CN1770264A - Noise removing method and device - Google Patents

Abstract

A noise removing method and device comprising: an incoming signal converter used for converting the incoming signal to frequency-region signal; a weighting calculator used for confirming the weighting according to the said frequency-region incoming signal; a weighting frequency-region signal calculator utilizing the said weighting for weighting the said frequency-region signal in order to confirm the weighted frequency-region signal; And a noise estimation unit used for confirming the noise of estimation according to the weighted frequency-region signal.

Description

Noise removing method and equipment

Technical field

The present invention relates to a kind of method and apparatus that is superimposed upon the noise on the needed voice signal that is used to estimate.

Background technology

Noise suppressor is the equipment that suppresses to be superimposed upon the noise on the needed voice signal.Noise suppressor carries out work, so that utilize the input signal that has been transformed into frequency-region signal to come the power spectrum of estimated noise component, and deducts the noise power spectrum of estimation from input signal, thus the noise that inhibition and needed voice signal mix.Noise suppressor can suppress the instability noise by noiseless part that detects speech and the power spectrum that upgrades noise component.

A kind of noise suppressor is published in the collection of thesis of in Dec, 1984 IEEE (IEEE) acoustics, speech and signal Processing, 6,1109 to 1121 pages (list of references 1) of 32 volume numberings.In this piece article, described noise suppressor uses usually said least mean-square error fast frequency spectrum amplitude treatment technology.Fig. 1 shows the noise suppressor structure that is described in the list of references 1.Comprise the signal that needed voice signal mixes with noise and hereinafter will be referred to as noisy voice signal.

The noise suppressor that Fig. 1 shows comprises input end 11, frame resolving cell 1, windows units 2, Fourier transform unit 3, voice activity detector 4, noise estimation unit 51, frequency-relevant signal to noise ratio (snr) counter 6, priori signal-to-noise ratio (snr) estimation device 7, spectrum gain generator 8, inverse Fourier transform unit 9, frame comprehensive unit 10, output terminal 12, counter 13 and multiplexed multiplier 16,17.In this noise suppressor, input end 11 provides the sequence of noisy voice signal as sampling.Then, the sampled value of noisy voice signal offers frame resolving cell 1, and it is separated into the frame with K/2 sampling with noisy voice signal, and wherein K represents even number.The sampling that the noise voice signal is arranged of separated framing offers windows units 2, and they are multiplied by window function w (t).By utilizing w (t) window operation input signal y _n(t) (t=0,1 ..., k/2-1) the n frame and the signal y that produces _n(t) represent with following equation:

y _n(t)＝w(t)y _n(t) (1)

In this noise suppressor, usually overlapping and two frames that window operation is continuous.Be used as overlap length if suppose 50% of frame length, then the y of windows units 2 outputs _n(t) (t=0,1 ..., K-1), with equation (2), (3) expression:

y _n(t)＝w(t)y _n-1(t) (2)

y _n(t+K/2)＝w(t+K/2)y _n(t) (3)

In the following description, suppose that 50% is overlapping.Hamming window is represented with equation (4), for example, can be used w (t):

The y of window output _n(t) offer Fourier transform unit 3, it exports y with window _n(t) convert noisy voice spectrum Y to _n(k).Noisy voice spectrum Y _n(k) be divided into a phase place and an amplitude.Noisy speech phase spectrum arg Y _n(k) offer inverse Fourier transform unit 9, noisy voice spectrum amplitude | Y _n(k) | offer voice activity detector 4, multiplexed multiplier 16 and multiplexed multiplier 17.

Voice activity detector 4 is according to noisy speech | Y _n(k) | spectrum amplitude determines whether speech exists, and a basis is determined that the voice activation of setting as a result detects sign and is transferred to noise estimation unit 51.Multiplexed multiplier 17 utilize provide the noise speech arranged | Y _n(k) | spectrum amplitude calculates noisy speech power spectrum, and the noise speech power spectrum of calculating is offered noise estimation unit 51 and frequency-relevant snr computation device 6.

Noise estimation unit 51 utilizes voice activation to detect sign, has the count value that the noise speech power is composed, sum counter 13 provides to come the estimated noise power spectrum, and the power spectrum of estimation is transferred to the noise power spectrum of frequency-relevant snr computation device 6 as estimation.It is each frequency computation part signal to noise ratio (S/N ratio) that frequency-relevant snr computation device 6 has noise speech power spectrum and already provided estimated noise power spectrum by utilization, and the signal to noise ratio (S/N ratio) that this calculating is provided as the posteriority signal to noise ratio (S/N ratio) to priori signal-to-noise ratio (snr) estimation device 7 and spectrum gain generator 8.

The posteriority signal to noise ratio (S/N ratio) that priori signal-to-noise ratio (snr) estimation device 7 utilization provides and estimate the priori signal to noise ratio (S/N ratio) from the spectrum gain that spectrum gain generator 8 provides, and the priori signal to noise ratio (S/N ratio) of estimation is offered spectrum gain generator 8 as feedback.

Spectrum gain generator 8 utilizes the posteriority signal to noise ratio (S/N ratio) and produces a spectrum gain as the estimation priori signal to noise ratio (S/N ratio) that input provides, and spectrum gain is offered priori signal-to-noise ratio (snr) estimation device 7 as feedback, but also the noise spectrum gain that generates is transferred to multiplexed multiplier 16.

The spectrum gain that multiplexed multiplier 16 utilizes spectrum gain generator 8 to provide

Weighting Fourier transform unit 3 provides the noise speech | Y _n(k) | spectrum amplitude, thus determine to strengthen speech | X _n(k) | spectrum amplitude, and will strengthen speech | X _n(k) | spectrum amplitude is transferred to inverse Fourier transform unit 9.The speech that strengthens | X _n(k) | represent with equation (5):

$\left|{\stackrel{\‾}{X}}_n\left(k\right)\right|={\stackrel{\‾}{G}}_n\left(k\right)\left|Y_n\left(k\right)\right|----\left(5\right)$

The enhancing speech that inverse Fourier transform unit 9 provides multiplexed multiplier 16 | X _n(k) | spectrum amplitude and Fourier transform unit 3 provide noise speech phase spectrum arg Y _n(k) multiply each other each other, therefore determine the speech X that strengthens _n(k).That is, inverse Fourier transform unit 9 is carried out according to equation (6) and is calculated:

X _n(k)＝| X _n(k)|arg?Y _n(k) (6)

The enhancing speech X that inverse Fourier transform unit 9 is producing _n(k) go up the execution inverse Fourier transform, produce the sample sequence x of a time domain _n(t) (t=0,1 ..., K-1), wherein a frame is made up of K sampled value, and transmits this time-domain sampling x _n(t) give frame comprehensive unit 10.Frame comprehensive unit 10 is from x _n(t) two adjacent frames take out k/2 sampled value, and this k/2 sampled value that superposes, and produce according to equation (7) to strengthen speech The enhancing speech that produces

(t=0,1 ..., K-1) be sent to output terminal 12 as the output of frame comprehensive unit 10.

${\hat{x}}_n\left(t\right)={\stackrel{-}{x}}_{n-1}(t+K/2)+{\stackrel{\‾}{x}}_n\left(t\right)----\left(7\right)$

The at length not open voice activity detector of how realizing being included in the noise suppressor shown in Figure 14 of list of references 1.Yet an example that can be used in the voice activity detector in the noise suppressor can obtain from " proceedings of the national meeting of Japanese acoustical society, in March, 2000,321-322 page or leaf (list of references 2) ".Voice activity detector shown in the list of references 2 will be described below as conventional embodiment of a voice activity detector 4.As shown in Figure 2, voice activity detector 4 comprises threshold value memory 401, comparer 402, multiplier 404, logarithmic calculator 405, power calculator 406, weighted summer 407, weighting storer 408 and not circuit 409.

In voice activity detector 4, the spectrum amplitude that the noise speech is arranged that Fourier transform unit 3 (Fig. 1) provides offers power calculator 406.Power calculator 406 calculates the spectrum amplitude power of noise speech from k=0 to k-1 | Y _n(k) | ²Sum, and will calculate and send logarithmic calculator 405 to.Logarithmic calculator 405 is provided by the logarithm that noise voice spectrum power is arranged that provides, and provides this logarithm to multiplier 404.Constant of multiplier 404 usefulness multiply by the logarithm that provides so that define noise speech power Q _n, and provide this noise speech power to be arranged to comparer 402 and weighted summer 407.Particularly, in the n frame noise speech power Q arranged _nBe to represent with following equation:

$Q_n=10{\log }_{10}(\underset{t=0}{\overset{K-1}{\mathrm{\Σ}}}{\left|Y_n\left(k\right)\right|}^2----\left(8\right)$

Disclosed voice activity detector utilizes time-domain sampling y according to equation (9) in list of references 2 _n(t) determine Q _n

$Q_n=10{\log }_{10}(\underset{t=0}{\overset{K-1}{\mathrm{\Σ}}}{\left|{\stackrel{\‾}{y}}_n^2\left(t\right)\right|}^2----\left(9\right)$

As Corona, in 1985 " digital signal processing " delivered (list of references 3), the 75-76 page or leaf is described, and well-known equation (8) and (9) are equivalent according to Parseval theorem (Parseval).

Threshold value memory 401 provides a threshold value TH _nGive comparer 402.Comparer 402 is with the output and the threshold value TH of multiplier 404 _nCompare.If TH _n＞Q _n, then the speech parts are represented in comparer 402 outputs " 1 ", and if TH _n≤ Q _n, then noiseless part is represented in comparer 402 outputs " 0 ", detects sign as a voice activation.The output of comparer 402 is used as the sign that voice activation detects, and also offers not circuit 409.What not circuit 409 provided it exports to the control signal 905 of weighted summer 407 as weighted summer.Weighted summer 407 is provided threshold value 902 and is provided weighting 903 by weighting storer 408 by threshold value memory 401.

