AT509570B1 - METHOD AND APPARATUS FOR ONE-CHANNEL LANGUAGE IMPROVEMENT BASED ON A LATEN-TERM REDUCED HEARING MODEL - Google Patents

Description

Austrian Patent Office AT 509 570 B1 2011-12-15

description

METHOD AND APPARATUS FOR ONE-CHANNEL LANGUAGE IMPROVEMENT BASED ON A LATEN-TERM REDUCED HEARING MODEL

FIELD OF THE INVENTION

The present invention relates to a method for improving a broadband audio signal with background noise, and more particularly, to a noise canceling system, a noise canceling method, and a noise canceling program. More particularly, the subject invention relates to latency reduced single channel noise reduction, using subband processing based on fading characteristics of the human auditory system.

BACKGROUND OF THE INVENTION

Additional background noise in the voice communication systems reduces the subjective quality and intelligibility of the perceived voice. Therefore, speech processing systems require noise reduction techniques, e.g. Methods aimed at processing to eliminate or attenuate the noise level in a noisy signal and to improve signal-to-noise ratio (SNR) without compromising the speech or its characteristics. Noise reduction is also commonly called noise suppression or speech enhancement.

For example, mobile phones are often used in environments such as public places with high background noise. The use of mobile phones and voice-activated devices and communication systems in cars has created a great need for hands-free installations for increasing safety and comfort in the car. For example, in many states and regions, the law prohibits the hand-held telephone in the car. Noise reduction becomes important for these applications, as their applications are necessary in acoustically unfavorable environments, especially at low SNR and high time varying noise floor characteristics, e.g. Rolling noise of cars.

In (hands-free) applications for teleconferencing, such as videoconferencing or speech recognition and interrogation systems, the background noise of fans from computers, printers or fax machines, which may be considered stationary (long-term). Conversation noise from (telephone) calls coming from colleagues sharing the room is often referred to as babble noise and consists of harmonic components and is therefore more difficult to attenuate by a noise reduction unit.

However, applications in hearing aids and car voice communication systems require noise suppression methods that can be performed in real time.

Nevertheless, the rapid development of the underlying hardware in terms of processing power and storage capacity supports the progress of the software implementations.

One of the most common methods of noise suppression in application-related applications is referred to in the jargon as spectral subtraction (see SF Boll, " Suppression of Acoustic Noise in Speech Using Spectral Subtraction, " IEEE Trans. Acoust. Speech and Sig. Proc., Vol. ASSP-27, pp. 113-120, Apr. 1979). In general, the spectral subtraction approach estimates the short-term spectral amplitude (STSA) of the clear speech from a disturbed speech signal, e.g. the desired noise-contaminated speech by subtraction of an estimated noise signal. Based on the assumption that the human ear is insensitive to phase distortions, the estimated magnitude of the speech signal is combined with the phase of the perturbed signal (see CL Wang et al., &Quot; The unimportance of phase in speech enhancement ", IEEE Trans Acoust Speech and Sig. Proc., 1/31 Austrian Patent Office AT 509 570 B1 2011-12-15 vol. ASSP-30, pp. 679-681, Aug. 1982). In practice, the spectral subtraction is accomplished by the multiplication of the input signal spectrum with a weighting function so as to suppress low SNR frequency components. This SNR-based weighting function is formed by estimates of the noise spectrum and the disturbed speech spectrum is broadly assumed to be stationary and the mean-free random signals, speech and noise signals to be uncorrelated. These conventional spectral subtraction methods offer significant noise suppression with the major disadvantage of reducing signal quality, perceived acoustically as musical sounds or musical noise. The musical sounds come from the spectral estimation errors. In recent years, many improvements of the simple spectral subtraction approach have been developed.

An often used method to reduce the musical sounds is to subtract an over-estimated noise spectrum to reduce the fluctuations in the DFT coefficients and to prevent the spectral components from going below a spectral lower limit (cf. M. Berouti et al., &Quot; Enhancement of speech, corrupted by acoustic noi-se, " in Proc. IEEE Int Conf. On Acoust., Speech and Sig. Proc. (ICASSP'79), vol. 4, pp 208-211, Washington DC, Apr. 1979). This approach successfully reduces the musical sounds at poor SNR ratios and periods of only noise. The main disadvantage is the distortion of the speech signal during speech. In practice, a compromise has been found between voice quality and the residual noise level. Other methods overcome this problem by introducing optimal and adaptive over-subtraction factors for poor SNR ratios and suggest sub-subtraction of the noise spectrum for good SNR ratios (see WM Kushner et al., The effects of subtractive-type speech enhancement / IEEE Int Conf Conf Acous-tics, Speech and Sig. Proc. (ICASSP'89), vol. 1, pp. 211- 214, 1989).

The use of a soft-decision based modification of the spectral weighting function (see R. McAulay and M. Malpass, " Speech en-hancement using a soft-decision noise suppression filter, " in IEEE Trans. Acoust., Speech and Sig. Proc, vol. 28, no. 2, pp. 137-145, 1980) has shown improvements in the noise cancellation characteristics of the amplifier system with respect to the suppression of musical sounds. These soft decision approaches depend mainly on the a priori likelihood of the lack of speech in each spectral component of the disturbed speech.

