JP2001516902A - How to suppress noise in digital audio signals - Google Patents

Abstract

(57) [Summary] The present invention calculates signal spectrum components (S _{n, f} , S _{n, i} ) in each frame and maximizes estimation of noise spectrum components included in a speech signal. Quantity [Equation 1] , Calculating the pitch by performing harmonic analysis, and estimating the maximum amount of noise corresponding to the spectral component from each spectral component (S _{n, f} ) of the speech signal in the frame. Performing a spectral subtraction comprising at least steps of subtracting a parameter-dependent quantity, including the adjusted pitch, and applying a time-domain transform to the subtracted result to enhance the speech signal (s ³ ). And suppressing noise in the digital that is processed by subsequent frames.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

TECHNICAL FIELD OF THE INVENTION

The present invention relates to a method for suppressing noise in a digital speech signal. More specifically, it relates to noise suppression by non-linear spectral subtraction.

[0002] Due to the widespread adoption of new forms of communication, especially mobile telephones, calls are increasingly being made in very noisy environments. Noise, in addition to speech, disrupts speech by preventing optimal compression of speech signals, resulting in unnatural background noise. The noise makes the spoken message difficult to understand and tires.

[0003]

[Prior art]

Many algorithms have been studied in an attempt to reduce the effects of noise in calls. S. F. Boll ("Suppression of Acoustic Noise in Speech Using Spectral Subtraction").
speech using spectral subtraction) "
IEEE Trans. For Sound, Speech and Signal Processing. (IEEE
Trans. on Acoustics, Speechand Signa
lProcessing), Vol. ASSP-27, no. 2 1979 4
Mon) Proposing an algorithm based on spectral subtraction. This technique consists of estimating the spectrum of the noise during the silence phase and subtracting it from the received signal. It reduces the noise level received. Its main drawback is musical noise, which is particularly noisy due to its unnatural nature.
).

[0004] This work is described in B. Paul (“Spectral Envelope Estimation Vocoder (Thes
principal envelope estimation vocoder)
IEEE Trans. For Sound, Speech and Signal Processing. (IEEE
ETrans. on Acoustics, Speechand Signa
l Processing), Vol. ASSP-29, no. 4 August 1981)), and P.S. Lockwood and J.M. Buddy ("Nonlinear Spectral Subtractor for Strong Speech Recognition in Cars, and Experiments with Hidden Markov Models and Projections (Experiments withtha)
nonlinear spectral subtractor (NSS)
, Hidden Markov Models and the project
tion, for robust speech recognition i
n cars), Speech Communication, vol. 11,
June 1992, pages 215-228, and EPO Patent Application Publication 05.
34837), which greatly improves the noise level, while retaining its natural features. Furthermore, this contribution has the advantage that it includes the principle of masking the computation of the noise suppression filter first. Based on this idea, to use explicitly calculated masking curves in spectral subtraction, first attempts were made by S.D. Nandkumar and J.M. H. L. Han
sen ("Speech enhancement on a new sequel"
tof auditory constrained parameters)
Proc. ICASSP 94, I.P. 1 to I. 4 pages). Despite the disappointing results of the above technique, this contribution had the advantage of emphasizing the importance of not degrading the speech signal during noise suppression.

[0005] Other methods based on splitting the speech signal into respective values, and thus directing the speech signal to a smaller space, are known as Bart De M
oore ("single value resolution and long and short spaces of noisy matrices (The single value decomposi
Tion and long and short spaces of no
isy matrices) ", IEEE Trans. (
IEEE Trans. on signal processing), Vo
l. 41, no. 9, September 1993, pp. 2826-2838), and S
. H. Jensen et al. (Reduction of broad-band noise in speech due to truncated QSVD).
in speech by truncated QSVD), IEEE Trans. for speech and audio processing. (IEEE Trans.
on Speech and Audio Processing), Vol.
3, No. 6, November 1995). The principle of the above technique assumes that the speech signal and the noise signal are not completely correlated, and that they have sufficient predictability to predict the speech signal based on a limited set of parameters. It is to consider. While this technique provides acceptable noise suppression for highly voiced signals, it completely changes the nature of the speech signal. When faced with relatively coherent noise, such as vehicle tire noise or engine noise, the noise can be predicted much more easily than unvoiced speech signals. This tends to project the speech signal onto a portion of the noise vector space. This method does not take into account speech signals, especially unvoiced speech regions, which are less predictable. Furthermore, predicting the speech signal based on a small set of parameters prevents taking into account all of the inherent richness of speech. The limitations of techniques based solely on mathematical considerations and monitoring the specific nature of speech are obvious.

[0006] Finally, other techniques are based on coherence criteria. The coherence function is described in J. A. Cadzoo and O.M. M. Particularly well developed by Solomon ("Linear modeling and coherence functions"
ng and the coherence function), IEEE Trans. for Sound, Speech and Signal Processing. Vol. ASSP-3
5, no. 1, January 1987, pp. 19-28). Also, its application to noise suppression is described in R.S. Developed by Le Rouquin ("Enhancing speech signals containing noise: Application to mobile radio communications (Enhancement)
of noisy speech signals: application t
mobile radio communications), Speech Communication, Vol. 18, pages 3-19). This method is based on the fact that if multiple independent channels are used, the speech signal is much more coherent than noise. The results obtained seem to be quite encouraging. However, this technique unfortunately requires multiple voice pickup points. This is not always possible.

US Pat. No. 5,228,088 describes a noise suppression device that operates in the frequency domain and incorporates a pitch detector. The results of such detection are used to adjust the noise suppression factor and look for "voice bands". The noise suppression factor is used by the spectral subtraction module to weight the noise estimator before subtracting it from the signal. The module that adjusts the suppression factor uses only information that indicates whether a pitch has been detected. However, the pitch has no effect on the suppression factor used. "Voice bands" determined with the aid of the detected pitch undergo global signal enhancement. It can instead be used to determine the "noise band" to which global attention is directed. Such emphasis or attenuation of parts of the spectrum, and of parts of the signal, is a very different noise suppression method than spectral subtraction.

[0008]

[Problems to be solved by the invention]

A main object of the present invention is to provide a novel noise suppression technique that enables efficient noise suppression without deteriorating speech perception, taking into account the characteristics of speech generation.

[0009]

[Means for Solving the Problems]

Therefore, the present invention provides:-a harmonic analysis of the speech signal to estimate the pitch frequency of the speech signal in each frame, characterized by internal speech activity, and the calculation of the spectral content of the speech signal for each frame Calculating, for each frame, an estimate of the spectral component of the noise contained in the speech signal, and from each spectral component of the speech signal in the frame, an estimate of the corresponding spectral component of the noise for the frame; Performing a spectral subtraction, including at least one step of subtracting a respective amount that depends on a parameter that includes at least an estimated pitch frequency value, and performing a spectral subtraction of the result of the spectral subtraction to form a noise-suppressed speech signal. Transformed into time domain, processed by successive frames It proposes a method of suppressing noise in digital speech signals.

[0010] Harmonic analysis of the speech signal is performed to estimate the pitch frequency of the speech signal over each frame characterized by voice activity. The parameters on which the amount to be subtracted depends include the pitch frequency so estimated.