Weighted summer 407 is upgraded the threshold value 902 that provides from threshold value memory 401 selectively according to weighted summer control signal 905, and will upgrade threshold value 904 and offer threshold value memory 401 as feedback.Be used to weighting 903, by carrying out threshold value TH from weighting storer 408 _N-1Determine the threshold value TH that upgrades with the weighting summation that noise speech power 901 is arranged _nThe threshold value TH that upgrades _nOnly when the weighted summer control signal 905 of not circuit 409 outputs equals " 1 ", just calculate, that is, only during noiseless part, calculate.Therefore, the threshold value 904 of renewal offers threshold value memory 401 as feedback.

As shown in Figure 3, power calculator 406 has demultiplexer 4061, a K multiplier 4062 ₀To 4062 _K-1With totalizer 4063.The multiplexed spectrum amplitude that the noise speech is arranged that Fourier transform unit 3 (Fig. 1) provides is separated into frequency-K relevant sampling by demultiplexer 4061, and it offers multiplier 4062 respectively ₀To 4062 _K-1 Multiplier 4062 ₀To 4062 _K-1Square input signal that provides respectively, and will square after signal send totalizer 4063 to, this determine the summation of input signal and will calculate and export.

As shown in Figure 4, weighted summer 407 has multiplier 4071,4073, constant multiplier 4075 and totalizer 4072,4074.From multiplier 404 (Fig. 2) noise speech power 901 arranged, from the threshold value 902 of threshold value memory 401 (Fig. 2), all offer weighted summer 407 as input from the weighting 903 of weighting storer 408 (Fig. 2) and from the weighted summer control signal 905 of not circuit 409 (Fig. 2).Weighting 903 with value β is transferred to constant multiplier 4075 and multiplier 4073.Constant multiplier 4075 usefulness-1 are multiplied by input signal and produce a value-β, and should send totalizer 4074 to by value-β, also provide 1 as another input.Therefore, totalizer 4074 output and 1-β should and offer multiplier 4071 again.On the other hand, multiplier 4071 usefulness have noise speech power Q _nMultiply by summation 1-β and import, produce product (1-β) Q who is transferred to totalizer 4072 as another _nThe value β that provides as weighting 903 is provided multiplier 4073 usefulness threshold values 902, and transmits product β TH _N-1Give totalizer 4072.Totalizer 4072 is with β TH _N-1With (1-β) Q _nAddition, and output is somebody's turn to do and threshold value 904 is upgraded in conduct.The threshold value TH that upgrades _nOnly when equaling " 1 ", weighted summer control signal 905 just calculates.That is, weighted summer 407 has renewal TH _N-1Function so that determine TH during weighting 903 numerical value that β represents in according to following equation _n:

${\mathrm{TH}}_n=\left\{\begin{array}{c}{\mathrm{TH}}_n,{\mathrm{TH}}_n\&GreaterEqual;Q_n\\ \mathrm{\&beta;}{\mathrm{TH}}_{n-1}+(1-\mathrm{\&beta;})Q_n,{\mathrm{TH}}_n<Q_n\end{array}\right.----\left(10\right)$

Fig. 5 shows the configuration example that is included in the multiplexed multiplier 17 in the noise suppressor shown in Figure 1.Multiplexed multiplier 17 has K multiplier 1701 ₀To 1701 _K-1, demultiplexer 1702,1703 and multiplexer 1704.In multiplexed multiplier 17, the multiplexed spectrum amplitude that the noise speech is arranged that Fourier transform unit 3 (Fig. 1) provides is separated into frequency-K relevant sampling by demultiplexer 1702,1703, and they offer multiplier 1701 respectively ₀To 1701 _K-1Multiplier 1701 ₀To 1701 _K-1Square input signal that provides respectively, and will square signal be transferred to multiplexer 1704, its multiplexed these input signal and export this multiplex signal as noise speech power spectrum is arranged.

As shown in Figure 6, the noise estimation unit 51 that is included in the noise suppressor shown in Figure 1 has demultiplexer 502, multiplexer 503 and K frequency-correlation noise evaluation unit 514 ₀To 514 _K-1In noise judging unit 51, the count value that the voice activation that voice activity detector 4 (Fig. 1) provides detects sign sum counter 13 (Fig. 1) to be provided all is transferred to frequency-correlation noise evaluation unit 514 ₀To 514 _K-1The noise speech power spectrum that has that multiplexed multiplier 17 (Fig. 1) provides is transferred to demultiplexer 502.Demultiplexer 502 will provide multiplexedly has noise speech power spectrum to be separated into K frequency-correlated components, and K frequency-correlated components is transferred to frequency-correlation noise evaluation unit 514 respectively ₀To 514 _K-1Frequency-correlation noise evaluation unit 514 ₀To 514 _K-1Utilization is calculated these noise power spectral components from the noise speech power spectrum that has that demultiplexer 502 provides, and the noise power spectral component that calculates is transferred to multiplexed 503.The calculating of noise power spectrum is to detect the value that indicates by count value and voice activation to control, and only just carries out when predetermined state satisfies.The multiplexer 503 multiplexed K that provides a noise power spectral component, and multiplexed noise power spectrum is used as the estimated noise power spectrum exports.

Fig. 7 shows the frequency-correlation noise evaluation unit 514 that is included in the noise judging unit 51 (Fig. 6) ₀To 514 _K-1Structure.Because frequency-correlation noise evaluation unit 514 ₀To 514 _K-1Structurally be identical each other, so they just are expressed as frequency-correlation noise evaluation unit 514 in Fig. 7.Disclosed noise estimation algorithm is used for upgrading the estimated noise value in the noiseless part in list of references 2, and uses by the average estimated noise instantaneous value of regressive filter as the estimated noise value.Another noise estimation algorithm is to publish in May, 1998, the collection of thesis of IEEE speech and Audio Processing, volume 6 numberings 3,287-292 page or leaf (list of references 4), the estimated noise instantaneous value that its description is averaged and uses.List of references 4 suggestions utilize transversal filter to realize average treatment, that is, wave filter comprises shift register, rather than regressive filter.Because the two realizes having equal function, so disclosed processing will be described below in the list of references 4.

Frequency-correlation noise evaluation unit 514 has the judging unit 521 of renewal, register length storer 5041, switch 5044, shift register 4045, totalizer 5046, minimum value selector 5047, separation vessel 5048, sum counter 5049.There is noise speech power spectrum to offer switch 5044 from the frequency of demultiplexer 502 (Fig. 6)-relevant.When its circuit was closed in switch 5044 passes, frequency-being correlated with had noise speech power spectrum to be transferred to shift register 5045.The control signal of judging unit 521 is upgraded in response, and shift register 5045 moves on to adjacent register cell with the storage numerical value in the internal register unit.The length of shift register 5045 equals to be stored in the numerical value in the register length storer 5941.Output from the totality register cell of shift register 5045 offers totalizer 5046.The output addition that totalizer 5046 will provide from all internal register unit, and with it be transferred to divider 5048.

On the other hand, from the count value of counter 13 with detect sign from the voice activation of voice activity detector 4 and offer and upgrade judging unit 521.Renewal judging unit 521 is always exported " 1 " and is reached a predetermined value up to count value.After count value reaches this predetermined value, when voice activation detection sign is " 0 ", promptly during noiseless part, upgrade judging unit 521 outputs " 1 ", otherwise output " 0 ".Renewal judging unit 521 is transferred to counter 5049, switch 5044 and shift register 5045 with its output.When upgrading judging unit 521 and provide signal to be " 1 ", closed its circuit of switch 5044, and the signal that provides when renewal judging unit 521 disconnects its circuit when being " 0 ".When renewal judging unit 521 provided signal to be " 1 ", counter 5049 increased its count values, and when the signal that provides from renewal judging unit 521 is " 0 ", does not change its count value.When the signal that upgrades judging unit 521 and provide is " 1 ", shift register 5045 from switch 5044 read a sampled signal and with the storage numeric shift of internal register unit to adjacent register cell.

The output of the output of counter 5049 and register length storer 5941 offers minimum value selector 5047.Minimum value selector 5047 is selected less in the count value that provides and the register length one, and selective value is transferred to divider 5048.In divider 5048 usefulness count values and the register length less one remove the frequency that totalizer 5046 provides-relevant have noise speech power spectrum and, and export this and discuss as frequency-correlation estimation noise power spectrum λ _n(k).If be stored in the frequency-relevant sampled value B that noise speech power spectrum component is arranged of shift register 5045 _n(k) (n=0,1 ..., n-1) expression, this frequency-correlation estimation noise power spectrum λ then _n(k) with equation (11) expression:

${\mathrm{\&lambda;}}_n\left(k\right)=\frac{1}{N}\underset{n=0}{\overset{N-1}{\mathrm{\&Sigma;}}}{B}_n\left(k\right)----\left(11\right)$

Wherein N represents less in count value and the register length one.The increase because count value is started from scratch monotonously is so to begin be to utilize count value to carry out and then utilize register length to carry out to divide operations.Utilize register length to carry out and remove the mean value that operation means storage numerical value in definite shift register.Beginning because do not have abundant value storage in shift register 5045, so with frequency-relevant have noise speech power spectrum and divided by the quantity of the register cell of actual storage numerical value.When count value during less than register length, the register cell quantity of actual storage numerical value equals count value, and when count value became bigger than register length, it equaled register length.

The structure of Fig. 8 display update judging unit 521.Upgrade judging unit 521 and have not circuit 5202, comparer 5203, threshold value memory 5204 and OR circuit 5211.In upgrading judging unit 521, the count value that counter 13 (Fig. 1) provides is transferred to comparer 5203.The threshold value of threshold value memory 5204 outputs also offers comparer 5203.Comparer 5203 compares count value that provides and the threshold value that provides mutually.If count value is less than threshold value, then OR circuit 5211 is given in comparer 5203 transmission " 1 ", and if count value greater than threshold value, then OR circuit 5211 is given in comparer 5203 transmission " 0 ".Offer the voice activation that upgrades judging unit 521 and detect sign and be transferred to not circuit 5202, it determine the logic reversal value of input signal and oppositely value be transferred to OR circuit 5211.Specifically, not circuit 5202 is transferred to voice activation with " 0 " and detects OR circuit 5211 in the speech part that sign is " 1 ", detects to indicate it is OR circuit 5211 in the noiseless part of " 0 " and will " 1 " be transferred to voice activation.As a result, during voice activation detects the noiseless part that sign is " 0 " or when count value OR circuit 5211 outputs " 1 " during, the switch closure and the count-up counter 5049 of Fig. 7 demonstration less than threshold value.

