIG. 2 in that the estimated noise calculation unit 5 is replaced by the estimated noise calculation unit 53 and that the voice detection unit 4 does not exist. To do. That is, the noise removal device of FIG. 21 does not need a voice detection unit for noise estimation. The operation will be described below focusing on these differences.

FIG. 22 is a block diagram showing the structure of the estimated noise calculation unit 53 shown in FIG. Estimated noise calculation unit 53 of FIG. 21, with respect to estimated noise calculator 5 shown in FIG. 6, the frequency domain estimated noise calculator 504 ₀ ~504 _K-1 frequency domain estimated noise calculator 508 ₀ ~508 _{K - The} point replaced by ₁ , and
The difference is that the estimated noise calculation unit 53 does not have a voice detection flag in the input signal.

FIG. 23 is a block diagram showing the configuration of the frequency _- dependent estimated noise calculation units 508 _{0 to} 508 _K-1 shown in FIG.
Frequency-dependent estimated noise calculation units 508 _{0 to} 508 in FIG.
_K-1 is that the update determining unit 520 is replaced by the update determining unit 522 in the frequency-specific estimated noise calculating unit 504 shown in FIG. 7, and the frequency-specific estimated noise calculating units 508 ₀ to 50 _0.
The difference is that 8 _K-1 does not have a voice detection flag at the input.

FIG. 24 is a block diagram showing the structure of the update determination unit 522 of FIG. The update determination unit 522 of FIG.
8 is that the logical sum calculation unit 5201 is replaced by the logical sum calculation unit 5221 with respect to the update determination unit 520 shown in FIG. 8 and that the update determination unit 522 does not have the logical negation circuit 5202. The difference is that the input signal does not have a voice detection flag. That is, the update determination unit 522 does not use the voice detection flag for updating the estimated noise. This point is different from the update determination unit 520 of FIG.

The logical sum calculation unit 5221 has a comparison unit 5205.
And the output value of the comparison unit 5203 is calculated, and
The calculation result is output to the switch 5044, the shift register 5045, and the counter 5049 in FIG. That is,
The update determination unit 522 always outputs “1” until the count value reaches a preset value, and after reaching the count value, outputs “1” when the deteriorated voice power is smaller than the threshold value.

As described with reference to FIG. 8, the comparison unit 5205 determines whether or not the deteriorated voice signal is noise. That is, it can be said that the comparison unit 5205 performs voice detection for each frequency. Therefore, it is possible to realize an update determination unit that does not have a voice detection flag as an input.

As described above, in each of the embodiments of the present invention, the case of using the minimum mean square error short-time spectrum amplitude method has been described as an example of the noise removal method, but it can be applied to other methods.

[0213] For example, "December 1979, Proceedings of the IEE, Vol. 67, No. 12"
(PROCEEDINGS OF THE IEEE, VOL.67, NO.12, PP.1586-160
4, DEC, 1979), pp. 1586-1604 "(Reference 9), and the Wiener filter method described in" April 1979, Eye.
EEE Transactions on Acoustics Speech and Signal Processing, Volume 27, Issue 2 (IEEE TRANSACTIONS ON ACOUSTICS, SP
EECH, AND SIGNAL PROCESSING, VOL.27, NO.2, PP.113-120,
APR, 1979), pp. 113-120 "(hereinafter referred to as" Reference 10 ").

FIG. 21 is an explanatory diagram of a schematic operation of the spectrum subtraction method described in Document 10. In FIG. 21, if the multiplex multiplication unit 16 is replaced with a multiple subtraction unit, the noise suppression coefficient generation unit 8 is replaced with a noise suppression amount calculation unit, and the suppression coefficient correction unit 15 is replaced with a suppression amount correction unit, the operation by the spectrum subtraction method is realized. be able to.

In the multiple subtraction section, the emphasized voice can be obtained by subtracting the corrected noise suppression amount from the deteriorated voice amplitude spectrum and performing the inverse Fourier transform on the obtained result. Here, after calculating the SNR, the SNR
Although the example in which the noise suppression amount is calculated based on the above is described, the estimated noise obtained by the estimated noise calculation unit 53 can be directly subtracted from the deteriorated speech amplitude spectrum.

[0216]

As described above, the present invention improves the continuity of signal samples at frame boundaries because the inverse Fourier transform outputs that form two adjacent frames are windowed and then superimposed and added. , Noise can be prevented.