The smallest mean square deviation of the short-term spectral amplitude estimator (MMSE-STSA, see Y. Ephraim and D. Malah, " Speech enhancement using a mini-mean-square error short-time estimator, " IEEE Trans Acoust, Speech and Sig. Proc., Vol. 32, no. 6, pp.1109-1121, 1984) and the least mean square deviation of the logarithmic amplitude spectral estimator (MMSE-LSA, Y. Ephraim and D. Malah, " Speech enhancement using a minimum mean-square error log spectral amplitude estimator, " IEEE Trans Acoust Speech and Sig. Proc, vol., 33, no. 2, pp. 443-445, 1985) minimizes the corresponding mean square deviation the estimated short-term spectral or logarithmic spectral amplitude. It has been recognized that the non-linear smoothing process of the MMSE-SP / LSA methods (the so-called decision-driven approaches) achieves a uniform estimation of the SNR that provides good noise suppression without unpleasant musical sounds (see O. Capp, " Elimination Ephraim and Malah noise suppressor " IEEE Trans. Speech and Audio Proc., vol. 2, no. 2, pp. 345-349, 1994). Both: Capp and Malah (See E. Malah et al., "Tracking speech-presence uncertainty to improve speech enhancement in non-stationary noise environments," in Proc. IEEE Int. Conf. Acoust., Speech and Sig. Proc. (ICASSP'99), vol. 2, pp. 789-792, 1999) propose limiting a priori SNR estimation to overcome the problem of perceptible low level musical noise during pauses in speech. The so-called a priori SNR provides the information about the unknown 2/31 Austrian Patent Office AT 509 570 B1 2011-12-15

Spectrum range collected from the previous frames and evaluated in the decision-driven approach (DDA). Because the smoothing performed by the DDA has irregularities, the musical noise may occur at a low level. A simple solution to this problem is to constrain the a priori SNR by a lower bound.

In single channel spectral subtraction, the noise spectrum is normally estimated during the speech pause that requires speech activity detection (VAD) methods (see R. McAulay and M. Malpass, "Speech enhancement using a soft-decision noise suppression filter" in IEEE Trans Acoust., Speech and Sig. Proc., Vol. 28, No. 2, pp. 137-145, 1980; and WJ Hess, "A pitch-synchronous digital feature extraction system for phonemic recognition of speech", in IEEE Trans. Acoust., Speech and Sig. Proc., Vol. 24, no. 1, pp. 14-25, 1976). This approach implies static noise characteristics during the periods of speech. Arslan et al. developed a robust noise estimation method that does not require speech activity detection methods due to the recursive averaging of level-dependent time constants for each subband (see L. Arslan et al., New Methods for Adaptive Noise Suppression, in Proc. Int. Conf. on Acoustics, Speech and Sig. Proc. (ICASSP-95), Detroit, May 1995). Martin proposes a noise estimation method based on minimum statistics and optimal power spectrum density smoothing (PSD, see R. Martin, "Noise power spectral density estimation based on optimal smoothing and minimum statistics," in IEEE Trans Audio Proc., Vol. 9, no. 5, pp. 512, July 2001). Furthermore, Ealey et al. a method for estimating non-stationary noise during the duration of the spoken words by using the harmonic structure of the spoken speech spectrum, also known as harmonic tunneling (see D. Ealey et al., "Harmony tunneling: tracking non-stationary noises during speech, " in Proc. Eurospeech Aalborg, 2001). Furthermore, if information from soft decisions is used, it is suggested by Sohn and Sung that the noise spectrum be continuously adapted whether speech is not present (see J. Sohn and W. Sung, " A voice activity detector employing soft decision based noise spectrum adaptation, " in Proc. IEEE Int Conf. Acoustics, Speech and Sig. Proc. (ICASSP'98), vol. 1, pp-365-368, 1998).

Ephraim and Van Trees propose another important signal subspace resolution based noise suppression method (see Y. Ephraim and HL Van Trees, " A Signal subspace approach for speech enhancement ", in IEEE Trans. Speech and Audio Proc, vol 3, pp. 251-266, July 1995). In this case, the noisy signal is decomposed into a signal plus noise subspace and a noise subspace, these two subspaces are orthogonal to each other. This makes it possible to estimate the clear speech signal from the used signal. The resulting linear estimator is a generic Wiener filter with adjustable noise level to adjust the trade-off between signal distortion and residual noise because they can not be simultaneously minimized.

Skoglund and Kleijn show the importance of temporarily masking out characteristics associated with feeding the spoken language (see J. Skoglund and WB Kleijn, " On Time-Frequency Masking in Voiced Speech ", IEEE Trans Audio Proc., Vol. 8, no. 4, pp. 361-369, July 2000). It is shown that noise between two feed pulses is perceived more strongly than noise near the pulses and this is especially the case for low word density speech for which the feed pulse is temporarily sparse. Temporary fading is not used by conventional noise reduction methods that use a frequency domain estimator. WO 2006 114100 discloses a signal subspace approach that takes into account temporary fading characteristics.

SCOPE AND SUMMARY OF THE INVENTION

The object of the present invention is to provide a single-channel-based noise cancellation method with latency-reduced processing of a broadband speech signal in the presence of background noise. In particular, the present invention is based on a spectral subtraction method using a modified decision-driven approach, including over-subtraction and adjustable noise level to avoid perceptible musical sounds. Furthermore, the present invention uses pre-filtering and post-filtering subband processing to account for temporary and simultaneous fading associated with human perception, particularly to minimize perceptible signal distortion during speech periods.

Frequency domain processing is performed by the proposed system, which is subdivided by means of a nonuniform gammaton filter bank (GTF), which is also called critical bands, often called bark bands. This analysis filter bank divides the noisy signal into a plurality of overlapping narrowband signals, taking into account the spectral (simultaneous) blanking characteristics of the human auditory senses.

A preprocessing unit that replicates the transfer behavior of the human outer and middle ear is applied to the time-discrete noisy input signal (e.g., to the desired speech-contaminated speech).

In each subband, the level of the noisy signal is detected and smoothed. These narrow-band level detectors are applied to a plurality of subbands to exploit the phase of the simple filter parts and to obtain the shortest signal processing times.

From the smoothed envelope of the subband signals, the noise floor is estimated using a heuristic recursive minimum-statistic based approach for each subband.

The immediate signal to noise ratio (SNR) for each subband is estimated from the envelope of the noisy signal and the noise floor.