Fig. 9 shows the configuration example of the frequency-relevant snr computation device 6 that is included in the noise suppressor shown in Figure 1.Frequency-relevant snr computation device 6 has K divider, 601 ₀To 601 _K-1, demultiplexer 602,603 and multiplexer 604.In frequency-relevant snr computation device 6, the noise speech power spectrum that has that multiplexed multiplier 17 (Fig. 1) provides is transferred to demultiplexer 602.The estimated noise power spectrum that noise estimation unit 51 (Fig. 1) provides is transferred to demultiplexer 603.Have noise speech power spectrum to be divided into K sampling, and this K sampled and offer each divider 601 corresponding to each frequency component by demultiplexer 602 ₀To 601 _K-1The noise power spectrum of estimation is divided into K sampling corresponding to each frequency component by demultiplexer 603, and this K sampled and offer each divider 601 ₀To 601 _K-1Divider 601 ₀To 601 _K-1Remove the noise speech power that has that provides with the estimated noise power spectrum that provides and compose, thereby determine frequency-relevant signal to noise ratio (S/N ratio) γ according to equation (12) _n(k), and with frequency-relevant signal to noise ratio (S/N ratio) γ _n(k) be transferred to multiplexer 604:

${\mathrm{\&gamma;}}_n\left(k\right)=\frac{{\left|Y_n\left(k\right)\right|}^2}{{\mathrm{\&lambda;}}_n\left(k\right)}----\left(12\right)$

λ wherein _n(k) noise power spectrum of expression estimation.K frequency of multiplexer 604 multiplexed transmission-relevant signal to noise ratio (S/N ratio), and export this multiplexed signal to noise ratio (S/N ratio) as the posteriority signal to noise ratio (S/N ratio).

As shown in figure 10, the priori signal-to-noise ratio (snr) estimation device 7 that is included in the noise suppressor shown in Figure 1 has multiplexed scope limiting processor 701, posteriority signal to noise ratio (S/N ratio) storer 702, spectrum gain storer 703, multiplexed multiplier 704,705, weighting storer 706, multiplexed weighted summer 707 and totalizer 708.

In priori signal-to-noise ratio (snr) estimation device 7, the posteriority signal to noise ratio (S/N ratio) γ that frequency-relevant snr computation device 6 (Fig. 6) provides _n(k) (k=0,1 ..., K-1) be transferred to posteriority signal to noise ratio (S/N ratio) storer 702 and totalizer 708.Posteriority signal to noise ratio (S/N ratio) γ in the posteriority signal to noise ratio (S/N ratio) storer 702 storage n frames _n(k) and transmit posteriority signal to noise ratio (S/N ratio) γ in (n-1) frame _N-1(k) to multiplexed multiplier 705.The spectrum gain that spectrum gain generator 8 provides (k=0,1 ..., K-1) be transferred to spectrum gain storer 703.Spectrum gain in the spectrum gain storer 703 storage n frames

And with the spectrum gain G in (n-1) individual frame _N-1(k) be transferred to multiplexed multiplier 704.704 squares of spectrum gain G that provide of multiplexed multiplier _N-1(k) to determine G ² _N-1(k) and the transmission G ² _N-1(k) give multiplexed multiplier 705.Multiplexed multiplier 705 is with G ² _N-1(k) and γ _N-1(k) (k=0,1 ..., k-1) multiply each other so that determine G ² _N-1(k) γ _N-1(k), and with G ² _N-1(k) γ _N-1(k) send multiplexed weighted summer 707 to as estimation back signal to noise ratio (S/N ratio) 922.Multiplexed multiplier 704,705 is structurally identical with the multiplexed multiplier of having described with reference to figure 5 17, therefore will no longer describe here.

An other end of totalizer 708 is provided-1, so and γ _n(k)-1 be transferred to multiplexed scope limiting processor 701.Multiplexed scope limiting processor 701 utilizes scope restriction operator P[.] processing totalizer 708 provide and γ _n(k)-1, so transmit this P[γ as a result _n(k)-1] give multiplexed weighted summer 707 as instantaneous estimation signal to noise ratio (S/N ratio) 921.P[x] definition as (13):

Multiplexed weighted summer 707 also provides the weighting 923 from weighting storer 706.Multiplexed weighted summer 707 is utilized in this instantaneous estimation signal to noise ratio (S/N ratio) 921 that provides, and the priori signal to noise ratio (S/N ratio) 924 of estimation is determined in estimation back signal to noise ratio (S/N ratio) 922 and weighting 923.Priori signal to noise ratio (S/N ratio) 924 usefulness of estimating if weighting 923 usefulness α represent Expression, then

According to equation 14) calculate:

${\widehat{\mathrm{\&xi;}}}_n\left(k\right)=\mathrm{\&alpha;}{\mathrm{\&gamma;}}_{n-1}\left(k\right){\stackrel{\&OverBar;}{G}}_{n-1}^2\left(k\right)+(1-\mathrm{\&alpha;})P[{\mathrm{\&gamma;}}_n\left(k\right)-1]----\left(14\right)$

Wherein ${\stackrel{\&OverBar;}{G}}_{-1}^2\left(k\right){\mathrm{\&gamma;}}_{-1}\left(k\right)=1$

As shown in figure 11, above-described multiplexed scope limiting processor 701 has constant storage 7011, a K maximum-value selector 7012 ₀To 7012 _K-1, demultiplexer 7013 and multiplexer 7014.In multiplexed scope limiting processor 701, demultiplexer 7013 provides the γ from totalizer 708 (Figure 10) _n(k)-1.The γ that provides is provided multiplexed separation vessel 7013 _n(k)-1 to K frequency-correlated components and provide these frequency-correlated components respectively to maximum-value selector 7012 ₀To 7012 _K-1, their another one input end provides from zero of constant storage 7011.Maximum-value selector 7012 ₀To 7012 _K-1With γ _n(k)-1 compare with zero, and transmit bigger value and give multiplexer 7014.This maximal value is selected to calculate corresponding to the calculating according to equation (13).Multiplexer 7014 multiplexed numerical value that these provide and export these multiplexed values.

As shown in figure 12, multiplexed weighted summer 707 has K weighted summer 7071 ₀To 7071 _K-1, demultiplexer 7072,7074 and multiplexer 7075.In multiplexed weighted summer 707, from the P[γ of multiplexed scope limiting processor 701 (Figure 10) _n(k)-1] offer demultiplexer 7072 as instantaneous estimation signal to noise ratio (S/N ratio) 921.Multiplexed separation vessel 7072 separates P[γ _n(k)-1] to K frequency-correlated components, and with the signal to noise ratio (S/N ratio) 921 of these frequency-correlated components as frequency-relevant instantaneous estimation ₀To 921 _K-1Send each weighted summer 7071 to ₀To 7071 _K-1From multiplexed multiplier 705 Figure 10) as the G that estimates back signal to noise ratio (S/N ratio) 922 ² _N-1(k) γ _N-1(k) offer demultiplexer 7074.Demultiplexer 7074 separates G ² _N-1(k) γ _N-1(k) be K frequency-correlated components, and these frequency-correlated components are estimated signal to noise ratio (S/N ratio) 921 as the past frequency dependence ₀To 921 _K-1Be transferred to each weighted summer 7071 ₀To 7071 _K-1Weighted summer 7071 ₀To 7071 _K-1 Weighting 923 also is provided.Weighted summer 7071 ₀To 7071 _K-1Carry out weighted addition according to equation (14), and with the result as frequency-correlation estimation priori signal to noise ratio (S/N ratio) 924 ₀To 924 _K-1Be transferred to multiplexer 7075.Multiplexer 7075 multiplexed frequencies-correlation estimation priori signal to noise ratio (S/N ratio) 924 ₀To 924 _K-1, and export the priori signal to noise ratio (S/N ratio) 924 of multiplexed result as estimation.Each weighted summer 7071 ₀To 7071 _K-1Operation all the weighted summer 407 with reference to figure 4 with described above is identical with structure, so be not described in detail.Yet this weighted addition is always being calculated.

Figure 13 shows the configuration example that is included in the spectrum gain generator 8 in the noise suppressor shown in Figure 1.Spectrum gain generator 8 has K spectrum gain sniffer 801 ₀To 801 _K-1, demultiplexer 802,803 and multiplexer 804.In spectrum gain generator 8, offer demultiplexer 802 from the posteriority signal to noise ratio (S/N ratio) of frequency-relevant snr computation device 6 (Fig. 1).Multiplexed separation vessel 802 the posteriority signal to noise ratio (S/N ratio) that provides is provided becomes K frequency-correlated components and these K frequency-correlated components is sent to spectrum gain sniffer 801 respectively ₀To 801 _K-1Estimation priori signal to noise ratio (S/N ratio) from priori signal-to-noise ratio (snr) estimation device 7 (Fig. 1) offers demultiplexer 803.Demultiplexer 803 the estimation priori signal to noise ratio (S/N ratio) that provides is provided becomes K frequency-correlated components and transmit these K frequency-correlated components respectively to spectrum gain sniffer 801 ₀To 801 _K-1Spectrum gain sniffer 801 ₀To 801 _K-1Search is corresponding to the spectrum gain of posteriority signal to noise ratio (S/N ratio) and already provided estimation priori signal to noise ratio (S/N ratio), and sends this result to multiplexer 804.Multiplexer 804 multiplexed spectrum gain that provide and export this multiplexed result.