[Brief description of drawings]

FIG. 1 is a block diagram showing a schematic configuration of a noise removal device according to a first exemplary embodiment of the present invention.

FIG. 2 is a block diagram showing a schematic configuration of a noise removal device according to a second exemplary embodiment of the present invention.

3 is a block diagram showing a configuration of a weighted deteriorated speech calculation unit 14 of FIG.

4 is a block diagram showing a configuration of a multiplex nonlinear processing unit 1405 of FIG.

FIG. 5 is a diagram showing an example of a non-linear function output from non-linear processing units 1485 _{0 to} 1485 _K-1 .

FIG. 6 is a block diagram showing a configuration of an estimated noise calculation unit 5 in FIG.

[7] the frequency domain estimated noise calculator 504 in FIG. 6 _0-50
It is a block diagram which shows the structure of _4K-1 .

8 is a block diagram showing a configuration of an update determination unit 520 of FIG.

[9] the frequency domain estimated noise calculator 504 in FIG. 6 _0-50
It is a block diagram which shows the structural example of _4K-1 .

10 is a block diagram showing a configuration of a suppression coefficient correction unit 15 of FIG.

11 is a block diagram showing the configuration of frequency-specific suppression coefficient correction units 1501 _{0 to} 1501 _K-1 included in the suppression coefficient correction unit 15 of FIG.

12 is a block diagram showing a configuration of a noise suppression coefficient generation unit 81 which is a modified example of the noise suppression coefficient generation unit 8 in FIG.

13 is a block diagram showing a configuration example of a frequency-based SNR calculation unit 61 which is a modified example of the frequency-based SNR calculation unit 6 of FIG.

FIG. 14 is a block diagram showing a configuration of a band-specific power calculation unit 611 of FIG.

FIG. 15 is a block diagram showing a schematic configuration of a noise removal device according to a third exemplary embodiment of the present invention.

16 is a block diagram showing a configuration of an estimated noise calculation unit 52 in FIG.

17 is a diagram illustrating frequency-dependent estimated noise calculators 506 ₀ to 506 _{0 in} FIG.
It is a block diagram which shows the structure of 506 _K-1 .

FIG. 18 is a block diagram showing a schematic configuration of a noise removal device according to a fourth exemplary embodiment of the present invention.

19 is a block diagram showing a configuration of an estimated a priori SNR calculation unit 71 in FIG.

FIG. 20 is a block diagram showing a schematic configuration of a noise removing device according to a fifth embodiment of the present invention.

FIG. 21 is a block diagram showing a schematic configuration of a noise removing device according to a sixth embodiment of the present invention.

22 is a block diagram showing a configuration of an estimated noise calculation unit 53 in FIG.

FIG. 23 is a diagram illustrating frequency-dependent estimated noise calculation units 508 ₀ to 508 ₀ shown in FIG.
It is a block diagram which shows the structure of 508 _K-1 .

24 is a block diagram showing a configuration of an update determination unit 522 of FIG.

FIG. 25 is a block diagram showing a schematic configuration of a conventional noise suppressor.

FIG. 26 is a block diagram showing a configuration of a voice detection unit 4 in FIG.

27 is a block diagram showing a configuration of a power calculation unit 406 in FIG.

28 is a block diagram showing a configuration of a weighted addition unit 407 of FIG.

FIG. 29 is a block diagram showing the structure of the multiplexing multiplication unit 17 of FIG. 25.

30 is a block diagram showing a configuration of an estimated noise calculation unit 51 of FIG. 25.

31 is a block diagram showing a configuration of a frequency-dependent estimated noise calculation unit 514 of FIG. 30. FIG.

32 is a block diagram showing a configuration of an update determination unit 521 in FIG.

FIG. 33 is a block diagram showing a configuration of a frequency-specific SNR calculation unit 6 of FIG. 25.

34 is a block diagram showing a configuration of an estimated a priori SNR calculation unit 7 in FIG. 25.

35 is a block diagram showing a configuration of a multiple value range limitation processing unit 701 of FIG. 34.

36 is a block diagram showing a configuration of a multiple weighted addition unit 707 of FIG. 34.

FIG. 37 is a block diagram showing the configuration of the noise suppression coefficient generation unit 8 in FIG. 25.

FIG. 38 is a suppression coefficient search unit 801 _{0 to} 801 _K-1 shown in FIG. 37.
3 is a block diagram showing the configuration of FIG.