It is generally desirable to overestimate the spectral envelope of the noise such that the overestimate obtained by overestimating the spectral envelope of the noise is robust to sudden changes in the noise. However, overestimation that is too large usually has the disadvantage of distorting the speech signal. This disadvantage is very troublesome on the phone, since the speech signal has the most energy in the voiced area. Taking into account the pitch frequency of the speech signal in noise suppression protects the harmonic content of the signal in those voiced regions.

As a general rule, in order to subtract a given spectral component from the speech signal, the spectral component must match the frequency being protected, ie
If it is closest to an integer multiple of the estimated pitch frequency, then an amount less than the amount when the spectral components do not match any of such protected frequencies is employed. This small amount can in particular be zero. In the latter case, the spectral subtraction does not affect the signal at the estimated pitch frequency and / or its harmonics. This eliminates some of the non-linearities introduced by noise overestimation, and they are particularly sensitive to voiced regions. Due to the more random nature of its excitation signal, unvoiced regions are less sensitive to this.

In one advantageous embodiment, after estimating the pitch frequency of the speech signal in a frame, oversampling the speech signal of the frame with an oversampling frequency that is a multiple of the estimated pitch frequency. , The speech signal is adjusted, and the spectral content of the speech signal in the frame is calculated based on the adjusted signal, and the amount is subtracted therefrom. The frequency closest to the pitch frequency estimated thereby is more favorable than the other frequencies. This avoids protecting harmonics that are far away from the harmonics of the pitch frequency. Thus, the harmonic nature of the speech signal is preserved as much as possible. To calculate the spectral content of the speech signal, the conditioned signal is distributed between blocks of N samples that are transformed into the frequency domain, and the difference between the oversampled frequency and the estimated pitch frequency is calculated. The ratio is selected as a factor of the number N.

The prior art estimates the time interval between two consecutive breaks in the signal that can be attributed to the glottal closure of the speaker during the frame: The estimated pitch frequency is inversely proportional to the time interval;-interpolating the speech signal within the time interval such that the adjusted signal as a result of the interpolation is constant between two successive breaks. To have a time interval of
By estimating the pitch of the speech signal over the frame at, it can be made even better.

This artificially composes a signal frame in which the speech signal has discontinuities at regular intervals. Therefore, any change in pitch during the duration of the frame is taken into account.

In another refinement, after processing each frame, the noise is reduced to an integer multiple of the ratio between the sampling frequency and the estimated pitch frequency of the speech signal provided by this processing. Several equal samples are retained. This avoids distortion problems caused by phase discontinuities between frames. It is generally not completely modified by conventional overlap-add techniques.

The adjustment of the signal by the oversampling technique gives a good measure of the voicedness of the speech signal in the frame from the calculation of the entropy of the autocorrelation of the spectral components calculated on the basis of the adjusted signal. As the spectrum becomes more disturbed, ie, the voicedness of the spectrum increases, the entropy value decreases. Adjusting the speech signal increases the spectral irregularities, and thus the change in entropy, so that the latter constitutes a sensitive measurement. To achieve the best performance, the autocorrelation is generally calculated based on the noise suppressed signal. However, it is possible to calculate them based on the conditioned signals before noise suppression.

To calculate the masking curve by applying an auditory model, the spectral components of the noise-suppressed signal obtained by subtracting said quantities from the speech signal spectral components can be used. The parameter on which the amount deducted from the speech signal spectral components in the frame depends preferably comprises the difference between the overestimation of the noise spectral components and the calculated masking curve. The amount subtracted is particularly limited to a portion of the overestimation of the spectral components corresponding to noise above the masking curve. This approach is based on the observation that it is sufficient to suppress audible noise frequencies. In contrast, suppression of noise not masked by speech has no utility.

In an advantageous embodiment, each overestimate of the noise contained in the speech signal is
It is obtained by combining a long-term estimator of the spectral component of noise with a measure of variability of the long-term estimator of the spectral component of noise.
This is a noise estimator that pays attention to long-term noise fluctuations.
The combination of two separate noise estimators, tor) and a noise estimator that pays attention to the short-term variability of the noise, results in a noise estimator that is particularly resistant to noise fluctuations.

[0020]

BEST MODE FOR CARRYING OUT THE INVENTION

Other features and other advantages of the present invention will become apparent in the following description of non-limiting embodiments of the present invention, given with reference to the accompanying drawings.

The noise suppression device shown in FIG. 1 processes a digital speech signal s. The windowing module 10 formats the signal s into successive windows or frames. Each frame is composed of a number N of digital signal samples. In the usual way, the frames can overlap each other. For the remainder of this description, it will be assumed that the frame is composed of N = 256 samples, the sampling frequency F _e is 8 kHz, each window is hamming weighted, and the overlap between successive windows is 50%. This is not a limitation of the present invention.

The signal frame is transformed to the frequency domain by module 11 using a conventional fast Fourier transform (FFT) algorithm to calculate the modulus of the signal's spectrum. After that, the module 11 outputs a set of N = 256 frequency components S _{n, f} of the speech signal. Where n is the number of the current frame and f is the frequency from the individual spectrum. Due to the characteristics of the digital signal in the frequency domain, only the first N / 2 = 128 samples are used.

Instead of using the frequency resolution available on the downstream side of the fast Fourier transform to calculate an estimate of the noise contained in the signal s, the bandwidth of the signal [0, _Fe / 2] is used. A lower resolution is used, determined by the number I of frequency bands that are covered. Each band i

[0024]

[Equation 19] Extends from a lower frequency f (i-1) to a higher frequency f (i). Here, f (0) = 0 and f (I) = F _e / 2. Subdivision into frequency bands uniform, a (f (i) -f (I -1) = F e / 2I). It can also be non-uniform (eg, according to the bulk scale). The module 12 averages each of the spectral components S _{n, f} of the in-band speech signal, for example,

[0025]

(Equation 20) Calculate with uniform weighting such as

This averaging reduces the variation between the bands by averaging the contribution of the noise in the bands. It reduces the diversity of the noise estimator. Also, this averaging greatly reduces the complexity of the device.

The averaged spectral components S _{n, i} are sent to a voice activity detector module 15 and a noise estimator module 16. Two modules 15 and 16 are used in which the speech activity γ _{n, i} measured by the module 15 for the different bands is used by the module 16 to estimate the long-term energy of the noise in the different bands, Long term estimator for a priori suppression of noise in speech signals in various bands to determine speech activity γ _{n, i}

[0028]

(Equation 21) Are used by the module 15.

The operation of modules 15 and 16 can be consistent with the flowcharts shown in FIGS.

In steps 17 to 20, module 15 performs a priori suppression of noise in the speech signal in band i for signal frame n. This a priori suppression is 1
It is performed in a conventional non-linear spectral subtraction manner based on an estimate of the noise contained in one or more previous frames. In step 17, the band I
Using the decomposition capability of

[0031]

(Equation 22) From the frequency response Hpn _{, i} of the a priori noise suppression filter. Where τ1 and τ2 are delays expressed as the number of frames

[0032]

(Equation 23) , Α ′ _{n, i} are noise excessive estimation coefficients determined as described later. The higher the reliability in detecting the voice activity, the smaller the value of τ1.

In steps 18 and 20, the spectral components

[0034]

(Equation 24) But

[0035]

(Equation 25) Is calculated from Here, to prevent the .beta.p _i is a floor coefficient close to 0 (floor coef ficient), takes an excessively small value, such as the spectrum of the signal noise is suppressed produce negative or tone noise, It has been conventionally used for this purpose.