Figure 14 shows spectrum gain sniffer 801 ₀To 801 _K-1Example of structure.Because spectrum gain sniffer 801 ₀To 801 _K-1Structurally be identical each other, so they are expressed as Figure 14 intermediate frequency spectrum gain sniffer 801.Spectrum gain sniffer 801 has spectrum gain table 8011 and address converter 8012,8013.In spectrum gain sniffer 801, offer address converter 8012 from the frequency of demultiplexer 802 (Figure 13)-relevant posteriority signal to noise ratio (S/N ratio).Address converter 8012 frequency-relevant posteriority the signal to noise ratio (S/N ratio) that provides is provided corresponding address, and transmits this address to spectrum gain table 8011.The frequency of demultiplexer 803 (Figure 13)-correlation estimation priori signal to noise ratio (S/N ratio) offers address converter 8013.Address converter 8013 is converted into corresponding address with the frequency-correlation estimation priori signal to noise ratio (S/N ratio) that provides, and transmits this address to spectrum gain table 8011.The spectrum gain of the region, address that provides corresponding to address converter 8012 and address converter 8013 is provided for spectrum gain table 8011 output, as the frequency dependence spectrum gain.

Conventional noise suppressor is described above.About above-described conventional noise suppressor, noise power spectrum is to upgrade in noiseless part according to the output of voice activity detector.Therefore, if incorrect from the testing result of voice activity detector, the power spectrum of estimated noise exactly then.When the speech part continued for a long time, because there is not noiseless part to exist, so noise power spectrum can not upgrade, the estimation precision of astable noise power spectrum reduced inevitably thus.Therefore, Chang Gui noise suppressor has residual noise and on the enhancing speech distortion is arranged.

According to the inhibition algorithm of routine, noise power spectrum is to utilize noisy speech power to compose to estimate.Therefore, for conventional algorithm, can not estimate accurately that the distortion trend is introduced into the speech of enhancing so the noise trend keeps under the speech power spectrum influence of noise power spectrum in including the noise speech.In addition, according to the noise suppression algorithm of routine, because squelch utilizes spectrum gain to finish, this spectrum gain is determined by the identical calculations method that is independent of signal to noise ratio (S/N ratio), so can not obtain sufficiently high quality for the speech that strengthens.

Summary of the invention

The purpose of this invention is to provide a kind of noise suppressing method, by with the irrelevant accurately estimated noise power spectrum of voice activity detector performance have the enhancing speech that reduces distortion and noise so that produce.

Another object of the present invention provides a kind of equipment that is used for squelch, composes by the accurately estimated noise power that is not subjected to the voice activity detector Properties Control, has the enhancing speech that reduces distortion and noise so that produce.

A further object of the invention provides a kind of noise suppressing method, even in the noise right and wrong stably during state, and this method is by estimated noise power spectrum accurately, so that produce the enhancing speech that reduces distortion and noise.

Also have, another object of the present invention provides a kind of equipment that is used for squelch, even when noise was non-plateau, the noise power spectrum of this device by accurately estimating had the enhancing speech that reduces distortion and noise so that produce.

Another object of the present invention provides a kind of noise suppressing method, has the enhancing speech that reduces distortion and noise so that utilize optimal spectrum with respect to all snr values to gain to produce.

A further object of the invention provides a kind of equipment that is used for squelch, has the enhancing speech that reduces distortion and noise so that utilize optimal spectrum with respect to all snr values to gain to produce.

According to a first aspect of the invention, a kind of method of squelch is provided, it comprises step: input signal is transformed into frequency domain and determines signal to noise ratio (S/N ratio) according to the signal of frequency domain, determine spectrum gain according to signal to noise ratio (S/N ratio), according to signal to noise ratio (S/N ratio) correction spectrum gain so that produce the spectrum gain of a correction, the spectrum gain that utilization is revised comes the signal of weighting frequency domain with definite second frequency-region signal, and converts second frequency-region signal to time-domain signal, so that produce the repressed output signal of noise.

According to a second aspect of the invention, provide a kind of equipment of squelch, it comprises: the snr computation device, and it is used for input signal being transformed into frequency domain and utilizing frequency-region signal to determine signal to noise ratio (S/N ratio); The spectrum gain generator, it is used for recently determining spectrum gain according to noise; The spectrum gain amending unit, it is used for according to signal to noise ratio (S/N ratio) correction spectrum gain so that produce the spectrum gain of a correction; Multiplier, it utilizes the spectrum gain of revising to come the weighting frequency-region signal so that determine second frequency-region signal; Inverter, it is used for converting second frequency-region signal to time-domain signal.

In the equipment and method of above-mentioned squelch, noise is to utilize the spectrum gain of the value correction that relies on signal to noise ratio (snr) to suppress.Specifically, the equipment for squelch has the spectrum gain amending unit, the spectrum gain that the numerical value of its received signal to noise ratio and spectrum gain and calculating are revised.Prevent noise by the spectrum gain that relies on snr value utilization correction, may obtain to suffer small distortion and whole relatively snr value to suffer the enhancing speech of less noise according to the present invention.

According to a third aspect of the present invention, a kind of method of squelch is provided, it comprises step: input signal is transformed into frequency domain and weighting frequency-region signal so that determine the weighted frequency-domain signal, utilize weighted frequency-domain signal estimation noise, utilize the noise and the frequency-region signal of estimation to determine signal to noise ratio (S/N ratio), determine spectrum gain according to signal to noise ratio (S/N ratio), utilize spectrum gain to come the weighting frequency-region signal, and change this second frequency-region signal and arrive the time domain signal so that produce the repressed output signal of noise with definite second frequency-region signal.

According to a fourth aspect of the present invention, provide a kind of equipment of squelch, comprise at least: the snr computation device, it is used for input signal being transformed into frequency domain and utilizing frequency-region signal to determine signal to noise ratio (S/N ratio); The spectrum gain generator, it is used for recently determining spectrum gain according to this noise; Multiplier, it utilizes spectrum gain to come the weighting frequency-region signal to determine second frequency-region signal; And inverter, it is used to change second frequency-region signal to the time domain signal, and wherein the snr computation device comprises: a weighting frequency-region signal counter, it is used for the weighting frequency-region signal so that determine the weighting frequency-region signal; The noise judging unit, it utilizes the weighting frequency-region signal to come estimated noise.

In the method and apparatus of above-mentioned squelch, the power spectrum of noise is to utilize weighting frequency-region signal (that is the noisy speech power spectrum of weighting) estimation.Especially, have the counter (that is, weighting has the counter of noise speech) of weighting frequency-region signal for the equipment of squelch, it calculates to have by oneself the weighted noise speech power spectrum of noise speech power spectrum and the noise power spectrum of estimation.Because noise power spectrum is to utilize the noise power spectrum of estimating in weighted noise speech power spectrum and the previous frame to estimate in the present frame, this weighted noise speech power composes self noise speech power spectrum, so the power spectrum of possible accuracy ground estimated noise and irrelevant, thereby produce the enhancing speech that suffers small distortion and noise with the character of noise.

According to a fifth aspect of the present invention, a kind of method of estimated noise is provided, it comprises step: utilize input signal and determine signal to noise ratio (S/N ratio) at the noise of determining to add the estimation that temporary exists, utilize signal to noise ratio (S/N ratio) to determine weighting, utilize weight weighting input signal so that determine the input signal of weighting, and the noise of determining estimation according to the input signal of weighting.

According to a sixth aspect of the invention, provide a kind of equipment of estimated noise, it comprises: the snr computation device, and it utilizes input signal and determines signal to noise ratio (S/N ratio) at the noise of determining to add the estimation that temporary exists; Weight calculator, it is used for determining weighting according to this signal to noise ratio (S/N ratio); The input signal counter, it utilizes weight weighting input signal so that determine the input signal of weighting; And the noise estimation unit, it is used for determining according to the input signal of weighting the noise of estimation.

Description of drawings

Fig. 1 is the block diagram that shows conventional noise suppressor;

Fig. 2 shows the block diagram that is included in the voice activity detector in the noise suppressor shown in Figure 1;

Fig. 3 shows the block diagram that is included in the power calculator in the noise suppressor shown in Figure 2;

Fig. 4 shows the block diagram that is included in the weighted summer in the voice activity detector shown in Figure 2;

Fig. 5 is the block diagram that a demonstration is included in the multiplexed multiplier in the noise suppressor shown in Figure 1;

Fig. 6 shows the block diagram that is included in the noise estimation unit in the noise suppressor shown in Figure 1;

Fig. 7 shows the block diagram that is included in the frequency-correlation noise evaluation unit in the noise estimation shown in Figure 6 unit;

Fig. 8 is the block diagram that shows the renewal judging unit that is included in frequency shown in Figure 7-correlation noise evaluation unit;

Fig. 9 is the block diagram that shows the frequency-relevant snr computation device that is included in the noise suppressor shown in Figure 1;

Figure 10 shows the block diagram that is included in the frequency-priori signal-to-noise ratio (snr) estimation device in the noise suppressor shown in Figure 1;

Figure 11 shows the block diagram that is included in the multiplexed scope limiting processor in frequency shown in Figure 10-priori signal-to-noise ratio (snr) estimation device;

Figure 12 shows the block diagram that is included in the multiplexed weighted summer in frequency shown in Figure 10-priori signal-to-noise ratio (snr) estimation device;

Figure 13 shows the block diagram that is included in the spectrum gain generator in the noise suppressor shown in Figure 1;

Figure 14 shows the block diagram that is included in the spectrum gain test cell in the spectrum gain generator shown in Figure 13;

Figure 15 is the noise suppressor block diagram that shows according to first embodiment of the invention;

Figure 16 shows the block diagram that is included in the weighted noise speech counter in the noise suppressor shown in Figure 15;

Figure 17 shows the block diagram that is included in the multiplexed nonlinear processor in the weighted noise speech counter;

Figure 18 is the figure that shows the nonlinear function example that uses multiplexed nonlinear processor;

Figure 19 shows the block diagram that is included in the noise estimation unit in the noise suppressor shown in Figure 15;

Figure 20 shows the block diagram that is included in the frequency-correlation noise evaluation unit in the noise estimation shown in Figure 19 unit;

Figure 21 is the block diagram that shows the renewal judging unit that is included in frequency shown in Figure 20-correlation noise evaluation unit;

Figure 22 shows the block diagram that is included in the frequency-correlation noise evaluation unit in the noise estimation shown in Figure 19 unit;

Figure 23 shows the block diagram that is included in the spectrum gain amending unit in the noise suppressor shown in Figure 15;

Figure 24 is the block diagram that shows the frequency-relevant frequency spectrum gain amending unit that is included in the spectrum gain amending unit shown in Figure 23;

Figure 25 is the block scheme that shows second example of spectrum gain generator structure;

Figure 26 is the block diagram that shows the frequency-bandwidth-relevant snr computation device of the frequency-relevant snr computation device that can replace in the noise suppressor shown in Figure 15;

Figure 27 shows the block diagram that is included in the frequency-bandwidth-related power counter in frequency-bandwidth shown in Figure 26-relevant snr computation device;

Figure 28 is the noise suppressor block diagram that shows according to second embodiment of the invention;

Figure 29 shows the block diagram that is included in the noise estimation unit in the noise suppressor shown in Figure 28;

Figure 30 shows the block diagram that is included in the frequency-correlation noise evaluation unit in the noise estimation shown in Figure 29 unit;

Figure 31 is the noise suppressor block diagram that shows according to third embodiment of the invention;

Figure 32 shows the block diagram that is included in the frequency-priori signal-to-noise ratio (snr) estimation device in the noise suppressor shown in Figure 31;

Figure 33 is the noise suppressor block diagram that shows according to fourth embodiment of the invention;

Figure 34 is the noise suppressor block diagram that shows according to fifth embodiment of the invention;

Figure 35 shows the block diagram that is included in the noise estimation unit in the noise suppressor shown in Figure 34;

Figure 36 shows the block diagram that is included in the frequency-correlation noise evaluation unit in the noise estimation shown in Figure 35 unit; And

Figure 37 shows the block diagram that is included in the renewal judging unit in frequency shown in Figure 36-correlation noise evaluation unit.