The a priori SNR is estimated from the immediate SNR through the use of the Ephraim-Malah Spectral Subtraction Rule (EMSR). To minimize the influence of estimation errors, an improved decision-driven approach (DDA) is proposed that introduces an underestimation parameter and a lower noise floor parameter.

The temporal hiding based on the human auditory sense is taken into account by adequately filtering the subband signals. These nonlinear auditory fade filters apply recursive averaging to falling edges of the signal levels detected in each subband; with the following effects: (a) overestimation variances of the jerky noise, (b) noise cancellation algorithms have no effect on signal below the timeout boundary, and (c) no additional signal delay is caused to transient signals that are important to speech perception.

A non-linear weighting function for each subband is estimated from the a priori SNR which includes oversubstraction of the estimated noise signal.

The disturbed signal in each subband is multiplied by a corresponding weighting factor to suppress the noise signal components.

An optimized, near-perfect reconstruction filter bank employs a signed-summation decision criterion for restoring the improved full-band speech signal.

Finally, a post-filter is applied to the improved full-band signal to compensate for the effect of the pre-filter.

Comments: The noise suppression methods cited above operate in the frequency domain and use the Discrete Time Fourier Transform (DTFT) which is based on block processing of the time discrete input signals. This block processing adds a frame size dependent signal delay. [0027] Austrian Patent Office AT 509 570 B1 2011-12-15 Subtraction type single-channel speech enhancement systems are efficient in reducing background noise; However, they contain noticeable, annoying residual noise. To cope with this problem, the characteristics of the hearing system are introduced into the amplification process. This phenomenon is modeled by calculating the Noise Blanking Limit in the frequency domain under which all components are inaudible (see N. Virag, " Single Channel Speech Enhancement Based on Masking Properties of the Human Auditory System ", IEEE Trans, on Speech and Audio Proc., Vol. 7, no. 2, pp. 126-137, March 1999).

In order to model the echo suppression in subtractive-type speech enhancement systems, filter bank implementations are particularly attractive because they can be adapted to the spectral and temporal resolution of the human ear. The authors propose a noise suppression method based on spectral subtraction combined with decomposition into critical bands Gammaton Filter Banks (GTF). The concept of critical bands describing the resolution of the human auditory system results in a nonlinear frequency scale, the so-called Bark scale (see JO Smith III and JS Abel, " Bark and ERB Bilinear Transforms ", IEEE Trans, on Speech and Audio Proc., vol. 7, no. 6, pp. 697-708, Nov. 1999).

The use of the Gammaton filterbank outperforms the DTFT based approaches in terms of computational complexity and overall system delay time. However, the GTF approaches allow short-term runs, low-computational complexity, and near-perfect reconstruction analysis-synthesis schemes. The proposed system filter generates the wideband output signal by simply summing the subband signals, introducing a criterion of the need to change the sign before summation. This approach outperforms the vocal-based approaches proposed by McAulay and Malpass (see RJ McAulay and ML Malpass, " Speech Enhancement Using a Soft-Decision Noise Suppression Filter ", IEEE Trans, on Acoust., Speech and Sig. Proc., Vol. ASSP-28, no. 2, pp. 137-145, April 1980). In this approach, the fullband reconstruction of the output signal is accomplished by summing alternate out-of-phase subband signals without regard to the real phase relationship between subbands. This brings big distortions to the output signal.

Important Note: Subband signals without downsampling, as often used in hearing aids, do not require a synthesis filter bank. Therefore, this approach is applicable to delay-reduced speech enhancement systems, but is computationally highly inefficient. The method proposed by the authors allows the calculation of the output signal from the subband signals by simple summation taking into account the phase differences! It is worth mentioning that there are many applications, such as hearing aids or hands-free kits in cars, where the computational complexity and signal delays are of utmost importance.

The main advantages of the present invention compared to conventional noise reduction techniques are the significant improvements in overall signal delays and computational efficiency.

The invention is not limited by the following embodiment. It is only intended to explain the inventive concept and to illustrate a possible application.

According to the invention, the method for reduced-time, auditory model-based single-channel noise reduction and reduction operates as an independent module and is intended for installations in digital signal processing chains, wherein a software-specified algorithm into a commercially available digital signal processor (DSP), in particular a DSP for audio applications. Remarks: The amplitude of the clear speech signal is estimated using the spectral ephraim-malah subtraction rule (EMSR) from the given amplitude of the noisy signal and the estimated noise variance. To avoid artifacts such as musical noise, modified decision-driven approaches (DDA), the over-subtraction (underestimation) of the noise variance are introduced with a lower noise level parameter.

In comparison to the prior art, both document for publication and in their synopsis, the solution is new and inventive. The essential difference is that the illustrated auditory gammaton analysis filter bank for subband decomposition of the input signal performs an additional phase shift on the subbands, by which an improved reconstruction of the output signal is achieved by simple and computationally summing the sub-sampled subband signals in the time domain.

Furthermore, in comparison with the present state of the art, the following difference can be established: The illustrated method for noise estimation using sensed threshold values is particularly efficient with regard to the memory requirement of the recursive embodiment. Recursive methods usually have a high degree of robustness and stability, but these advantages are always associated with a high memory requirement and a high computing time. The combination of intermeshing processing stages (means at small signal levels - keeping the estimates at high but temporally limited signal levels - overestimating at sustained high signal levels) is not yet available in this form in the prior art. Due to the recursive structure of the signal processing algorithms, a long storage of signal level values can be avoided and the required amount of computation can be minimized.

BRIEF DESCRIPTION OF THE DRAWING FIGURES

FIG. 1 is a schematic diagram of a one-channel subband speech enhancement unit of the present invention. FIG.

Fig. 2 is a schematic representation of the nonlinear calculation of the gain factor for the noise reduction used for each subband.