[Explanation of symbols]

1 frame division unit 2, 22 windowing processing unit 3 Fourier transform unit 4 voice detection unit 5, 51, 52, 53 estimated noise calculation unit 6, 61, 715, 1402 frequency SNR calculation unit 7, 71 estimated a priori SNR calculation Section 8, 81 noise suppression coefficient generation section 9 inverse Fourier transform section 10 frame synthesis section 11 input terminal 12 output terminal 13, 5049 counter 14 weighted deteriorated speech calculation section 15 suppression coefficient correction section 16, 17, 704, 705, 716 1404 Multiple multiplication units 401, 1593, 5204, 5206 Threshold storage units 402, 1594, 5203, 5205 Comparison units 404, 4075 Constant multiplier 405 Logarithmic calculation unit 406 Power calculation units 407, 5071, 7071 _{0 to} 7071 _K-1 Weighted Adders 408, 706, 5072 Weight storages 409, 5202 Logical NOT circuit 502 505,602,603 separation unit 503,604,804,147 multiplexer _{504 0 ~504 K-1, 506} 0 ~506 K-1, 507,5
08 _{0 to} 508 _K-1 , 514 _{0 to} 514 _K-1 Frequency _- dependent estimated noise calculation unit 520, 521, 522 Update determination unit 601 _{0 to} 601 _K-1 , 5041, 5048 Division unit 611, 612 Frequency-specific power calculation unit 701 Multiple range limitation processing section 702 Acquired SNR storage section 703 Suppression coefficient storage section 707 Multiple weighted addition sections 708, 4063, 4072, 4074, 5046, 6
110 _{0 to} 6110 _M-1 adder 712, 1401, 5942 Estimated noise storage unit 713 Enhanced speech power spectrum storage unit 801 _{0 to} 801 _K-1 suppression coefficient search unit 811 Gain function value calculation unit 812 Generalized likelihood ratio calculation unit 813 voice existence probability storage unit 814 suppression coefficient calculation unit 901 deteriorated voice power 902 threshold 903, 923 weight 904 update threshold 905 weighted addition unit control signals 910 _{0 to} 910 _K-1 , 910 _{0 to} 910 _ML-1 deteriorated voice by frequency Power spectrums 911 _{0 to} 911 _K-1 , 911 _{0 to} 911 _ML-1 Band-specific deteriorated speech power spectrum 921 Instantaneous estimation SNR 921 _{0 to} 921 _K-1 Instantaneous estimation SNR 922 by frequency Past estimation SNR 922 _{0 to} 922 _{K- 1} past of the frequency domain estimated SNR 924 estimated apriori SNR 924 ₀ ~924 _K-1 frequency domain estimated inherent SNR 1 05 Multiple nonlinear processor 1485 ₀ ~1485 _K-1, 5042 nonlinear processor 1501 ₀ ~1501 _K-1 frequency-suppression coefficient correction unit 1591,7012 ₀ ~7012 _K-1 maximum value selection unit 1592 suppression coefficient lower-limit value storage unit 1595, 5044 Switch 1596 Correction amount storage unit 1597, 1701 _{0 to} 1701 _K-1 , 4062 _{0 to} 406
2 _K-1 , 4071, 4073, 5043 Multiplier 5045 Shift register 5047 Minimum value selection unit 5201, 5211, 5221 Logical sum calculation unit 5207 Threshold calculation unit 5941 Register length storage unit 7011 Constant storage unit 8011 Suppression coefficient table 8012, 8013 Address Converter

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Masahiro Serizawa             5-7 Shiba 5-1, Minato-ku, Tokyo NEC Corporation             Inside the company F-term (reference) 5D015 EE05

Claims (18)