Embodiment

The noise suppressor of the first embodiment of the invention that shows according to Figure 15 shows that with Fig. 1 the noise suppressor of routine is similar, but different is that the noise estimation unit has different inner structures, and has increased weighting noise speech counter 14 and spectrum gain amending unit 15 are arranged.Specifically, the noise estimation unit 51 of using noise estimation unit 5 to replace in the noise suppressor shown in Figure 1 according to the noise suppressor of first embodiment.Weighting has noise speech counter 14 to calculate the noise power spectrum that the weighting of having noise speech power spectrum by oneself has noise speech power spectrum and estimation, and the weighting that will calculate has noise speech power spectrum to output to noise estimation unit 5.Spectrum gain amending unit 15 is according to the spectrum gain of the priori snr computation correction of spectrum gain and estimation.Replace offering multiplexed multiplier 16 and priori signal-to-noise ratio (snr) estimation device 7 with the spectrum gain of revising by the spectrum gain that spectrum gain generator 8 produces.Voice activity detector 4, noise estimation unit 5, frequency-relevant snr computation device 6, counter 13, weighting have noise speech counter 14, and the multiplexed multiplier 17 common snr computation devices 101 of forming.Priori signal-to-noise ratio (snr) estimation device 7 and the spectrum gain generator 8 common spectrum gain generation units 102 of forming.

In the following description, those components of representing with reference character the same with shown in Fig. 1 to 14.According to the noise suppressor of present embodiment, mainly the difference with respect to conventional noise suppressor is described below.

As shown in figure 16, weighting has noise speech counter 14 to have estimated noise storer 1401, frequency-relevant snr computation device 1402, multiplexed nonlinear processor 1405 and multiplexed multiplier 1404.1401 storages of estimated noise storer are composed from the estimated noise power that noise estimation unit 5 (Figure 15) provides, and estimated noise power spectrum in the previous frame of storage is exported to frequency-relevant snr computation device 1402.Frequency-relevant snr computation device 1402, the noise speech power that has that the estimated noise power of utilizing estimated noise storer 1401 to provide is composed and multiplexed multiplier 17 (Figure 15) provides is composed, be each frequency computation part signal to noise ratio (S/N ratio), and the signal to noise ratio (S/N ratio) that output is calculated is to multiplexed nonlinear processor 1405.Multiplexed nonlinear processor 1405 utilizes from the noise that frequency-relevant snr computation device 1402 provides and recently calculates the weighting factor vector, and exports this weighting factor vector and give multiplexed multiplier 1404.Multiplexed multiplier 1404 calculating are composed the product of each frequency from the noise speech power that has that multiplexed multiplier 17 (Figure 16) provides, and the weighting factor vector that provides from multiplexed nonlinear processor 1405 is provided, and the output weighting has the noise speech power to compose to estimated noise storer 5 (Figure 15).Weighting has the amplitude component of noise speech power spectrum corresponding to weighting.

Have in the noise speech counter 14 in weighting, frequency-relevant snr computation device 1402 is structurally identical with above-described frequency with reference to figure 9-relevant snr computation device 6, and multiplexed multiplier 1404 is structurally with above-mentioned identical with reference to multiplexed multiplier among the figure 5 17.Therefore, will be not described in detail.

Being included in weighting has the structure detail of the multiplexed nonlinear processor 1405 in the noise speech counter 14 and work to be described in detail below with reference to Figure 17.As shown in figure 17, multiplexed nonlinear processor 1405 has demultiplexer 1495, a K nonlinear processor 1485 ₀To 1485 _K-1And multiplexer 1475.Demultiplexer 1495 separation become frequency-relevant signal to noise ratio (S/N ratio) from the signal to noise ratio (S/N ratio) that frequency-relevant snr computation device 1402 (Figure 16) provides, and difference output frequency-relevant signal to noise ratio (S/N ratio) is to nonlinear processor 1485 ₀To 1485 _K-1According to nonlinear function, nonlinear processor 1485 ₀To 1485 _K-1The actual value number of input value is depended in output.Figure 18 shows a nonlinear function example.As input value f ₁During expression, nonlinear function shown in Figure 180 has an output valve f with equation (15) expression ₂:

Each nonlinear processor 1485 ₀To 1485 _K-1Frequency-relevant signal to noise ratio (S/N ratio) of utilizing nonlinear function to handle to provide from demultiplexer 1495 is so that determine weighting factor, and exports this weighting factor and give multiplexer 1475.Specifically, nonlinear processor 1485 ₀To 1485 _K-1Rely on signal to noise ratio (S/N ratio) to export from 1 to 0 weighting factor scope, so they export 0 than hour output 1 in signal to noise ratio (S/N ratio) when signal to noise ratio (S/N ratio) is big.Multiplexer 1475 is multiplexed from nonlinear processor 1485 ₀To 1485 _K-1The weighting factor of output and output weighting factor vector are given multiplexed multiplier 1404.

With there being noise speech power spectrum to multiply each other by multiplexing multiplier 1404 in road (Figure 16) and weighting factor, it has the numerical value that depends on signal to noise ratio (S/N ratio).(, have the noise speech to comprise bigger voice components) weighting factor has smaller value when signal to noise ratio (S/N ratio) is big.The noise of estimation normally utilizes noise speech power spectrum to upgrade.There is noise speech power frequency spectrum usually to upgrade estimated noise by weighting, is used to estimate high-precision noise, can reduce to be included in the influence of voice components in the noise speech power frequency spectrum with signal to noise ratio (S/N ratio).When nonlinear function was used for calculating weighting factor in this example, the signal to noise ratio (S/N ratio) function rather than the nonlinear function that can use another kind of form to represent were such as linear function or height-order polynomial.

Figure 19 shows the structure that is included in the noise judging unit 5 in the noise suppressor.Noise judging unit 5 is similar to the noise judging unit 51 that is used for conventional noise suppressor shown in Figure 6, except it has demultiplexer 505, and frequency-correlation noise judging unit 514 ₀To 514 _K-1With frequency-correlation noise judging unit 504 ₀To 504 _K-1Replace.Noise judging unit 5 mainly is that these differences are described below relatively.

Noise speech power spectral decomposition will be provided from the weighting that weighting has noise speech counter 14 (Figure 15) to provide demultiplexer 505 is that frequency-related weighing has noise speech power frequency spectrum, and the noise speech power frequency spectrum of these frequency-related weighings is outputed to frequency-correlation noise judging unit 504 respectively ₀To 504 _K-1Frequency-correlation noise judging unit 504 ₀To 504 _K-1According to the frequency that provides from demultiplexer 502-being correlated with has noise speech power frequency spectrum, according to the frequency-associated weight that provides from demultiplexer 505 noise speech power frequency spectrum is arranged, detect sign according to the voice activation that provides from voice activity detector 4 (Figure 15), and come calculated rate-correlation estimation noise power spectrum, and the frequency-correlation estimation noise power spectrum that calculates is exported to multiplexer 503 according to the count value that provides from counter 13 (Figure 15).Multiplexer 503 multiplexed frequency-correlation noise judging units 504 ₀To 504 _K-1The frequency that provides-correlation estimation noise power spectrum, and export a synthetic estimated noise power and compose noise speech counter 14 (Figure 15) to frequency-relevant snr computation device 6 (Figure 15) and weight is arranged.Frequency-correlation noise judging unit 504 ₀To 504 _K-1Structure will be described below.

Figure 20 display frequency-correlation noise judging unit 504 ₀To 504 _K-1Structure.Because frequency-correlation noise judging unit 504 ₀To 504 _K-1Structurally be mutually the same, so they are expressed as the frequency-correlation noise judging unit 504 among Figure 20.Frequency-correlation noise judging unit 504 is different from frequency shown in Figure 7-correlation noise judging unit 514 as used herein, in frequency-correlation noise judging unit 504, has estimated noise storer 5942, upgrade judging unit 520 and replaced renewal judging unit 521, and offer switch 5044 be frequency-related weighing noise speech power frequency spectrum arranged, rather than frequency-relevant has noise speech power frequency spectrum.These differences be because frequency-correlation noise judging unit 504 ₀To 504 _K-1That has used weighting has a noise speech power frequency spectrum, rather than noise speech power frequency spectrum is arranged, and in calculating estimated noise, and uses the noise of estimation and noise speech power frequency spectrum is arranged in the renewal of determining estimated noise.Frequency-correlation estimation noise power spectrum that estimated noise storer 5942 storage separation vessels 5048 provide, and frequency-correlation estimation noise power spectrum of exporting storage in the mode of above-mentioned frame is to upgrading judging unit 520.