Thus, steps 17 to 20 subtract the noise spectrum estimator estimated a priori from the signal spectrum and add a factor

[0037]

(Equation 26) , Which are almost overweighted.

In step 21, module 15 determines the energy of the a priori noise suppressed signal in various bands i for frame n.

[0039]

[Equation 27] Is calculated. It is also calculated by summing the global average of the energies of the a priori noise-suppressed signals with the energy En _{, i} for each band, weighted by the width of the band. Index i = 0 is used to indicate the global bandwidth of the signal.

In steps 22 and 23, the module 15

[0041]

[Equation 28] In contrast, the magnitude ΔE _{n, i} representing a short-term change in the energy of the noise-suppressed signal in band i and the long-term value of the energy of the noise-suppressed signal in band i

[0042]

(Equation 29) Is calculated. The magnitude ΔE _{n, i} is a simplified equation

[0043]

[Equation 30] Can be calculated from Long time energy

[0044]

(Equation 31) Can be calculated using the forgetting factor B1 such that 0 <B1 <1; that is,

[0045]

(Equation 32) The energy E _{n, i of the} noise-suppressed signal and its short-term change ΔE _{n, i} ;
Its long time value

[0046]

[Equation 33] After calculating in the manner shown in FIG. 2, module 15

[0047]

(Equation 34) The evolution of the energy of the signal in which the noise has been suppressed (evolution)
Calculating the value [rho _i representing the. This calculation is performed in steps 25 to 36 in FIG. 3 and is performed for each band i from i = 0 to i = I. The calculation and long noise envelope estimator ba _i, an internal estimator bi _i, and a frame counter b _i noisy used.

In step 25, the magnitude ΔE _{n, i} is compared with a threshold value ε1. When not reached its threshold value .epsilon.1, counter b _i is allowed incremented by one unit in step 26. In step 27, the long-term estimator ba _i is a smoothed energy value

[0049]

(Equation 35) Is compared to

[0050]

[Equation 36] If, then the estimator ba _i is the smoothed value in step 28

[0051]

(37) And the counter b _i is reset to zero. Then,

[0052]

(38) (Step 36), the magnitude ρ _i is equal to one.

[0053]

[Equation 39] When showing the steps 27 that is, the counter b _i is compared with the limit Sakaichi bmax in step 29. If bi> bmax, the signal is considered too constant to support voice activity. Step 28 above is then performed, which causes the frame to be considered to contain only noise. In step 29

[0054]

(Equation 40) If so, the internal estimator bi _i is given by the equation

[0055]

[Equation 41] Is calculated from In the above equation, Bm represents an update coefficient. Its value depends on the state of the speech activity detector automaton (steps 30-32). The state δ _n-1 is determined during processing of the preceding frame. Assuming that the automaton is in the speech detection state (δ _n-1 = 2 in step 30), the noise estimator is updated very slightly in the presence of speech since the coefficient Bm takes a value Bmp very close to 1. Is done. Otherwise, the coefficient Bm takes on a smaller value Bms to allow a more meaningful update of the noise estimator in the silence phase. In step 34, the difference ba _i -bi _i between the long time estimator and the internal noise estimator is compared to a threshold ε2. If the threshold value ε2 has not been reached, then in step 35, the long time estimator ba _i is updated with the value of the internal estimator bi _i . Otherwise, the long term estimator ba _i remains unchanged. This prevents sudden changes due to speech signals that cause the noise estimator to update.

After the magnitude ρ _i has been obtained, the module 15 proceeds to the voice activity determination step 37. Module 15 first updates the state of the detected automaton according to the magnitude ρ ₀ calculated for all bands of the signal. As shown in FIG. 4, the new state δ _n of the automaton depends on the preceding states δ _n−1 and ρ ₀ .

Four states are possible: δ = 0 detects silence, ie absence of speech, δ = 2 detects the presence of voice activity, and states δ = 1 and δ = 3 are intermediate. A rising state and a falling state. Assuming that the automaton is in a silent state (δ _n-1 = 0), it will stay there if ρ ₀ does not exceed the initial threshold SE1, otherwise it will go to the rising state. In the rising state (δ _n-1 = 1), if ρ ₀ is smaller than the first threshold value SE1, it returns to the silence state, and if ρ ₀ is larger than the second threshold value SE2 which is larger than the threshold value SE1, speech is given. Go to the state,

[0058]

(Equation 42) Then stay in the rising state. If the automaton is in a speech state (δ _n-1 = 2), it will stay there if ρ ₀ exceeds a third threshold SE3 which is smaller than the threshold SE2, otherwise it will fall to go into. In the falling state (δ _n-1 = 3), if ρ ₀ is larger than the threshold SE2, the automaton returns to the speech state, and if ρ ₀ is smaller than the second threshold SE4 smaller than the threshold SE2. Return to silence,

[0059]

[Equation 43] Then, it stays in the falling state.

In step 37, the module 15

[0061]

[Equation 44] Also calculates the voice activity γ _{n, i} . This time, γ _{n, i} is a non-binary parameter, that is, swelling γ _{n, i} = G (ρ _i ) ranges from 0 to 1 continuously as a function of the value taken by the magnitude ρ _i . Preferably, it is a function that changes. This function has, for example, the form shown in FIG.

The module 16 calculates noise estimators on a band-by-band basis, and these estimators are used in the noise suppression process using the subsequent values of the components S _{n, i} and the voice activity γ _{n, i} . This corresponds to steps 40 to 42 in FIG. Step 40 determines whether the voice activity detector automaton has just advanced from a rising state to a falling state. If so, each band

[0063]

[Equation 45] The last two estimators previously calculated for

[0064]

[Equation 46] When

[0065]

[Equation 47] Estimator preceded by

[0066]

[Equation 48] Will be modified according to The modification is that in the rising phase (δ = 1), a long-term estimate of the energy of the noise in the voice activity detection process (steps 30 to 33) is as if the signal contained only noise (Bm = Bms). Calculated and done to allow for the fact that they result in errors.

In step 42, module 16 calculates the expression

[0068]

[Equation 49]

[0069]

[Equation 50] Is used to update the amount of noise estimation on a band-by-band basis. In the equation, λ _B represents a forgetting factor such that 0 <λ _B <1. Equation (6) shows that the nonlinear speech activity γ _{n, i} is taken into account.

As described above, the long-term estimated amount of noise

[0071]

(Equation 51) But before the noise suppression by nonlinear spectrum subtraction, the module 45 (FIG. 1)
Is overestimated by Module 45 calculates the overestimated coefficient α ′ _{n, i} as

[0072]

(Equation 52) Overestimate that almost matches

[0073]

(Equation 53) Calculate with

FIG. 6 shows the configuration of the over-estimation module 45. Estimator

[0075]

(Equation 54) Is the long-term estimator

[0076]

[Equation 55]And a measure of the variability ΔB of the component of the noise in band i around its long term estimator ^max _{n, i} Are obtained by combining In the example considered here,
The alignment is an almost simple addition performed by the adder 46. It instead of it
Can be weighted addition.

The overestimated coefficient α ′ _{n, i} is calculated by adding the sum

[0078]

[Equation 56] And the long-term estimator delayed

[0079]

[Equation 57] And has a maximum limit α _max , for example α _max = 4 (block 48). If necessary, the delay τ3 corrects the value of the overestimation coefficient α ′ _{n , i in} the rising phase (δ = 1) before the long-term estimator is corrected by steps 40 and 41 from FIG. (For example, τ3 = 3).