[Claims] 1. An input signal is converted into a frequency domain signal, a signal-to-noise ratio is obtained based on the frequency-domain signal, a suppression coefficient is determined based on the signal-to-noise ratio, and the frequency is calculated based on the suppression coefficient. When obtaining a noise-removed output signal by weighting the domain signal and converting the weighted frequency domain signal into a time domain signal, the frequency domain is applied after the first windowing is applied to the input signal. Convert it to a signal,
A noise removing method, wherein the output signal is obtained by applying a second windowing to the time domain signal. 2. The first window function when applying the first windowing and the second window function when applying the second windowing, when the suppression coefficient is set to 1 The noise removing method according to claim 1, wherein the noise removing method has a characteristic that an input signal and the output signal match each other. 3. The first window function when applying the first windowing and the second window function when applying the second windowing are equal to each other. Noise removal method. 4. The noise removing method according to claim 1, wherein the frequency domain signal is weighted using a corrected suppression coefficient obtained by correcting the suppression coefficient. 5. The noise removing method according to claim 1, 2 or 3, wherein the suppression coefficient is determined based on a corrected signal-to-noise ratio obtained by correcting the signal-to-noise ratio. 6. A weighted frequency domain signal is obtained by weighting the frequency domain signal, an estimated noise is obtained using the weighted frequency domain signal, and a signal-to-noise is obtained using the estimated noise and the frequency domain signal. The method according to claim 1, 2, 3, 4, or 5, wherein a ratio is obtained. 7. An input signal is converted into a frequency domain signal, an estimated noise is obtained based on the frequency domain signal, and a value corresponding to the estimated noise is subtracted from the frequency domain signal to obtain an emphasized voice in the frequency domain, When the noise-removed output signal is obtained by converting the emphasized speech into a time domain signal, the input signal is first windowed and then converted into a frequency domain signal, and the time domain signal is converted into a second domain signal. A method of removing noise, characterized in that an output signal is obtained by applying the windowing of. 8. The noise removing method according to claim 7, wherein the frequency domain signal is weighted to obtain a weighted frequency domain signal, and the estimated noise is obtained using the weighted frequency domain signal. 9. The noise removing method according to claim 6, wherein when the weighted frequency domain signal is obtained by weighting the frequency domain signal, the weight is obtained using a signal-to-noise ratio. 10. When the frequency domain signal is weighted to obtain a weighted frequency domain signal, a weight is obtained using a signal-to-noise ratio, the weight is processed by a non-linear function, and a correction weight is obtained. 9. The noise removing method according to claim 6, wherein the frequency domain signal is weighted using a correction weight. 11. A first windowing processing unit that performs windowing processing on an input signal, and a conversion unit that converts the input signal that has been windowed by the first windowing processing unit into a frequency domain signal. A signal-to-noise ratio calculation unit that obtains a signal-to-noise ratio based on the amplitude component of the frequency-domain signal that has been converted by the conversion unit; And a multiplication unit for weighting the amplitude component of the frequency domain signal based on the suppression coefficient generated by the suppression coefficient generation unit, and the frequency domain weighted by the multiplication unit. An inverse transform unit that transforms the amplitude component of the signal and the phase component of the frequency domain signal transformed by the transform unit into a time domain signal; and a second windowing process for the time domain signal transformed by the inverse transform unit. Equipped with a window processing unit A noise eliminator characterized by. 12. A first window function when performing windowing processing in the first windowing processing unit and a second window function when performing windowing processing in the second windowing processing unit 12. The noise removing apparatus according to claim 11, wherein the noise removing device has a characteristic that an input signal and an output signal match when the suppression coefficient is set to 1. 13. A first window function used when performing windowing processing in the first windowing processing unit and a second window function used when performing windowing processing in the second windowing processing unit The noise removing device according to claim 11 or 12, wherein the noise removing devices are equal to each other. 14. A suppression coefficient correction unit that corrects the suppression coefficient, wherein the multiplication unit weights the amplitude component of the frequency domain signal using the corrected suppression coefficient corrected by the suppression coefficient correction unit. The noise removing device according to claim 11, 12, or 13. 15. The noise eliminator according to claim 14, wherein the suppression coefficient generation unit includes a correction signal-to-noise ratio calculation unit that corrects the signal-to-noise ratio to obtain a correction signal-to-noise ratio. . 16. The signal-to-noise ratio calculation unit includes an estimated noise calculation unit that estimates noise by using an amplitude component of the frequency domain signal, and the signal-to-noise ratio is calculated based on an estimation result of the estimated noise calculation unit. The ratio is calculated as claimed in claim 1.
The noise removing device according to 1, 12, 13, 14 or 15. 17. The noise removal according to claim 16, wherein the estimated noise calculation unit includes a weighted deteriorated speech calculation unit that weights an amplitude component of the frequency domain signal to obtain a weighted amplitude component. apparatus. 18. The weighted deteriorated speech calculation unit includes a non-linear processing unit that processes the signal-to-noise ratio by a non-linear function to obtain a weighting coefficient.
The noise removal device described.