The structure of Figure 21 display update judging unit 520.Renewal judging unit 520 is different from the renewal judging unit 521 among Fig. 8, has comparer 5205 in upgrading judging unit 520, threshold value memory 5206 and threshold calculations device 5207, and OR circuit 5201 has been replaced OR circuit 5211.Upgrading judging unit 520 mainly is that these differences are described below relatively.

Threshold calculations device 5207 calculates a numerical value that depends on frequency-correlation estimation noise power spectrum that estimated noise storer 5942 (Figure 20) provides, and it is exported to threshold value memory 5206 as a threshold value.According to the simple process of calculated threshold, with multiplying each other of a constant and frequency-correlation estimation noise power spectrum as a threshold value.According to other processing, can utilize high-order moment or nonlinear function to come calculated threshold.Threshold value memory 5206 is stored from the threshold value of threshold calculations device 5207 outputs, and the threshold value of storage is outputed to comparer 5205 in above-mentioned frame mode.The frequency that the threshold value that comparer 5205 provides threshold value memory 5206 and demultiplexer 502 (Figure 19) provide-relevant has the noise voice spectrum to compare.If frequency-being correlated with has the noise voice spectrum less than threshold value, OR circuit 5201 is given in comparer 5205 outputs " 1 ".If frequency-being correlated with has the noise voice spectrum greater than threshold value, OR circuit 5200 is given in comparer 5205 outputs " 0 ".Therefore, comparer 5205 is determined noisy speech signal or noise according to the amplitude of estimated noise power spectrum.OR circuit 5201 calculate from the logic of comparer 5203 outputs, not circuit 5202 outputs and comparer 5205 outputs and, and export this result to switch 5044, shift register 5045, sum counter 5049 (Figure 20).

If there is the noise speech power not only less in original state and noiseless part, and also less in the speech part, then upgrade judging unit 520 outputs " 1 ", thereby upgrade estimated noise.Because threshold value is each frequency computation part, so estimated noise can be upgraded each frequency.

In Figure 20, suppose that counter 5049 has a count value CNT, shift register 5045 has the accumulator register length N and shift register 5045 storing frequencies-related weighing has noise speech power frequency spectrum B _n(k) (n=0,1 ..., N-1).The frequency that separation vessel 5048 provides-correlation estimation noise power spectrum λ _n(k) with equation (16) expression:

In other words, frequency-correlation estimation noise power spectrum λ _n(k) expression is stored in the mean value that frequency-associated weight in the shift register 5045 has noise speech power frequency spectrum.This mean value can utilize weighted summer (regressive filter) to calculate.Use the Structure Calculation frequency-correlation estimation noise power spectrum λ of weighted summer _n(k) be described below.

Figure 22 display frequency-correlation noise judging unit 504 ₀To 504 _K-1The structure of second example.Because frequency-correlation noise judging unit 504 ₀To 504 _K-1Structurally mutually the same, they are expressed as the frequency-correlation noise judging unit 507 among Figure 22.Frequency shown in Figure 22-correlation noise judging unit 507 has weighted summer 5071 and weighting storer 5072, add them to replace shift register 5045, totalizer 5046, minimum value selector 5047, separation vessel 5048, counter 5049 and the register length storer 5941 in frequency-correlation noise judging unit 504 shown in Figure 20.

The frequency dependence estimated noise power spectrum that weighted summer 5071 utilizes estimated noise storer 5942 to provide with above-mentioned frame form, frequency-associated weight of utilizing switch 5044 to provide has noise speech power frequency spectrum, come calculated rate-correlation estimation noise with the weighting factor that utilizes weighting storer 5072 output, and frequency-correlation estimation noise of exporting this calculating is given multiplexer 503.Specifically, frequency-associated weight has noise speech power frequency spectrum to use if be stored in that weighting factor in the weighting storer 5072 is represented with δ Expression, the then frequency-correlation estimation noise power spectrum λ that exports from weighted summer 5071 _n(k) with equation (17) expression.Because weighted summer 5071 is identical with above-described weighted summer 407 with reference to figure 4 structurally, so weighted summer 5071 will be not described in detail.Yet weighted addition is always calculated in weighted summer 5071.

${\mathrm{\&lambda;}}_n\left(k\right)=\mathrm{\&delta;}{\mathrm{\&lambda;}}_{n-1}\left(k\right)+(1-\mathrm{\&delta;}){\left|{\stackrel{\&OverBar;}{Y}}_n\left(k\right)\right|}^2----\left(17\right)$

Spectrum gain in the noise suppressor shown in Figure 15 is improved unit 15 and will be described below.Spectrum gain amending unit 15 is revised and is depended on the signal to noise ratio (S/N ratio) spectrum gain, so that prevent that residual noise is introduced in the inhibition owing to deficiency when signal to noise ratio (S/N ratio) is low, and prevent from when signal to noise ratio (S/N ratio) is higher, to reduce owing to excessive inhibition generation voice distortion causes speech quality.As an example of spectrum gain correction, when signal to noise ratio (S/N ratio) was low, modified value was affixed to spectrum gain suppressing residual noise, and during signal to noise ratio (S/N ratio), the minimum value of spectrum gain is limited so that prevent voice distortion.As shown in figure 23, spectrum gain amending unit 15 has K frequency-relevant frequency spectrum gain amending unit 1501 ₀To 1501 _K-1, demultiplexer 1502,1503 and multiplexer 1054.

Demultiplexer 1502 is separated into frequency-correlated components with the estimation priori signal to noise ratio (S/N ratio) that priori signal-to-noise ratio (snr) estimation device 7 (Figure 15) provides, and exports this frequency-correlated components respectively and give frequency-relevant frequency spectrum gain amending unit 1501 ₀To 1501 _K-1Demultiplexer 1503 is separated into frequency-correlated components with the spectrum gain that spectrum gain generator 8 (Figure 15) provides, and exports this frequency-correlated components respectively and give frequency-relevant frequency spectrum gain amending unit 1501 ₀To 1501 _K-1Frequency-relevant frequency spectrum gain amending unit 1501 ₀To 1501 _K-1In each, the frequency that frequency-correlation estimation priori signal to noise ratio (S/N ratio) that provides according to demultiplexer 1502 and demultiplexer 1503 provide-relevant frequency spectrum gain, calculated rate-relevant is revised spectrum gain, and frequency-relevant spectrum gain of revising of exporting this calculating is given multiplexer 1504.Multiplexer 1504 multiplexed frequencies-relevant frequency spectrum gain amending unit 1501 ₀To 1501 _K-1The frequency that provides-be correlated with and revise spectrum gain, and multiplexed correction spectrum gain is outputed to multiplexed multiplier 16 and priori signal-to-noise ratio (snr) estimation device 7.

Figure 24 display frequency-relevant frequency spectrum gain amending unit 1501 ₀To 1501 _K-1Structure.Because frequency-relevant frequency spectrum gain amending unit 1501 ₀To 1501 _K-1Structurally be identical each other, so they are expressed as the frequency-relevant frequency spectrum gain amending unit 1501 among Figure 24.Frequency-relevant frequency spectrum gain amending unit 1501 has maximum-value selector 1591, spectrum gain lower limit storer 1592, threshold value memory 1593, comparer 1594, switch (selector switch) 1595, modified value storer 1596 and multiplier 1597.

Frequency-correlation estimation priori signal to noise ratio (S/N ratio) that threshold value that comparer 1594 provides threshold value memory 1593 and demultiplexer 1502 (Figure 23) provide compares mutually.If frequency-correlation estimation priori signal to noise ratio (S/N ratio) is greater than this threshold value, then comparer 1594 provides " 0 " to give switch 1595.If frequency-correlation estimation priori signal to noise ratio (S/N ratio) is less than this threshold value, then comparer 1594 provides " 1 " to give switch 1595.When the output of comparer 1594 was " 1 ", switch 1595 was exported to multiplier 1597 with the signal that demultiplexer 1503 (Figure 23) provides.When comparer 1594 outputs were " 0 ", the signal that switch 1595 provides demultiplexer 1503 was exported to maximum-value selector 1591.In other words, when frequency-correlation estimation priori signal to noise ratio (S/N ratio) during less than this threshold value, spectrum gain is corrected.Because this spectrum gain is in hour correction of signal to noise ratio (S/N ratio), thus inhibition exceedingly of voice components, and also the numerical value of residual noise reduces.The product of the output valve of multiplier 1597 compute switches 1595 and the output valve of modified value storer 1596, and export this result of calculation to maximum-value selector 1591.In order to reduce spectral gain value, modified value is usually less than 1.Yet, whether can select this modified value to depend on the purpose of noise suppressor.In conventional noise suppressor shown in Figure 1, spectrum gain provides to multiplexed multiplier 16 and priori signal-to-noise ratio (snr) estimation device 7.Yet in the noise suppressor according to first embodiment, what offer multiplexed multiplier 16 and priori SNR estimation device 7 is the spectrum gain of revising, rather than spectrum gain.

Spectrum gain lower limit storer 1592 provides a lower limit of storing for spectrum gain to maximum-value selector 1591.The gain of the frequency that maximum-value selector 1591 provides switch 1595-relevant frequency spectrum compares mutually with the spectrum gain lower limit that spectrum gain lower limit storer 1592 provides, and exports bigger among them one and give multiplexer 1504 (Figure 23).Here it is, and spectrum gain is always greater than the lower limit of storing in spectrum gain lower limit storer 1592.Therefore, prevent owing to excessive inhibition makes voice distortion.

Figure 25 shows second example that is included in the spectrum gain generator structure in the noise suppressor shown in Figure 1.Have least mean-square error STSA gain function value counter 811, normalization likelihood ratio calculator 812, speech at the spectrum gain generator 81 of this graphic extension and have probability storer 813 and spectral gain calculator 814.Spectrum gain generator 81 is different from shown in Figure 15 by searching for the spectrum gain generator 8 of determining spectrum gain, because noise spectrum gain generator 81 recently calculates spectrum gain according to estimation priori signal to noise ratio (S/N ratio) with the posteriority noise that provides.The process of calculating spectrum gain according to the equation of describing in the list of references 1 will be described below.