Figures 3 and 4 show the roof-shaped MMSE-SP attenuation area as a function of the a posteriori (yk) and the a priori (ξϋ SNR. To cover all values 0 <yk <oo, the x-axis refers to yk and not (yk -1) as in the literature The dot-dash line in Fig. 3 marks the transition between the regions and, the dotted line shows the spectral attenuation subtraction contour. 4 above the MMSE-SP suppression surface The dashed lines in Fig. 4 show the average of the dynamic relationships between yk and The solid lines show the static conditions.

Figs. 5 and 6 are illustrations of the combined (modified) DDA and MMSE SP

Assessment behavior. The dashed lines in Fig. 5 show the average of the dynamic relationship between yk and ξι &lt;. The solid lines show the static conditions. Two fictitious hysteresis loops in Figure 6 match the observations of informal experiments.

Fig. 7 shows a block diagram of the complete system.

FIG. 8 shows the complete system comprising a hearing frequency analysis and a reassembly as input and output, as well as a special delay time-reduced voice amplification with little effort between them

Austrian Patent Office Figure 9 Figure 10 [0046] Figure 12 [0048] Figure 13 [0049] Figure 14 AT 509 570 B1 2011-12-15. A combination of a sophisticated noise suppression law with a human auditory model enables high quality performance. shows an outer ear and a middle ear filter composed of three second order SOS sections. shows an example: three-zero gamma-tone filter of order 3. The common zero at z = 1 is not shown in this figure. shows a known type of level detection. When using the signal power, the square of the amplitude is detected. shows the runtime reduced FIR level detector. shows a non-linear recursive post-masking auditory filter that responds to falling edges. shows a recursive noise floor estimator that uses three time constants and a counter threshold.

DETAILED DESCRIPTION

In this description, new aspects are presented concerning the Ephraim Malah Noise Suppression Rule (EMSR) and the Decision Based Approach (DDA) for a priori S / N estimation. After splitting the range of the amplitude estimator, it becomes clear that the combined DDA estimation follows an unconfigured hysteresis cycle. The introduction of a hysteresis full-width paremeter improves the hysteresis and reduces the musical noise. Finally, we get a more flexible noise canceler with less dependency on the system's sample rate.

I. INTRODUCTION

The Ephraim-Malah Amplitude Estimator and the Decision-Driven Ephraim-Malah a priori SNR Estimate (Y. Ephraim and D. Malah, Speech Enhancement Using a Minimum Mean-Square Error Short-Time Spectral Amplitude Estimator, IEEE Transactions on Acoustics, Speech, and Signal Processing, No. 6, vol. ASSP-32, pp. 1109-1121, Dec. 1984 and Y. Ephraim and D. Malah, &quot; Speech Enhancement Using a Minimum Mean Square Error Log. Spectral Amplitude Estimator ", IEEE Transactions on Acoustics, Speech, and Signal Processing, No. 2, vol. ASSP-33, pp. 443-445, Apr. 1985.) are powerful noise suppression tools in speech signal processing. At present, there is a great deal of recent work published on both topics, as the combined algorithm is a powerful tool on the one hand (O. Cappe, "Elimination of the Musical Noise Phenomenon with the Ephraim and Malah Noise Suppressor", IEEE Transactions on Speech and 2, vol. 2, pp. 345-349, Apr. 1994), but on the other hand simplifications (PJ Wolfe and SJ Godsill, "Simple Alternatives to the Ephraim and Malah Suppression Rule for Speech Enhancement", Proc 11th IEEE Signal Processing Workshop, pp. 496-499, 6-8 Aug. 2001) as well as further developments (I. Cohen and B. Berdugo, &quot; Speech Enhancement for non-stationary noise environments &quot;, Signal Processing, no. pp. 2403-2418, Elsevier, Nov. 2001; I. Cohen, &quot; Speech Enhancement Using a Noncausal A priori SNR estimator &quot;, IEEE Signal Processing Letters, no. 9, pp. 725-728, Sep. 2004; Cohen, &quot; Relaxed Statistical Model for Speech Enhanc ement and A Priori SNR Estinnation, Center for Com- munication and Information Technologies, Israel Institute of Technology, Oct, 2003, CCIT Report no. 443; Μ. K. Hasan, S. Salahuddin, M.R. Khan, &quot; Modified A Priori SNR for Speech Enhancement Using Spectral Subtraction Rules &quot;, IEEE Signal Processing Letters, vol. 11, no. 4, pp 450-453, April 2004) are desirable.

In the amplitude estimation part of the algorithm, a signal model is used in which a noise signal y [n] consisting of speech x [n] and additive noise d [n], the time index n. The signals x [n] and d [n ] are considered statistically independent 7/31 Austrian Patent Office AT 509 570 B1 2011-12-15

Gaussian random variables are assumed. Because of certain properties of the Fourier transform, the same statistical model for the corresponding short-term spectral amplitudes Xk [m] and Dk [m] can be assumed in each frequency interval k at the time of analysis m. Xk [m] is a complex variable in our notation.To simplify the notation, let Xk [m] represent the amount | Xk [m] |). For given speech and noise variations, a2xk and σ ] k, the speech amplitude Xk [m] can be estimated from the noisy language Yk [m].

A suitable estimator [m] for the clear speech amplitude is described in section I-A.

The unknown variances of the clear language a2xk are implicitly determined in the a priori SNR estimation part of the algorithm, wherein the noise variance adk is to be determined in advance, e.g. through the use of minimum statistics (PJ Wolfe and SJ God-sill, "Simple Alternatives to the Ephraim and Malah Suppression Rule for Speech Enhancement", Proc. 11th IEEE Signal Processing Workshop, pp. 496-499, 6-8. Aug 2001), MCRA (Cohen and B. Berdugo, &quot; Speech Enhancement for non-stationary noise environments &quot;, Signal Processing, no. 11, pp. 2403-2418, Elsevier, Nov. 2001) or harmonic tunneling (D. Ealey, H. Kelleher, D. Pearce, &quot; Harmony Tunneling: Tracking Non-Stationary Noises Düring Speech &quot;, Proc. Eurospeech, 2001).