Overestimation

[0081]

[Equation 58] Is finally

[0082]

[Equation 59] (Multiplier 49).

The measure of noise variability ΔB ^max _{n, i} reflects the variation of the noise estimator. There, the value of _{Sn, i} calculated for a certain number of preceding frames, such that the speech signal does not characterize any voice activity in band i,

[0084]

[Equation 60] It is obtained as a function of the value of It is the difference calculated for a number K of silence frames

[0085]

[Equation 61] Is a function of

[0086]

(Equation 62) . In the example shown, this function is simply the maximum (block 50). Each frame n
, The voice activity γ _{n, i} is compared to a threshold (block 51) and the difference calculated in

[0087]

[Equation 63] Must be loaded into a queue 54 with K locations configured in a first-in, first-out (FIFO) mode. If γ _{n, i} does not exceed a threshold (which can be equal to 0 if the function g () has the form as shown in FIG. 5), the FIFO 54 is not loaded, otherwise It is loaded. Thereafter, the maximum value contained in the FIFO 54 is determined by the measured variability Δ
B ^max _{n, i} .

The measured variability ΔB ^max _{n, i} is instead replaced by the values S _{n, f} (not S _{n, i} ) and

[0089]

[Equation 64] Can be obtained as a function of Then, the FIFO 54 for each band i

[0090]

[Equation 65] Instead of,

[0091]

[Equation 66] The procedure is the same except that

Long-term fluctuation of noise

[0093]

[Equation 67] And the independent estimation of the short-term variability ΔB ^max _{n, i}

[0094]

[Equation 68] Makes the noise suppression process extremely resistant to musical noise.

The module 55 shown in FIG. 1 performs an initial spectral subtraction step. At this stage, band i

[0096]

[Equation 69] Provides the frequency response H ¹ _{n, i} of the first noise suppression filter as a function of the components S _{n, i} , B _{n , i} and the overestimated coefficient α ′ _{n, i} . This calculation is based on the equation for each band i.

[0097]

[Equation 70] Can be performed using Where τ4 is

[0098]

[Equation 71] (For example, τ4 = 0). Factor beta ¹ _i in equation (7), as the coefficient .beta.p _i in equation (3) represents a floor which has been conventionally used in order to avoid negative values or excessively small value of the signal with suppressed noise .

In a manner known in the art (see EPO Patent Application Publication No. 0534837), the overestimated coefficient α ′ _{n, i} of equation (7) is replaced by α ′ _{n, i} and the estimator of the signal-to-noise ratio (
For example,

[0100]

[Equation 72] ) Can be replaced by another coefficient equal to the function of This function is a sub-function of the estimated value of the signal-to-noise ratio. This function is then equal to α ′ _{n, i} for the lowest signal to noise ratio. Assuming the signal is very noisy, reducing the overestimation factor clearly has no value. This function is advantageous because it decreases towards zero for the highest value of the signal / noise ratio. This protects the highest energy region of the spectrum in which the speech signal is most meaningful. The quantity is then subtracted from the signal going to zero.

This approach can be refined if the latter is characterized by voice activity, by selectively applying it to harmonics of the pitch frequency of the speech signal.

Thus, in the embodiment shown in FIG. 1, a second noise suppression stage is performed by the harmonic protection module 56. This module converts the frequency response H ² _{n, f} of the second noise suppression filter, with the resolution of the Fourier transform _, to the parameters H ¹ _{n, i} , α
′ _{N, i} ,

[0103]

[Equation 73], Δ_n, S_{n, i}Out of silence by the function of and the harmonic analysis module 57
Pitch frequency f_p= F_e/ T_pCalculate as a function of Silence phase (δ_n= 0)
Then, the module 56 does not operate. That is, for each frequency f in band i, H ^Two _{n, f} = H¹ _{n, i}It is. Module 57 analyzes the speech signal of the frame,
Pitch period T expressed as an integer or fractional sample_{p '}In order to determine the,
Any conventional method for analyzing the speech signal of a frame, e.g.
Measurement method can be used.

The protection provided by module 56 is that for each frequency f belonging to band i,

[0105]

[Equation 74] Can be configured.

Δf = F _e / N represents the spectral resolution of the Fourier transform. Assuming that H ² _{n, f} = 1, the amount subtracted from the component S _{n, f} is zero. In this calculation, the floor - coefficient beta ² _{i ^{(e.g., β 2 i = β 1 i}} ) is the number of harmonics of the pitch frequency f _p can be masked by the noise, thus useful to protect them No, represents the fact that.

This protection strategy preferably applies to each of the frequencies closest to the harmonic of f _p , ie to any integer η.

[0108] When the frequency resolution in the analysis module 57 of the pitch frequency f _p estimated caused shall indicate that delta] f _p, i.e., the actual pitch frequency f _p -.DELTA.f _p / 2 and f _p + _{δf p} / 2 assuming that lies between, that the difference between the actual pitch frequency eta following harmonics and its estimator eta × f _p (condition (9)) proceeds to ± f _p × δf _p / 2 it can. For high values of η, the difference can be higher than the spectral half-resolution of the Fourier transform. To add this uncertainty into consideration, and the actual in order to ensure better protection of harmonics of the pitch, the range _{[η × f p -η × δf} p / 2, η × f p + η × δf p / 2 ] Can be protected, that is, the above condition (9)

[0109]

[Equation 75] This approach (9 ') is particularly advantageous if the value of η can be increased, especially if the device is used in a broadband device.

For each frequency to be protected, the modified frequency response H ² _{n, f} can be equal to one, as described above. This corresponds, in terms of spectral subtraction, to zero-subtraction, ie to ending protection of the frequency in question. More generally, the modified frequency response H ² _{n, f} can be taken to be equal to a value from 1 to H ¹ _{n, f} , depending on the degree of protection sought. This corresponds to subtracting an amount less than would be subtracted if the frequency in question had not been protected.

The multiplier 58 calculates the spectral component S ² _{n, f} of the noise-suppressed signal.

S ² _{n, f} = H ² _{n, f} · S _{n, f} This signal S ² _{n, f} is a psychoacoustic model (p.
Applying a sychoacoustic model is provided to a module 60 that calculates a masking curve for each frame.

The masking phenomenon is a well-known operating principle of the human ear. Assuming that two frequencies are present at the same time, one of those frequencies may not be audible.
Then it is said that it was masked.

There are various methods for calculating a masking curve. For example, in J.I. DJohnst
on "(Transform Coding of Audio S using the perceptual noise criterion).
signals Using Perceptual Noise Criter
ia) ", an IEEE journal on selected areas of communication (IEEE
EJournal on Selected Areas in Commun
ications), Vol. 2, February 1988). The method works on a bulk frequency scale. The masking curve can be seen as a convolution of the spectral extension function of the basilar membrane in the bulk region with the excitation signal, in this application the signal S ² _{n, f} . The spectral extension function can be modeled as shown in FIG. For each bulk band, the contribution of the lower and higher bands convolved with the basement membrane expansion function is calculated from the equation.

[0115]

[Equation 76] In this equation, the indices q and q 'are the bulk bands

[0116]

[Equation 77] , And S ² _{n, q} represents the average of the components S ² _{n, f} of the excitation signal in which noise for the individual frequency f belonging to the bulk band q ′ is suppressed.