Suppose that the numbering of frame represents that with n frequency number is represented with k, k, γ _n(k) frequency that the expression frequency-relevant snr computation device 6 (Figure 15) provides-relevant posteriority signal to noise ratio (S/N ratio), and Frequency-correlation estimation priori signal to noise ratio (S/N ratio) that expression priori signal-to-noise ratio (snr) estimation device 7 (Figure 15) provide.Also can suppose:

${\mathrm{\&eta;}}_n\left(k\right)={\widehat{\mathrm{\&xi;}}}_n\left(k\right)/q,$

υ _n(k)＝η _n(k)·γ _n(k)/(1+η _n(k))υ _n(k)

Least mean-square error STSA gain function value counter 811 is according to posteriority signal to noise ratio (S/N ratio) that frequency-relevant snr computation device 6 provides, the estimation priori signal to noise ratio (S/N ratio) that priori signal-to-noise ratio (snr) estimation device 7 provides, and the speech that speech exists probability storer 813 to provide exists probability q to come to be each frequency computation part least mean-square error STSA gain function value, and the least mean-square error STSA gain function value of exporting this calculating is given spectral gain calculator 814.Least mean-square error STSA gain function value G for each frequency _n(k) provide by equation (18).In equation (18), I ₀(z) Bessel's function of expression the 0th rank correction, and I ₁(z) Bessel's function of expression the 1st correction.Modified Bessel function is to be described in 1985,374 pages of G (list of references 5) of " mathematics dictionary " that Iwanami Shoten publishes.

$G_n\left(k\right)=\frac{\sqrt{\mathrm{\&pi;}}\sqrt{{\mathrm{\&upsi;}}_n\left(k\right)}}{2\&CenterDot;{\mathrm{\&gamma;}}_n\left(k\right)}\exp (-\frac{{\mathrm{\&upsi;}}_n\left(k\right)}{2})\&CenterDot;[(1+{\mathrm{\&upsi;}}_n\left(k\right))I_0\left(\frac{{\mathrm{\&upsi;}}_n\left(k\right)}{2}\right)+{\mathrm{\&upsi;}}_n\left(k\right)I_1\left(\frac{{\mathrm{\&upsi;}}_n\left(k\right)}{2}\right)]----\left(18\right)$

Normalization likelihood ratio calculator 812 is according to posteriority signal to noise ratio (S/N ratio) γ that frequency-relevant snr computation device 6 provides _n(k), the estimation priori signal to noise ratio (S/N ratio) that provides of priori signal-to-noise ratio (snr) estimation device 7

And the speech that speech exists probability storer 813 to provide exists probability q to come to be each frequency computation part normalization likelihood ratio, and the normalization likelihood ratio of exporting this calculating is given spectral gain calculator 814.Normalization likelihood ratio Λ for each frequency _n(k) be to represent by equation (19):

${\mathrm{\&Lambda;}}_n\left(k\right)=\frac{\mathrm{q\; exp}\left({\mathrm{\&upsi;}}_n\left(k\right)\right)}{1-\mathrm{<figure-callout\; id="1"\; label="q"\; filenames="A20051012008400351.png,A20051012008400371.png"\; state="\{\{state\}\}">q</figure-callout>}1+{\mathrm{\&eta;}}_n\left(k\right)}----\left(19\right)$

The least mean-square error STSA gain function value G that spectral gain calculator 814 provides according to least mean-square error STSA gain function value counter 811 _nAnd the normalization likelihood ratio Λ that provides according to normalization likelihood ratio calculator 812 (k), _n(k), be each frequency computation part spectrum gain, and the spectrum gain of exporting this calculating is given spectrum gain amending unit 15 (Figure 15).Spectrum gain G for corresponding frequencies _n(k) with equation (20) expression:

${\stackrel{\&OverBar;}{G}}_n\left(k\right)=\frac{{\mathrm{\&Lambda;}}_n\left(k\right)}{{\mathrm{\&Lambda;}}_n\left(k\right)+1}{G}_n\left(k\right)----\left(20\right)$

In noise suppressor shown in Figure 15, for the frequency band scope that comprises a plurality of frequencies, can determine and use common signal to noise ratio (S/N ratio), rather than frequency-relevant signal to noise ratio (S/N ratio).Second example that is used to calculate the frequency of signal to noise ratio (S/N ratio)-relevant snr computation device 6 for frequency band separately will be described below.

The structure of frequency-bandwidth shown in Figure 26-relevant snr computation device 61 can be used for replacing the frequency of noise suppressor shown in Figure 15-relevant snr computation device 6.Frequency-bandwidth-relevant snr computation device 61 is different with frequency-relevant snr computation device 6 shown in Figure 9, and wherein it has frequency-bandwidth-related power counter 611,612.The frequency that frequency-bandwidth-related power counter 611 provides according to demultiplexer 602-relevant has noise speech power frequency spectrum to come calculated rate-bandwidth-related power, and frequency-bandwidth-related power of exporting this calculating is respectively given separation vessel 601 ₀To 601 _K-1The frequency that frequency-bandwidth-related power counter 612 provides according to demultiplexer 603-correlation estimation noise power spectrum comes calculated rate-bandwidth-related power, and frequency-bandwidth-related power of exporting this calculating is respectively given separation vessel 601 ₀To 601 _K-1

The structure of Figure 27 display frequency-bandwidth-related power counter 611.In the example of graphic extension, whole power spectrum is divided into M the frequency band with equal bandwidth L, and wherein L, M are the satisfied K=LM that concerns of natural number.

Frequency-bandwidth-related power counter 611 has M totalizer 6110 ₀To 6110 _M-1The frequency that provides from demultiplexer 602 (Figure 26)-relevant has noise speech power spectrum component 910 ₀To 910 _K-1(910 ₀To 910 _ML-1) be transferred to totalizer 6110 respectively corresponding to each frequency ₀To 6110 _M-1For example, because be 0 to L-1, so frequency-relevant has noise speech power spectrum component 910 corresponding to the frequency number of frequency band number 0 ₀To 910 _L-1Be transferred to totalizer 6110 ₀Similarly, for example, because be that L is to 2L-1, so frequency-relevant has noise speech power spectrum component 910 corresponding to the frequency number of frequency band number 1 _LTo 910 _2L-1Be transferred to totalizer 6110 ₁Totalizer 6110 ₀To 6110 _M-1The frequency that calculating provides-relevant have noise speech power spectrum component each and, and output frequency-bandwidth-relevant has noise speech power spectrum component 911 ₀To 911 _ML-1(911 ₀To 911 _K-1) to separation vessel 601 ₀To 601 _K-1(Figure 26).Provide from totalizer 6110 ₀To 6110 _K-1Result of calculation as frequency-bandwidth-relevant noise speech power spectrum component is arranged, for the frequency that relies on each frequency band number.For example, from totalizer 6110 ₀Result of calculation output as frequency-bandwidth-relevant noise speech power spectrum component 911 is arranged ₀To 911 _L-1, and from totalizer 6110 ₁Result of calculation output as frequency-bandwidth-relevant noise speech power spectrum component 911 is arranged _LTo 911 _2L-1

Frequency-bandwidth-related power counter 612 is equivalent to frequency-bandwidth-related power counter 611 in structure and operation.Therefore, frequency-bandwidth-related power counter 612 below will not done detailed description.

A plurality of frequency band ranges of describing before whole power spectrum is divided into simultaneously, also can use other frequency band segregation method, such as, be used to separate the method that whole power spectrum becomes critical band, it is disclosed in electronics, information and communication enineer association in 1980, the 115-118 page or leaf of " monitoring and speech " (list of references 6), perhaps be used to separate the method that whole power spectrum becomes the octave frequency band, as be disclosed in nineteen eighty-three, U.S. Prentice-Hall company, " multi-rate digital signal processing " (list of references 7).

The second embodiment of the present invention will be described below.Be different from noise suppressor according to first embodiment shown in Figure 15 according to the noise suppressor of second embodiment shown in Figure 28, wherein noise judging unit 5 replaces with noise judging unit 52 and saves weighting noise speech counter 14.Noise suppressor according to second embodiment mainly is described below with respect to these differences.

Figure 29 shows the structure that is included in according to the noise judging unit 52 in the noise suppressor of second embodiment.Noise judging unit 52 is different from noise judging unit 5 shown in Figure 19, its medium frequency-correlation noise judging unit 504 ₀To 504 _K-1With frequency-correlation noise judging unit 506 ₀To 506 _K-1Replace, and the input signal of noise judging unit 52 does not have weighting that noise speech power frequency spectrum is arranged.This is because in view of the frequency in the noise judging unit 5 shown in Figure 19-correlation noise judging unit 504 ₀To 504 _K-1Needing input to have frequency-related weighing has noise speech power frequency spectrum, and the signal noise judging unit 506 in the noise judging unit 52 ₀To 506 _K-1Do not need input to have frequency-related weighing noise speech power frequency spectrum is arranged.

Figure 30 is that a demonstration is included in the frequency-correlation noise judging unit 506 in the noise judging unit 52 shown in Figure 29 ₀To 506 _K-1Block diagram.Because frequency-correlation noise judging unit 506 ₀To 506 _K-1Structurally mutually the same, so they are expressed as the frequency-correlation noise judging unit 506 among Figure 30.Frequency-correlation noise judging unit 506 is different from frequency shown in Figure 20-correlation noise judging unit 504, wherein it does not have use to have the input signal that weighting has noise speech power frequency spectrum, and it has separation vessel 5041, Nonlinear Processing machine 5042 and multiplier 5043.Noise judging unit 506 is primarily aimed at these difference and is described below.

Separation vessel 5041 provides noise power spectrum in the previous frame by estimation from estimated noise storer 5942, separating the frequency that demultiplexer 502 (Figure 29) provides-relevant has noise speech power frequency spectrum, and the result that output separates gives Nonlinear Processing machine 5042.Nonlinear Processing machine 5042, it is identical in structure and function aspects and Nonlinear Processing machine 1485 shown in Figure 17, calculate the weighting factor that depends on separation vessel 5041 outputs, and the weighting factor that output is calculated is given multiplier 5043.Frequency-relevant weighting factor product that has noise speech power frequency spectrum and Nonlinear Processing machine 5042 to provide that demultiplexer 502 (Figure 28) provides is provided multiplier 5043, and exports this product and give switch 5044.