The decision-controlled estimation, described in Section l-B, determines the a priori SNR ξ1ζ = σ2χ1ζΙ a2dk in each frequency interval k. In addition, the noise canceler uses an immediate estimate, the so-called a posteriori SNR estimator, which relates the square of the current noise floor to the noise variance 7k [m] = Yk [m] la2dJi.

Section II gives an overview of the combined estimate and presents the hysteresis form. Subsequently, Section III shows how a small modification can reduce unwanted estimation behavior and allow smoother hysteresis. A. THE EPHRAIM-MALAH SUPPRESSION GESTURE (EMSR) As described above, the EMSR reconstructs the magnitude of the clear speech signal Xk \ m \ from the noisy observation Yk [m], because the amounts are different

Times m were assumed to be statistically independent, the time index m can be omitted for simplicity of notation.

The MMSE SA estimator by Ephraim and Malah (Y. Ephraim and D. Malah, &quot; Speech Enhancement Using a Minimum Mean-Square Error Short-Time Spectral Amplitude Estimator &quot;, IEEE Transactions on Acoustics, Speech, and Signal Processing, No. 6, vol., ASSP-32, pp. 1109-1121, Dec. 1984) solves the Bayesian formula Λ = 4ν, ΐη} by the amplitude of the clear ones

Estimate language Xk. When different distortions are applied to the amplitude, other estimators are derived in a similar way, e.g. MMSE-LSA Y. Ephraim and D. Malah, "Speech Enhancement Using a Minimum Mean-Square Error Log-Spectral Amplitude Estimator", IEEE Transactions on Acoustics, Speech, and Signal Processing, no. 2, vol. ASSP-33, pp. 443-445, Apr. 1985) Xk = eE ^ Xt ^ t \ and the MMSE-SP by Wolfe and Godsill (P.J.

Wolfe and S.J. Godsill, "Simple Alternatives to the Ephraim and Malah Suppression Rule for Speech Enhancement", Proc. 11th IEEE Signal Processing Workshop, pp. 496-499, 6-8. Aug 2001) Xk = yjE {x2 \ Yk \. For a more detailed description, see Cohen (I. Cohen, &quot; Relaxed Statistical Model for Speech Enhancement and Priori SNR Estimation &quot;, Center for 8/31 Austrian Patent Office AT 509 570 B1 2011-12-15

Communication and Information Technologies, Israel Institute of Technology, Oct, 2003, CCIT Report no. 443).

According to Ephraim and Malah, the noisy phase is an optimal estimate of the clear phase. Therefore, the reconstruction operator is a real-valued spectral weight G [m]: G [m] Kk [m]

Xk [m] Yk [m] G [m \ -Zfc [m]. (1) (2) Because of its simplicity, we have chosen the MMSE SP (3) from Wolfe and Godsill as the basis for our consideration. The corresponding weight rule can be specified as follows:

GMMSE sp [m]

(3) using the Wiener filter equation (4) _ χΑ = 6

Grv - 1 + & 'To simplify the application, we decompose the reconstruction operator into some regions • (7k - 1) &lt; 1/6: Gmmse-sp • (7¾ - 1) 1/6: Gmmse-sp ~ Gw • - 1) = 1/6: GmMSE-SP => / Gw 2 / 7fc- [0061] Additionally, we can do that Wiener filters by • 6 ^ 1 'Gw ~ 6 • 6 ^ 1: Gw ~ 1 approximate. With the combination of both, we can logarithmically decompose the MMSE SP area into flat parts (see also Fig. 3): 1) (7fe-1) <C 1/6, ξΐι <1. 1 = &gt; Gmmse-sp «\ / 6 / 7fc 2) (7¾ - 1) &lt; 1/6, ξ * »1 = &gt; Gmmse-sp ~ \ fillk 3) (7k - 1) &gt; 1/6 »6 &lt; 1 = * Gmmse-sp «6 4) (7 *, - 1) &gt; 1/6, 6 &gt; 1 = &gt; Gmmse-sp ~ 1 In the following sections we use the short form G when referring to GMmse-sp. B. THE DECISION-DRIVEN APPROACH (DDA) The DDA combines two simple SNR estimators into a new estimator for a priori SNR ξκ.

The first estimator is the immediate SNR. With only positive SNR values, one obtains (ni) = k-9/31 Austrian Patent Office AT 509 570 B1 2011-12-15 (5) SNRinst = max (7fc-1.0 ), Which can be calculated before the noise suppression. This immediate SNR differs from the true SNR in the following cases: if the analysis time window is too short as to the stationarity of the signals x [n] and d [n], if a nonstationary noise is not can be identified in detail, or · if noise and speech signal are highly correlated.

The second order estimator describes the recovered SNR which is calculated after noise suppression as follows: SNRrec = n = 1k-G2 (6) 'd, k For poor SNR ratios, e.g. 0 &lt; 7k &lt; 2, the a posteriori SNR 7k shows relative variations with time smaller than those of (7k - 1). (Relative variations, eg 10 log (7k [m]) - 10 log (7k [m-1]), are more significant than linear variations in human hearing.) Ideally, G 5 provides consistently high attenuation for poor SNR ratios , Therefore, the recovered SNRrec gives more consistent values than SNRinst in bad SNR cases.

Finally, to estimate a priori SNR SNRinst and SNRrec, the DDA combines: (7) £ fc [m] = (1-a) * SNRin8t [m] + a * SNRrec [τη-1] - [0073] The specific properties of the estimator can be observed upon onset of suppression enhancement in the DDA.