The module 60 obtains a masking threshold M _{n, q} for each bulk band q from the equation M _{n, q} = C _{n, q} / R _q . In this equation, R _q is
It depends on whether the signal contains more or less voiced sounds. As known in the art, is a form of R _q is _{10 · log 10 (R q)} = (A + q) · χ + B · (1-χ). A = 14.5 and B = 5.5. χ indicates the voicedness of the speech signal that changes from 0 (no voice) to 1 (a signal with a very high voicedness). Parameter χ is the form known in the art.

[0118]

[Equation 78] Here, SFM represents the ratio, expressed in decibels, between the arithmetic mean and the geometric mean of the bulk band energy, where SFM _max = −60 dB.

The noise suppression device compares the frequency response of the noise suppression filter with the masking curve M _{n, q} calculated by the module 60 and the over-estimation quantification calculated by the module 45.

[0120]

[Expression 79] And a modification module 62 as a function of The noise in the signal is overestimated by comparing the envelope of the noise overestimator with the envelope formed by the masking threshold _{Mn, q.}

[0121]

[Equation 80] Is determined only in a range above the masking curve. This avoids unnecessary suppression of noise masked by speech.

The new response H ³ _{n, f} for the frequency f belonging to the band i and the bulk band q defined by the module 12 is an overestimate of the corresponding spectral component of the noise.

[0123]

[Equation 81] And the masking curve M _{n, q} as follows.

[0124]

(Equation 82) In other words, the amount subtracted from the spectral components S _{n, f} in the spectral subtraction process with the frequency response H ³ _{n, f} is the amount subtracted from the spectral components in the spectral subtraction process with the frequency response H ² _{n, f} , Overestimation of the corresponding spectral components of the noise, possibly beyond the masking curve M _{n, q}

[0125]

[Equation 83] Is approximately equal to the smaller of

FIG. 8 shows the principle of the modification applied by the module 62. That is, the spectral component S ² _{n, f} of the noise-suppressed signal and the overestimated amount of the noise spectrum

[0127]

[Equation 84] An example of the masking curve M _{n, q} calculated based on the above is shown in a schematic form. The last amount subtracted from the component Sn _{, f} is indicated by the shaded area, ie, it is an overestimate of the noise spectral components.

[0128]

[Equation 85] Are limited to the portion above the masking curve.

The subtraction is performed by multiplying the frequency response H ³ _{n, f} of the noise suppression filter by the spectral component S _{n, f} of the speech signal (multiplier 64). Then, an inverse fast Fourier transform (IFFT) is applied to the sample of frequency S ³ _{n, f} supplied by multiplier 64.
, The module 65 reconstructs the noise-suppressed signal in the time domain. For each frame, only the first N / 2 = 128 samples of the signal generated by module 65 are, after superposition-addition with N / 2 = 128 samples after the preceding frame, Provided as the last signal with noise suppressed (module 66).

FIG. 9 shows a preferred embodiment of a noise suppression device using the present invention. This device includes several components that are similar to the corresponding components of the device shown in FIG. The same reference numbers are used for those parts. Thus, the modules 10, 11, 1, 15, 16, 45 and 55 comprise the quantities S _{n, i} used for selective noise suppression,

[0131]

[Equation 86] , Α ′ _{n, i} ,

[0132]

[Equation 87] And H ¹ _{n, f} in particular.

The frequency resolution of the fast Fourier transform 11 constitutes a limitation of the device shown in FIG. The frequency protected by module 55 is not necessarily the exact pitch frequency f
It is not p but the frequency closest to it in the individual spectrum. In some cases,
Harmonics that are relatively far from the pitch frequency may be protected. The device shown in FIG. 9 reduces this drawback by properly adjusting the speech signal.

[0134] This adjustment is an integer sample time of a signal period 1 / f _p is adjusted just to cover, to correct the sampling frequency of the signal.

Many harmonic analysis methods that can be used by the module 57 can provide a fractional value of the delay T _{p ′} , represented as a number of samples at the initial sampling frequency F _e . Thereafter, a new sampling frequency fe equal to an integer multiple of the estimated pitch frequency is selected. That is, f _e = p · f _p = p · F _e / T _p = K · F _e where p is an integer. To avoid signal samples is lost, f _e must be higher than F _e. In particular, in order to facilitate adjustment, range f _e is from F _e to 2F _e

[0136]

[Equation 88] Can be imposed.

Of course, if no voiced activity is detected in the current frame (δ _n ≠ 0), or if the delay Tp estimated by the module 57 is an integer delay, there is no need to adjust the signal.

To match each pitch harmonic to an integer sample of the adjusted signal, the integer p is
The factor of the dimension N of the signal window generated by the module 10 must be: N = αp. Here, α is an integer. This dimension N must normally be a power of two for the realization of the FFT. In the example considered here, it is 256.

The spectral resolution Δf of the individual Fourier transform of the adjusted signal is given by the following equation: Δf = pf _p / N = f_p/ Α. Therefore, reduce p and minimize α
It is advantageous, but large enough for oversampling. F_e= 8 kHz and N = 256, in the example considered here for the parameters p and α
Table 1 shows the values selected.

[0140]

[Equation 89] Selected according to the value of the delay T _p supplied is performed by the module 70 by harmonic analysis module 57. Module 70 supplies the ratio K between the sampling frequencies to the three frequency changer modules 71,72,73.

The module 71 has a value S _{n, i} associated with the band i defined by the module 12,

[0142]

[Equation 90] , Α ′ _{n, i} ,

[0143]

[Equation 91] And H ¹ _{n, f} are converted to a modified frequency scale (sampling frequency f _e ). This transform simply extends band i by a factor K. The converted value is supplied to the harmonic protection module 56.

Thereafter, the latter module operates as before and supplies the frequency response H ² _{n, f} of the noise suppression filter. This response H ² _{n, f} is obtained in the same manner as in FIG. 1 (conditions (8) and (9)). However, the condition (9), except that the pitch frequency f _p = f _e / p is determined according to the value of the integer delay p supplied by the module 70. Module 70 also provides frequency resolution Δf.

Module 72 oversamples the frame of N samples provided by windowing module 10. Oversampling by rational coefficient K (K = K1 / K2)
First, oversampling is performed using the integer coefficient K1, and then undersampling is performed using the integer coefficient K2. This oversampling and undersampling by integer coefficients can be performed in a conventional manner by a bank of polyphase filters.

The conditioned signal frame s ′ provided by the module 72 contains KN samples at frequency f _e . The samples are sent to a module 75 that calculates the Fourier transform of the samples. The transformation can be performed based on two blocks of N = 256 samples. One block consists of the first N samples of a frame of length KN of the adjusted signal s', and the other block consists of the N samples after the frame. Therefore, the two blocks are (2-K) × 10
It has 0% overlap. A set of Fourier components S _{n, f} is obtained for each of the two blocks. The component Sn _{, f} is supplied to the multiplier 58. The multiplier multiplies those components by the spectral response H ² _{n, f} to provide the first noise suppressed signal spectral component S ² _{n, f} .

The component S ² _{n, f} is a module 6 for calculating a masking curve in the manner described above.
Sent to 0.