Be equivalent to frequency-related weighing noise speech power spectrum component frequency shown in Figure 180-correlation noise judging unit 504 from the signal of multiplier 5043 output.Specifically, frequency-related weighing has noise speech power frequency spectrum to calculate in frequency-correlation noise judging unit 506.Therefore in the noise suppressor according to second embodiment, can save weighting has noise speech counter.

The third embodiment of the present invention will be described below.Be different from noise suppressor according to first embodiment shown in Figure 15 according to the noise suppressor of third embodiment of the invention shown in Figure 31, wherein priori signal-to-noise ratio (snr) estimation utensil has different inner structures.Figure 32 shows the structure of the priori signal-to-noise ratio (snr) estimation device 71 that is used for the 3rd embodiment.Priori signal-to-noise ratio (snr) estimation device 71 is different from priori signal-to-noise ratio (snr) estimation device 7 shown in Figure 10, wherein it has estimated noise storer 712, the multiplexed multiplier 716 that strengthens speech power frequency spectrum storer 713, frequency-relevant snr computation device 715 and replace posteriority signal to noise ratio (S/N ratio) storer 702, spectrum gain storer 703 and multiplexed multiplier 705,704.In addition, in view of the input signal of priori signal-to-noise ratio (snr) estimation device 7 shown in Figure 10 comprises spectrum gain, the input signal of priori signal-to-noise ratio (snr) estimation device 71 shown in Figure 32 comprises enhancing voice spectrum amplitude and the estimated noise power spectrum that replaces spectrum gain.

Multiplexed multiplier 716 provides the enhancing voice spectrum amplitude of each frequency to carry out square multiplexed multiplier 16 (Figure 31), and exports this enhancing speech power frequency spectrum of determining and give and strengthen speech power frequency spectrum storer 713.Because multiplexed multiplier 716 is with the multiplexed multiplier 17 with reference to figure 5 is structurally identical as mentioned above, thus below be not described in detail multiplexed multiplier 716.Strengthen the enhancing speech power frequency spectrum that the multiplexed multiplier 716 of speech power frequency spectrum storer 713 storages provides, and give frequency-relevant snr computation device 715 with the enhancing speech power frequency spectrum that the mode of previous frame is exported storage.Because frequency-structurally with as mentioned above frequency-relevant snr computation the device 6 with reference to figure 9 is identical for relevant snr computation device 715, so below be not described in detail frequency-relevant snr computation device 715.The estimated noise power spectrum that estimated noise storer 712 storage noise judging units 5 (Figure 31) provide, and the mode that the estimated noise power of storage is composed with previous frame outputs to frequency-relevant snr computation device 715.Frequency-relevant snr computation device 715, be each frequency, the signal to noise ratio (S/N ratio) of the speech power frequency spectrum that rises that calculating enhancing speech power frequency spectrum storer 713 provides, and the signal to noise ratio (S/N ratio) of calculating the estimated noise power spectrum that estimated noise storer 712 provides, and the signal to noise ratio (S/N ratio) that output is calculated is given multiplexed weighted summer 707.

The output signal of the frequency in priori signal-to-noise ratio (snr) estimation device 71 shown in Figure 32-relevant snr computation device 715 is equivalent to the output signal of the multiplexed multiplier 705 in the priori signal-to-noise ratio (snr) estimation device 7 shown in Figure 10.Therefore, according to the 3rd embodiment, priori signal-to-noise ratio (snr) estimation device 7 can replace with aforesaid priori signal-to-noise ratio (snr) estimation device 71.

The fourth embodiment of the present invention will be described below.Be different from noise suppressor according to the noise suppressor of fourth embodiment of the invention shown in Figure 33, wherein use the priori signal-to-noise ratio (snr) estimation device 71 (seeing Figure 32) in the 3rd embodiment to be used as a priori signal-to-noise ratio (snr) estimation device according to second embodiment shown in Figure 28.Noise judging unit 52 in structure and work with use in a second embodiment similar, and priori signal-to-noise ratio (snr) estimation device 71 in structure and work with use similar in the 3rd embodiment.Therefore, the function that noise suppressor is finished shown in Figure 33 is equivalent to the function according to the first embodiment noise suppressor shown in Figure 15.

The fifth embodiment of the present invention will be described below.Be different from noise suppressor according to first embodiment shown in Figure 15 according to the noise suppressor of the 5th embodiment shown in Figure 34, wherein noise judging unit 5 replaces with noise judging unit 53 and saves weighting noise speech counter 4.Therefore, noise suppressor so arranges it not need to use the voice activity detector of estimated noise.Be primarily aimed in these differences according to the noise suppressor of the 5th embodiment and be described below.

Figure 35 shows the structure that is used in the noise judging unit 53 among the 5th embodiment.Noise judging unit 53 is different from noise judging unit 5 shown in Figure 19, its medium frequency-correlation noise judging unit 504 ₀To 504 _K-1With frequency-correlation noise judging unit 508 ₀To 508 _K-1Replace, and input signal does not comprise voice activation detection sign.

Figure 36 shows each frequency-correlation noise judging unit 508 ₀To 508 _K-1Structure.Because frequency-correlation noise judging unit 508 ₀To 508 _K-1Structurally mutually the same, so they are expressed as the frequency-correlation noise judging unit 508 among Figure 36.Frequency-correlation noise judging unit 508 is different from frequency shown in Figure 20-correlation noise judging unit 504, wherein upgrade judging unit 520 usefulness and upgrade judging unit 522 replacements, and this input signal does not comprise voice activation detection sign.The structure of judging unit 522 is upgraded in graphic extension among Figure 37.Upgrade judging unit 522 and be different from renewal judging unit 520 shown in Figure 21, wherein OR circuit 5221 replaces OR circuit 5201, saves not circuit 5202, and input signal does not comprise voice activation detection sign.Specifically, upgrade judging unit 522 and be different from renewal judging unit 520 shown in Figure 21, it does not use voice activation to detect sign in upgrading estimated noise.The logic that OR circuit 5221 is calculated comparer 5205 output valves and comparer 5203 output valves with, and the result exported to trigger 5044, shift register 5045, adds counter 5049 (Figure 36).Renewal judging unit 522 is always exported " 1 " and is reached a predetermined value up to count value.After count value reaches predetermined value, upgrade judging unit 522 outputs " 1 " during less than threshold value as the noise speech power.With reference to aforesaid Figure 21, it is noise that comparer 5025 defines the noise speech signal.That is, comparer 5205 detects the speech of each frequency.Therefore, utilize said structure, may realize upgrading judging unit, it does not need voice activation detection sign to be included in the input signal.

Noise suppressor is described in the above according to the preferred embodiment of the invention.In the superincumbent description, supposed that least mean-square error fast frequency spectrum amplitude is used as the algorithm of squelch.Yet the present invention also is applicable to other noise suppression algorithm.Such noise suppression algorithm is a kind of to be that Wiener filtering is handled, and it is disclosed in Dec, 1979, IEEE proceedings, 67 volume numberings 12,1586-1604 page or leaf (list of references 8).

Industrial applicibility

According to the present invention, as mentioned above because the power spectrum of noise be utilize weighting the noise speech arranged Power spectrum is estimated, so can accurately estimating, the power spectrum of noise has nothing to do with Noise Characteristic, because of This utilizes minimizing distortion and noise to produce the enhancing speech. In addition, according to the present invention, because noise is Utilization is depended on what the improvement spectrum gain of snr value (signal-right-noise ratio) suppressed, so Can utilize the distortion and the noise that reduce with respect to whole snr values to produce the enhancing speech.

Claims (10)

1. the method for an estimated noise comprises step:

Convert input signal to frequency-region signal;

Determine weighting according to described frequency-region signal;

According to described weighting to described frequency-region signal weighting to determine the weighted frequency-domain signal; And

Determine the noise of estimation according to described weighted frequency-domain signal.

2. the method for estimated noise according to claim 1 is characterized in that describedly determining that the step of described weighting comprises step:

Determine the frequency dependence signal to noise ratio (S/N ratio) according to described frequency-region signal; And

According to the described weighting of described frequency dependence snr computation.

3. the method for estimated noise according to claim 2 is characterized in that describedly determining that the step of described weighting comprises:

Utilize nonlinear function to handle described frequency dependence signal to noise ratio (S/N ratio) to determine the step of described weighting.

4. according to the method for claim 2 or 3 described estimated noise, it is characterized in that the step of described definite described frequency dependence signal to noise ratio (S/N ratio) comprises step: determine described frequency dependence signal to noise ratio (S/N ratio) according to the noise of described frequency-region signal and the described estimation in former frame.

5. according to the method for any one described estimated noise among the claim 1-4, it is characterized in that the step of the noise of described definite described estimation comprises step: the noise of determining described estimation according to the moving average of described weighted frequency-domain signal.

6. equipment that is used for estimated noise is characterized in that comprising:

The input signal transducer, it is used for converting input signal to frequency-region signal;

Weight calculator, it is used for determining weighting according to described frequency-region signal input signal;

Weighting frequency-region signal counter, it utilizes described weighting to described frequency-region signal weighting, to determine the weighted frequency-domain signal; And

The noise estimation unit, it is used for determining according to described weighted frequency-domain signal the noise of estimation.

7. the equipment of estimated noise according to claim 6, it is characterized in that described weight calculator comprises: frequency dependence snr computation device is used for determining the frequency dependence signal to noise ratio (S/N ratio) according to described frequency-region signal.

8. the equipment of estimated noise according to claim 7 is characterized in that described weight calculator comprises: nonlinear processor is used to utilize nonlinear function to handle described frequency dependence signal to noise ratio (S/N ratio) to determine described weighting.

9. according to the equipment of claim 7 or 8 described estimated noise, it is characterized in that described frequency dependence snr computation device determines described frequency dependence signal to noise ratio (S/N ratio) according to described frequency-region signal and the estimated noise in former frame.

10. according to the equipment of any one the described estimated noise among the claim 6-9, it is characterized in that the noise estimation unit comprises: the moving average counter is used for determining according to the moving average of described weighted frequency-domain signal the noise of described estimation.

Images