II. Combination of DDA and EMSR

Inserting the parts of the reconstruction operator Gmmse-sp by Wolfe and Godsill from section IA into the DDA equation (7) of Ephraim and Malah gives the following ranges of action for the combined a priori SNR estimation: 1) (7k-1) &lt; 1 / f, 6 &lt; 1. Gαy / ξΐί / 7fc £ fc [m] «(1 - a) · max (qfk [m] - 1,0) + (8) a - £ fc [m - 1]. 2) (7fc - 1) <C 1 / &, 6 »1, G« y / lpik £ fc [ra] «(1 - a) · max (7 * [πι] - 1,0) + a (9) ~ a. 10/31 Austrian Patent Office AT 509 570 B1 2011-12-15 3) (7fc - 1) »1 / &, ξ * [τη]« (1 - a) max (7 * [m] - 1,0) + or £ * [m - 1] - 7 * [m - 1] (10)

(ID «(1 - a) (7k [m] - 1).« (1 - a) · max (7fc [m] -1,0) + a 7k [m - 1] «a-7 *; [ ml]. 5) (7fc-1) = 1/6, &amp; &Lt; 1 = ► G => / 2 · efc / 7fc w (1-a) · (TfcH "1) + (12) 2a ^ fc [m-1].

The characteristic of the combined approach can be considered in FIG. Considering the amplitude of the speech signal and a constant noise level, e.g. a time-varying a posteriori SNR 7k as input sequence, one can imagine a kind of hysteresis loop development on the MMSE SP area. Besides obvious discontinuities in this loop, other properties are shown (O. Cappe, &quot; Elimination of the Musical Noise Phenomenon with the Ephraim and Malah Noise Suppressor &quot;, IEEE Transactions on Speech and Audio Processing, No. 2, vol. 2, pp. 345-349, Apr. 1994).

A. RECURSIVE AVERAGE EDUCATION

1) EXPECTATIONS OF RECURRENT AVERAGING: In the above enumeration it can be seen that the a priori SNR estimate in part 1 corresponds to the recursive mean (8) of the immediate SNRinst (5). It is possible to generalize the averaging process by introducing a time constant Tavg, which determines the mean parameter α = exp [-1 / (Tavg fs)]. Here, the sampling rate fs = 1 / T denotes the number of time-frequency transformations per second.

2) THE CONSTANT EFFECT: If the a priori SNR ξκ has a constant value in part 1, e.g. in the case of large time constants Tavg or at the edges of the ξκ value range, the estimator could work strangely. At small and constant ξκ, the system keeps the output at a constant level. This happens when the entrance is small enough (κ * 2 [ι »] / σ ^ -l)« l / £ t ^> Yk2 [m] «G2dk / Gw (using (8) and its presuppositions ):

11/31 (13) [0079] In certain circumstances, the disturbing, additional broadband Störteisches Patentamt AT 509 570 B1 2011-12-15 may cause noises which may be worse than a constant output which, because of the limitation of 6 to a minimum ζ for F / [m] &lt; a] k I ζ is caused.

3) INSTABILE RECURSIVE AVERAGE FORMATION: Following (12), part 5 may result in a priori SNR estimation by unstable recursive averaging of SNRins, if α &gt; 1/2, e.g. 6 can suddenly rise in this part.

B. PARTS WITHOUT RECURRENT AVERAGE EDUCATION

In parts 2, 3, and 4 the interpretation of the recursive averaging is not useful. Namely, in (9), the a priori SNR estimate 6 takes a constant value, and in (10), ξκ is determined by a simple delay time. It is strange that SNR ξκ is in (10) reduced version of SNRins.

C. SUMMARY OF PROPERTIES

Actually, every part other than 1 and 4 (Eqs. (8) and (11)) has unexpected behavior. The definition of α by a time constant yields generalized averaging properties of (8), whereas a sample-rate-dependent behavior is obtained by the Eqs. (9) - (12) defined estimate. This form of sample rate excludes a generally fitting parameter set for different time step analysis and transformation quantities.

Unfavorable estimation behavior, e.g. the &quot; constant &quot; effect &quot;, and the discontinuities in the hysteresis loop (Figure 4) increase the consideration of modifying the DDA and retesting current and minimum a priori SNR magnitudes.

III. A MODIFY, QUICKLY ANSWERING DDA

To minimize the influence of unexpected estimators, the decision-driven approach is modified: £ * [m] = (1-a) * (p * SNRmstfwi] + C) + &amp; · SNRrec [m-1], (14) with ζ as the lower noise level parameter (O. Cappe, "Elimination of the Musical Noise Phenomenon with the Ephraim and Malah Noise Suppressor", IEEE Transactions on Speech and Audio Processing, no. 2, vol. 2, pp. 345-349, Apr. 1994) and p and underestimation parameters of the immediate SNR. Similar to the parts in Section II one can find: 1) (tk - 1) «1/6. &Amp; C 1, G »VWrÄ £ fc [ra]« p (1 - α) · max (7fc [m] - 1,0) + a 6 [m - 1]. (15) 2) (lk-1) &lt; 1/6, 6 »1, G« y / l / lfk (16) 6H ~ «· 3) (7 * -1) &gt; 1 / 6.6 &lt; ~ 6 6 [m] «ρ (1 - a) (7fcH - 1) · U7) 4) (7 * -1) &gt; 1 / 6,6 »1, G« i (18) 6M ~ a · 7k [m - 1]. With regard to the divisions of the new estimator, one can consider the scheme of the total estimator in FIG. 5. Instead of the time constant in the quasi-stationary domain of the language Tavg = 2 ms is now used, p = 10'15 / 10 guarantees that the scaling factor in (17) is approximated by p (la) «p, which fixes the discontinuities in the estimation hysteresis , You can choose the lower noise level ζ = 10'2T ° so small that the maximum attenuation ζ is at the lower end of the dynamic range of the frequency interval. These measures largely reduce the sample rate dependence described in Section II-C and the &quot; Constant Effect &quot; from section II-A.2.