When the masking curve is calculated, the magnitude indicating the voicedness of the speech signal χ (formula (
13)) preferably takes the form χ = 1-H. Here, H is the entropy of the autocorrelation of the spectral component S ² _{n, f} of the adjusted signal in which noise has been suppressed.
The autocorrelation A (k) is, for example,

[0149]

(Equation 92) Calculated by the module 76 using

The module 77 then calculates the normalized entropy H and supplies it to the module 60 for calculating the masking curve (SA McCl
Ellan et al .: "Spectral Entropy: An Alternative Indicator for Rate Allocation?", Pr
oc. ICASP'94, pages 201-204).

[0151]

[Equation 93] Due to the signal conditioning and the noise suppression by the filter H ² _{n, f} , the normalized entropy H constitutes a very voicing measure against noise and pitch changes.

The correction module 62 operates in the same way as the device shown in FIG. 1, and the over-estimated noise rescaled by the frequency changer module 71.

[0153]

[Equation 94] Tolerate. It provides the last noise suppression filter frequency response H ³ _{n, f} . It is multiplied by the spectral component S _{n, f of} the signal conditioned by module 64. The resulting component S ³ _{n, f} is processed by the IFFT module 65 and returned to the time domain. The module 80 at the output terminal of the IFFT module 65
The two signal blocks resulting from the processing of the two overlapping blocks supplied by the FT 75 are combined for each frame. This combination can be comprised of a Hamming weighted sum of the samples to form a tuned signal frame with suppressed noise for the KN samples.

The module 73 changes the sampling frequency of the adjusted signal in which the noise supplied by the module 80 is suppressed. The sampling frequency is returned to F _e = f _e / K by an operation that is the reverse of the operation performed by module 75. Module 7
3 provides N = 256 samples per frame. N after preceding frame
/ 2 = after 128 superimposed addition reconstitution with specimens, only N / 2 = 128 samples at the beginning of the current frame is the last being held, the last signal s ³ which noise is suppressed Form (module 66).

[0155] In a preferred embodiment, is formed by the module 10, the window that has been held by the module 66 and module 82 are managed, the T _p = F _e / f number equal to an integral multiple of _p M samples Hold. This avoids the problem of phase discontinuities between frames. In a corresponding manner, the management module 82 sets the windowing module 10 so that the overlap between the current frame and the next frame corresponds to NM.
Control. This overlap of the NM samples is taken into account in the superposition and addition operation performed by module 66 when processing the next frame. From the value of T _p supplied by the harmonic analysis module 57, the module 82 calculates the number of samples to be retained, M = T _p × E [N / (2T _p )], where E [] indicates the integer part. And controls the modules 10 and 66 accordingly.

In the embodiment just described, the pitch frequency is estimated as an average over the frame. The pitch can change slightly over this duration. In order to obtain a constant pitch in the frame by artificial means, it is possible in the context of the present invention to allow those variations.

This requires that the harmonic analysis module 57 provide a time interval between successive breaks in the speech signal that can be due to the speaker's voiceprint closure occurring during the duration of the frame. Methods that can be used to detect such short breaks are well known in the art of harmonic analysis of speech signals. In this connection, the following paper can be referenced: M. BASSEVILLE et al., "Sequential detection of abstraction of sudden changes in the spectral characteristics of digital signals."
changes in spectral characteristicsof
digital signals), IEEE Tra about Information Theory
ns. 1983, Vpl. IT-29, No. 5,708-723; R
. ANDRE-OBRECHT, "New statistical approach for automatic segmentation of continuous speech signals (A new approach for the atom)
atic segmentation of continuous speed
ch signals), IEEE Trans. on Acous. , Sp
ech ad Sig. Proc. Vol. 36, no. January 19
88; MURGIA et al., “An algorithm for glottal closure estimation using sequential detection of sudden changes in speech signals (An algorithm for the same).
Estimation of total closure insta
nt using the sequential detectionof
abrupt change in speech signals), Sig
nal Processing VII, 1994, pages 1685-1688.

The principle of the above method is to perform a statistical test between the short-term model and the long-term model. Both models are adaptive linear prediction models. Statistical test values wm were corrected by Kullback divergence, 2
It is the cumulative sum of the recursive likelihood ratios of two distributions. For a distribution of residues with Gaussian statistics, the value w _m is

[0159]

[Equation 95] Given by Here e ⁰ _m and sigma ² ₀ represents the remainder calculated when the change of the sample m and long model frame, e ¹ _m and sigma ² ₁ likewise represents a change remainder short model. The closer the two models are, the closer the statistical test value w _m approaches zero. In contrast, the two models is to away from each other, the value w _m is negative. It indicates a break R in the signal.

Accordingly, FIG._mFig. 3 shows one possible example of the evolution of the, and shows a break R in the speech signal. Two pulls
Time interval t between successive breaks R_r(R = 1, 2, etc.) is calculated and expressed as the number of sample samples of the speech signal. Each interval t_rIs the pitch frequency f_pInversely proportional to
. Therefore, it is estimated locally: f at r-th interval_p= F_e/ T _r .

[0161] The time variation of the pitch (ie, the fact that the intervals _tr are not always all equal in a given frame) can then be corrected in order to obtain a constant pitch frequency in each analysis frame. This modification, the sampling frequency is performed by modifying standing each interval t _r yard, obtain a constant distance between two glottal closure after oversampling. Therefore, the duration between two breaks is modified by oversampling at a variable ratio to lock to the maximum interval. Also, adjustment constraints are satisfied such that the oversampling frequency is a multiple of the estimated pitch frequency.

FIG. 11 shows the means used to adjust the signal in the latter case. Supplying interval t _r of harmonic analysis module 57 using the above analysis, associated with the signal frame generated by the module 10. For each of those intervals, module 70 (block 90 of FIG. 11) calculates the _oversampling ratio K _r = p _r / t _r . Here, the integer p _r, if takes a value t _r is shown in the second column of Table 1 are given by the third column of Table 1. These oversampling ratio K _r, as interpolated by the sampling ratio K _r is performed over the corresponding time interval t _r, it is supplied to a frequency changer module 72 and 73.

The time interval t supplied by the module 57 for the frame_rLongest time interval T_pIs selected by the module 70 to obtain the pair p, α as shown in Table 1 (block 91 in FIG. 11). Then the modified sampling frequency
Is f as before_e= PF_e/ T_pIt is. The spectral resolution Δf of the individual Fourier transform of the adjusted signal is Δf = F_e/ (Α ・ T_p) Is still given by
. For the frequency changer module 71, the oversampling ratio K is K = p / T_p(Block 92). Module 56 for protecting pitch harmonics is
For case (9), the spectral resolution Δf provided by block 91 and the
Pitch frequency f determined according to the value of integer delay p supplied by lock 91 _p = F_eIt operates as before using / p.

This embodiment of the present invention also includes the application of the window management module 82. To be held over the current frame, in this case the number M of samples of the signal which the noise is suppressed, subsequent matching integer time interval t _r between two glottal closure (block 10). This avoids the phase discontinuity problems between frames, yet allowing the possible changes in the time interval t _r over the frame.

[Brief description of the drawings]

FIG. 1 is a block diagram of a noise suppression device that realizes the present invention.

FIG. 2 is a flow chart of a procedure used by the voiced activity detector of the device shown in FIG.

FIG. 3 is a flow chart of a procedure used by the voiced activity detector of the device shown in FIG.

FIG. 4 is a diagram illustrating a state of a voiced activity detection automaton.