It will be appreciated that rising immediate SNRs are now better attenuated in Figure 5 than in Figure 4. Therefore, a strong attenuation for musical sounds, e.g. inconsistent high immediate SNR, while a signal of consistently high SNR can pass through the noise canceler. The two crimped loops in Figure 6 give an example of an approximated hysteresis loop during system operation.

The parameter p can directly control the suppression hysteresis width and the suppression of musical noise. Our modifications allow separate control of the averaged time constant and noise suppression.

IV. CONCLUSION

We have found a tractable way to graphically describe the properties of the spectral amplitude estimate of Wolfe and Godsill, as well as the decision-driven a priori SNR estimate of Ephraim and Malah. This description can similarly be used for other amplitude estimation rules and provides a new insight into the Ephraim and Malah noise cancelers.

So far, the suppression of musical noise has been a compromise between the suppression of musical noise and transient distortion. Small modifications in the decision-driven estimation rule allow for a more flexible handling of the suppression of musical noise, while reducing the dependencies of the time step analysis and the &quot; constant effect &quot;. An informal hearing test with modified algorithm and adjustable analysis time / frequency resolution (filter bank approach) already showed useful improvements in the overall system.

Our future work will turn our descriptive methods into more sophisticated estimation approaches by Cohen (I. Cohen, &quot; Speech Enhancement Using a Noncausal A priori SNR estimator &quot;, IEEE Signal Processing Letters, no. 9, pp. 725-728, Sep. 2004) or Hasan (Μ.K. Hasan, S. Salahuddin, MR Khan, &quot; Modified A Priori SNR for Speech Enhancement Using Spectral Subtraction Rules &quot;, IEEE Signal Processing Letters, vol. 11, no. 4, pp 450- 453, April 2004).

APPARATUS FOR RUNNING REDUCED ONE-CHANNEL LANGUAGE GAIN

Hereinafter, a preferred embodiment will be described, but the invention is not limited to this embodiment.

The reduction of musical noise in noise reduction algorithms is still a key issue for noise reduction. Although the Ephraim-Malah suppression rule (EMSR) and the decision-driven approach (DDA) have good performance, additional aids must be used. In addition, processing times from signal analysis (fast Fourier transform, FFT) present a problem for real-time applications. Significant improvements in both can be achieved by implementing signal analysis and filtering approaches with human hearing models and run time reduction. 13/31 Austrian Patent Office AT 509 570 B1 2011-12-15

V. INTRODUCTION

The main part of this description is devoted to the preparation and the analysis of the audio signal using efficient algorithms with short delay times. Our system combines an auditory gammaton filter bank (RF Lyon, "The All-Pole Gammatone Filters and Auditory Models", Proc. Forum Acusticum, Antwerp 1996; L. Lin, E. Ambikairajah, WH Holmes, &quot; Auditory Filterbank Design EUROSPEECH Scandinavia, 7th European Conference on Speech Communication and Technology, 2001; L. Lin, E. Ambikairajah, WH Holmes, &quot; Perceptual Domain Based Speech and Audio Coder &quot;, Proc. Of the Third International Symposium DSPCS 2002, Sydney, Jan. 28-31, 2002) using the Ephraim-Malah noise suppression rule (Y. Ephraim and D. Malah, &quot; Speech Enhancement Using a Minimum Mean-Square Error Short-Time Spectral Amplitude Estimator &quot;, IEEE Transactions on Acoustics, Speech, and Signal Processing, No. 6, vol., ASSP-32, pp. 1109-1121, Dec. 1984; Ephraim and D. Malah, "Speech Enhancement Using a Minimum Mean-Square Error Log-Spectral Amplitude Estimator" ;, IEEE Transactions on Acoustic s, Speech, and Signal Processing, no.2, vol. ASSP-33, pp. 443-445, Apr. 1985; P.J. Wolfe and S.J. Godsill, &quot; Simple Alternatives to the Ephraim and Malah Suppression Rule for Speech Enhancement &quot;, Proc. 11th IEEE Signal Processing Workshop, pp. 496-499, 6-8. Aug 2001). This combination was recently presented by the authors using the combination of an auditory gammatone filter bank with a Wiener noise canceler of (L. Lin, E. Ambikairajah, &quot; Speech Denoising Based on an Auditory Filter Bank &quot;, 6th ICSP, International Conference on Signal Processing, (552-555), 26-30 Aug. 2002) and a frequency domain solution of WO 00/30264 (International Applicants No. PCT / SG99 / 00119). Furthermore, the integration of an outer and middle ear filter in the time domain and the integration of a nonlinear temporary post-masking filter (G. Stall, JG Beerends, R. Bitto, K. Brandenburg, C. Colomes, B. Feiten, M. Keyhl, C Schmidmer, T. Sporer, T. Thiede, WC Treurniet, &quot; PEAQ - the new ITU standard for the objective measurement of perceived audio quality &quot;, RTM - Rundfunktechnische Mitteilungen, the trade journal for radio and television technology, Volume 43, ISSN 0035- 9890 (81-120), Mensing GmbH + Co. KG, Publishing Division, Sept. 1999; L. Lin, E. Ambikairajah, WH Holmes, &quot; Perceptual Domain Based Speech & Audio Coder &quot;, Proc. Of the Third International Symposium DSPCS 2002, Sydney, Jan. 28-31, 2002) into a noise cancellation system. In addition, a short-latency narrow-band level detector utilizing the phase of a first-order simple filter is first introduced. Finally, we present a simple scheme for signal reconstruction (recovery) while avoiding band edge signal cancellations.