FIG. 5 is a graph showing a change in voiced activity.

FIG. 6 is a block diagram of a module for overestimating noise of the apparatus shown in FIG. 1;

FIG. 7 is a graph showing calculation of a masking curve.

FIG. 8 is a graph showing the use of a masking curve in the device shown in FIG.

FIG. 9 is a block diagram of another noise suppression device which realizes the present invention.

FIG. 10 is a graph showing a harmonic analysis method that can be used in the method of the present invention.

FIG. 11 shows a part of a modification of the block diagram shown in FIG. 9;

[Procedural Amendment] Submission of translation of Article 34 Amendment of the Patent Cooperation Treaty

[Submission date] March 21, 2000 (2000.3.21)

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] Claim 5

[Correction method] Change

[Correction contents]

[Procedure amendment 2]

[Document name to be amended] Statement

[Correction target item name] Claim 22

[Correction method] Change

[Correction contents]

Is the long term estimator of the spectral component of the noise

22. A method as claimed in claim 1, obtained by combining ## EQU4 ## with a measure (ΔB ^max _{n, i} ) of the variability of the spectral component of the noise for a long-term estimate of the noise. Item 2. The method according to item 1.

[Procedure amendment]

[Submission date] September 13, 2000 (2000.9.13)

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] Claim 1

[Correction method] Change

[Correction contents]

[Procedure amendment 2]

[Document name to be amended] Statement

[Correction target item name] Claim 21

[Correction method] Change

[Correction contents]

(Equation 1) Each of the first quantities depending on parameters including the estimated pitch frequency (fp) is subtracted from each spectral component (S _{n, f} ) of the speech signal in the frame to reduce noise. A first subtraction step for obtaining a spectral component (S ² _{n, f} ) of the first signal; and an auditory model based on the spectral component (S ² _{n, f} ) of the first signal in which noise is suppressed. Calculating the masking curve (M _{n, q} ) by applying the following:-an overestimate of the corresponding spectral component of the noise for the frame

(Equation 2) Comparing with the calculated masking curve (M _{n, q} ); the corresponding first quantity and the overestimated part of the corresponding spectral component of the noise above the masking curve Is subtracted from the spectral component (S _{n, f} ) of the speech signal to obtain the spectral component (S ³ _n ) of the noise-suppressed second signal. 21. A method as claimed in any one of claims 1 to 20, comprising: a second subtraction step to obtain _f ).

[Procedure amendment 3]

[Document name to be amended] Statement

[Correction target item name] Claim 22

[Correction method] Change

[Correction contents]

(Equation 3) Is a long-term estimator of the spectral component of the noise

(Equation 4) 22. A method as claimed in any preceding claim, obtained by combining a measure of the variability of the spectral component of the noise (ΔB ^max _{n, i} ) for a long term estimator of the noise. the method of.

[Procedure amendment 4]

[Document name to be amended] Statement

[Correction target item name] Claim 26

[Correction method] Change

[Correction contents]

(Equation 5) And comparing the long-term estimator with the instantaneous estimator of the energy (E _{n, i} ) calculated for the frame, the speech activity of the speech signal for the frame n in frequency band i The method according to claim 24 or 25, wherein (γ _{n, i} ) is obtained.

[Procedure amendment 5]

[Document name to be amended] Statement

[Correction target item name] Claim 29

[Correction method] Change

[Correction contents]

[Procedure amendment 6]

[Document name to be amended] Statement

[Correction target item name] Claim 30

[Correction method] Change

[Correction contents]

[Procedure amendment 7]

[Document name to be amended] Statement

[Correction target item name] 0009

[Correction method] Change

[Correction contents]

[0009]

SUMMARY OF THE INVENTION Accordingly, the present invention provides: a harmonic analysis of a speech signal to estimate a pitch frequency of the speech signal in each frame characterized by internal speech activity; Calculate the spectral component of the speech signal for each frame, calculate the estimated amount of the spectral component of the noise contained in the speech signal for each frame, and calculate the noise component for the frame from each spectral component of the speech signal in the frame the estimated amount of the corresponding spectral component and the value of the estimated pitch frequency, including at least one step subtracting the amount of each depending on at least comprises parameters, cormorants line spectral subtraction, It proposes a method of suppressing noise in digital speech signals processed by successive frames. Results of the spectral subtraction is transformed to the time domain to form a Speech signal noise is suppressed.

──────────────────────────────────────────────────続 き Continuation of front page (81) Designated country EP (AT, BE, CH, CY, DE, DK, ES, FI, FR, GB, GR, IE, IT, LU, MC, NL, PT, SE ), OA (BF, BJ, CF, CG, CI, CM, GA, GN, GW, ML, MR, NE, SN, TD, TG), AP (GH, GM, KE, LS, MW, SD, SZ, UG, ZW), EA (AM, AZ, BY, KG, KZ, MD, RU, TJ, TM), AL, AM, AT, AU, AZ, BA, BB, BG, BR, BY, CA, CH, CN, CU, CZ, DE, DK, EE, ES, FI, GB, GE, GH, GM, HR, HU, ID, IL, IS, JP, KE, KG, KP , KR, KZ, LC, LK, LR, LS, LT, LU, LV, MD, MG, MK, MN, MW, MX, NO, NZ, PL, PT, RO, RU, SD, SE, SG, SI, SK, SL, TJ, TM, TR, TT, UA, UG, US, UZ, VN, YU, ZWF terms (reference) 5D015 CC03 CC14 EE05 FF03 5K046 AA05 HH11

Claims (28)

[Claims] And 1. A row cormorants step harmonic analysis of the speech signal to estimate a pitch frequency (f _p) of the speech signal in each frame, wherein the voice activity inside, for each frame Calculating the spectral components (S _{n, f} , S _{n, i} ) of the speech signal; and calculating, for each frame, the calculated value of the spectral component of the noise included in the speech signal; From each spectral component of the speech signal in the frame (S _{n, f} ), it depends on parameters including at least the estimated amount of the corresponding spectral component of noise for the frame and the estimated value of the pitch frequency. Performing a spectral subtraction that includes at least one step of subtracting the respective amounts to be subtracted; Constructing a noise-suppressed speech signal (s ³ ) by applying a time-domain transformation to the result, comprising: a method for suppressing noise in a digital speech signal (s) processed by a subsequent frame. . 2. Using the estimated pitch frequency (f _p ) to select a frequency to be protected from the set of frequencies for which the spectral components of the speech signal are calculated, _Given spectral component (S _{n, f} )
2. The method according to claim 1, wherein if the spectral component corresponds to a protected frequency, a smaller amount is used than the amount used when the spectral component does not correspond to a protected frequency. Method. 3. The frequency to be protected such that the spectral components of the speech signal corresponding to each of the protected frequencies exceed a noise level determined from a corresponding estimator of a corresponding spectral component of noise. Claim 2 for selecting
The described method. 4. Within the set of frequencies for which the spectral content of the speech signal is calculated, each protected frequency is closest to an integer multiple of the estimated pitch frequency (f _p ). A method according to claim 2 or 3, wherein Wherein said speech signal in the set of frequencies as the spectral components are calculated in for it, each frequency to be protected _{[η × f p -η × δf} p / 2,
closest to the frequency of the form in the range of _{η × f p + η × δf} p / 2], f p represents the frequency resolution of the previous SL pitch frequency estimated, eta is an integer, and claim 2 or 3, wherein the method of. 6. The method according to claim 2, wherein the amount subtracted from the spectral component (S _{n, f} ) of the speech signal at the protected frequency is substantially zero. 7. After estimating the pitch frequency (f _p ) of the speech signal in a frame, the speech signal of the frame is converted to an oversampling frequency (f _e ) that is a multiple of the estimated frequency. ) By oversampling it and calculating the spectral component (S _{n, f} ) of the speech signal in the frame based on the conditioned signal (s ′) 7. The method according to claim 1, wherein the amount is subtracted therefrom. 8. Spectral components (S _{n, f} ) of the speech signal are calculated by distributing the adjusted signal (s ′) into a block of N samples subjected to frequency domain transformation, and The method according to claim 7, wherein the ratio (p) between the normalized frequency (f _e ) and the estimated pitch frequency is a factor of the number N. 9. The voicedness (度) of the speech signal is calculated for the frame based on a calculation of the entropy (H) of the autocorrelation of the spectral components calculated based on the adjusted signal. The method according to claim 7 or 8, wherein the estimation is performed. 10. The spectral component whose autocorrelation (H) is calculated.
S ² _{n, f)} the method of claim 9, wherein calculating by the basis of the adjusted signal (s') after the amount of subtraction. 11. The voicing degree (χ) is given by:Where N is the spectral component (S) based on the adjusted signal (s'). _{n, f} ) Is used to calculate the number of samples, A (k),Normalized autocorrelation defined by^Two _{n, f}Is based on the adjusted signal
11. The method according to claim 9 or 10, which shows the spectral components of the hierarchy f calculated by means of:
. 12. After the processing of each frame, the pitch frequency (f) estimated as the sampling frequency (F _e ) among the noise-suppressed speech signal samples supplied by the processing. the method according to any one of claims 1 to 11 comprising the step of holding the number (M) of equal specimen to an integer multiple of the ratio (T _p) between _p). 13. The estimation of the pitch frequency of the speech signal in a frame comprises:-two consecutive breaks (R) of the signal which can be attributed to the glottal closure of the speaker in the frame.
Estimating the time interval (t _r ) during which the estimated pitch frequency is inversely proportional to the time interval; and interpolating the speech signal into the time interval and interpolating the same. 12. A method as claimed in any one of the preceding claims, wherein the adjusted signal (s') as a result of has a fixed time interval between two consecutive breaks. . 14. After the processing of each frame, the number (M) of samples of the noise-suppressed speech signal supplied by said processing corresponding to the estimated time interval (t _r ) 14. The method of claim 13, comprising the step of: 15. The signal-to-noise ratio of the speech signal in each frame is estimated in the spectral domain and the parameter on which the amount to be subtracted comprises the estimated signal-to-noise ratio, and over the frame A method according to any one of the preceding claims, wherein the amount subtracted from each spectral component of the speech signal is a corresponding estimated sub-function of the signal-to-noise ratio. 16. The method of claim 15, wherein said function decreases toward zero to maximize said signal to noise ratio. 17. The noise-suppressed signal spectral component (S _{n, f} ) obtained by subtracting the quantity from the speech signal spectral component (S _{n, f} ).
With S ² _{n, f),} the method according to any one of claims 1 to 16 to calculate a masking curve (M _{n, q)} by applying an auditory model. 18. The method according to claim 11 _{, wherein} the calculation of the masking curve (M _{n, q} ) uses the voicedness (度) measured by the normalized entropy H.
7. The method according to 7. 19. The parameter on which the amount is subtracted from the spectral component (S _{n, f} ) of the speech signal in a frame is an overestimate of the corresponding spectral component of the noise. ] Method according to claim 17 or 18, including the difference between the calculated masking curve ( _{Mn, q} ). 20. The overestimate of the spectral component of noise for a frame. Is compared with the calculated masking curve (M _{n, q} ) and the amount subtracted from the spectrum component (S _{n, f} ) of the speech signal to calculate the component (S ³ _n ) converted into the time domain. 20. The method of claim 19, wherein _f ) is obtained and is limited to a portion of an overestimate of a corresponding spectral component of the noise that is above the masking curve. 21. Spectral subtraction: An overestimate of the corresponding spectral component of the noise for said frame. Each of the first quantities depending on parameters including the estimated pitch frequency (fp) is subtracted from each spectral component (S _{n, f} ) of the speech signal in the frame to reduce noise. A first subtraction step for obtaining a spectral component (S ² _{n, f} ) of the first signal; and an auditory model based on the spectral component (S ² _{n, f} ) of the first signal in which noise is suppressed. Calculating the masking curve (M _{n, q} ) by applying the following:-an overestimate of the corresponding spectral component of the noise for the frame. Comparing with the calculated masking curve (M _{n, q} ); the corresponding first quantity and the overestimated part of the corresponding spectral component of the noise above the masking curve Is subtracted from the spectral component (S _{n, f} ) of the speech signal to reduce the time domain transformed noise of the second signal. the method according to any one of claims 1 to 20 comprising a second subtraction step of obtaining a spectral component (S ³ _{n, f),} the. 22. An overestimation of each of the estimators of the spectral components of the noise taken into account in the spectral subtraction, and an overestimation of each of the spectral components of the noise contained in the speech signal. ] Is a long term estimator of the spectral component of the noise 22. A method as claimed in any preceding claim, obtained by combining a measure of the variability of the spectral component of the noise (ΔB ^max _{n, i} ) for a long term estimator of the noise. the method of. 23. A long-term estimator of a spectral component of the noise in a frame n corresponding to a frequency included in a band i. Is given by Where γ _{n, i} denotes the non-binary speech activity of the speech signal, determined for the frame n for the frequency band i, and S _{n, i} denotes the frame n of frame n in band i. The method of claim 22, wherein the average of the spectrum amplitude of the speech signal is indicated, and λ _B is a forgetting factor. 24. An estimate of the noise obtained during at least one preceding frame. By performing a priori noise suppression of the speech signal of frame n on the basis of and by analyzing the energy change of the a priori noise suppression signal, the speech activity γ _{n, i} for said frame n 24. The method of claim 23, wherein is determined. 25. The method of claim 24, wherein the speech activity (γ _{n, i} ) associated with frequency band i is a function that varies continuously from 0 to 1. 26. A long-term estimator of the energy of the a priori noise-suppressed signal. And comparing the long-term estimator with the instantaneous estimator of the energy (E _{n, i} ) calculated for the frame, the speech activity of the speech signal for the frame n in frequency band i The method according to claim 24 or 25, wherein (γ _{n, i} ) is obtained. 27. Long-term estimation of the noise frame For the noise, the measure of the variability of spectral components (ΔB ^max _{n, i} ) corresponding to the frequencies contained in band i, in which the speech signal is the speech activity in band i A given number that does not feature gender The difference calculated for the frame of A method according to any one of claims 23 to 26, which is a function of 28. Long-term estimation of the noise frame The measured value (ΔB ^max _{n, i} ) of the variability of the spectral components, corresponding to the frequencies contained in band i, of the noise for, where the speech signal is within band i A given number that does not characterize speech activity The largest difference calculated for the frame of S _{nk, f} indicates a spectral component corresponding to a frequency for frame _nk, and a frequency range [f (i−1), f (i)] corresponds to band i. 27. The method according to any one of claims 26.