The combination of an auditory gammaton filter bank and an EMSR noise canceler in a time domain approach. Integration of an outer and middle ear filter into the suppression system in one

Time domain approach · integration of a post-masking auditory filter

Short-latency narrow-band level detector Wolfe and Godsill signal recovery at low cost Short-latency upsampling Short-latency recovery despite hindering destructive interference [0099] [00101] The references "Speech denoising based on auditory filterbank" ; Lin et al., 2002, "Nonlinear Adaptive Speech Enhancement Inspired by Early Auditory Processing". by Hussain et al., 2005 and WO 0205262 A2 each describe a method for noise suppression of audio signals in the time domain. In this case, a splitting of the audio input signal into a plurality of frequency subbands, in each of which a noise reduction is performed, then the filtered frequency subbands are reassembled into an output signal. Such methods for reconstructing the output signal from frequency subbands allow without the use of a synthesis AT 509 570 B1 2011-12-15

Filterbank no undersampling of the subband signals, which leads to a comparatively high computing power requirements. Thus, these methods have a major drawback, especially in modern telecommunications and mobile systems.

VI. SYSTEM OVERVIEW

The overall system is shown as a block diagram in Fig. 7 and may be implemented as an analog or digital effects processor or as part of a software algorithm. Within the overall system are several subsystems (Figure 8): an outer and middle ear filter (Home), [00105] a Gammaton Filter Bank Analysis (GFB) section, [00106] the low latency level detector (LD · Auditory post-masking filter (PM), · a recursive noise spectrum estimator (NE), [00109] · spectral subtraction weight (EMSR), [00110] · short latency upsampling (L t ), The vocoder state and [00112] the inverse outer and middle ear filters (Η, ομε) -

VII. OUTER AND MIDDLE EAR FILTER

An outer and middle ear filter comprises three second order (SOS) parts representing the physiological part of the human ear (E. Zwicker, H. Fastl, &quot; Psychoa-coustics, facts and models &quot;, Springer, Berlin Heidelberg E. Terhardt, &quot; Acoustic Communications &quot;, Springer, Berlin Heidelberg, 1998): 1) The high-pass attenuation curve below 1KHz models the 100-phonon curve which includes the outer ear acoustic impedance and mechanical impedance Auditory ossicles in the middle ear represent [00115] 2) the resonance of the ear canal and [00116] 3) the low-pass attenuation curve above 1 kHz models the threshold of hearing.

The last two filters are optional, with the high-pass component being mandatory and reducing the influence of low-frequency noise on the noise canceler.

A filter structure with an adequate size transfer function might eventually look like FIG. 9. All three filter sections must have second order sections to ensure proper edges. The outer filter edges can be modeled as second order low and high pass cow tail filters, where the resonances can be modeled as a parametric bell filter (P. Dutilleux, U. Zölzer, "DAFX", Wiley & Sons, 2002).

The filter inversion is straightforward. If zeros at e.g. z = 1 in the z-range, the inverse filter can not do this. Perhaps z = 0.99 is a suitable choice for a starting value for inversion a z = 1 zero.

VIII. FREQUENCY GROUPS / BAND WIDTHS

Frequency grouping is an important effect in human perception of volume. The perceived volume includes special volumes for different frequency ranges. An audible frequency scale can be used to model the frequency group effects, the units of which can be considered the frequency resolution of human volume perception (E. Zwicker, H. Fastl, "Psychoacoustics, facts and models", Springer, Berlin Herdeiberg, 1999). We denote any audible frequency transformation with ffi {·} and the corresponding inverse frequency transformation with A reasonable frequency scale uses a small number of frequency groups according to the formula of Traunmüller (e. Terhardt, "Acoustic Communication", Springer, Berlin Heidelberg, 1998) * 7 [Bark] = »{// [Hz]} = 11 ^ -0.53. (19)

Accordingly, the inverse transformation ffi'1 {} // [Hz] =! 8 W [Bark]} = 1960 &lt; 20 &gt; The center frequencies fk of the auditory filter bank can be calculated using the inverse transformation fk = 0, 'r - \\ pk) on an equidistant scale vk (with distances dv, eg dv = 1 [Bark]) in the Bark space Similarly, the bandwidths Bk of Bk = R'1 {vk + dv / 2} - ffi-1 {vk-dv / 2} can be calculated. Other Bark scales (e.g., E. Zwicker, H. Fastl, &quot; Psycho-Acoustics, Facts and Models &quot;, Springer, Berlin Heidelberg, 1999) use smaller bandwidths and result in higher group delay auditory filters; therefore, the above distance is preferred.

In order to avoid confusion with the variable z of the z-range, v is used instead of z for the bark frequencies.

IX. GEHOR Gammatone FILTERS

[00123] Auditory gamma-tone filters (R.F. Lyon, "The All-Pole Gammatone Filter and Auditory Models", Proc. Forum Acusticum, Antwerp 1996) can be efficiently implemented in the time domain and allow the separation of a wideband audio signal in auditory canal signals. The response size of the gammaton filter corresponds to the immediate fording properties of the human ear. The size of this filter plotted over the audible frequency scale remains the same regardless of the center frequency the filter has been designed for. The arbitrary shape represents a family of gamma-tone filters of order m and is further illustrated, where k is the filter bank channel index. A corresponding z-transformation, where * GF denotes any Gammaton filter (eg GF, APGF, OZGF, TZGF): H * GF, k (^) 9 * GF '-iÄium.ki'Z) Π _1_ 1 - 2 · Rk cos (0fc) · z · x + rk · z ~ 2 (21) Digital center frequencies 0k and pole radii rk are determined by the continuous time quantities center frequency fk, bandwidth Bk, band edge suppression CdB (eg CdB - 5) [dB]) and the sampling rate fs:

9k = 2π * JS rk = 1 - 2ττ * 7 * Js (22) An auditory gain-tomon filter bank represents a group of overlapping gam-matone filters which subdivides the audible frequency scale into equidistant frequency bands. The order m = 4 is often used in the literature, where the order m = 3 has been proposed to minimize the computational power. The term g * GF should be adjustable so that the unity gain at the center frequency fk is achieved. For a particular form of gamma-tone filter, the system Hnum, k (z) must be suitably adapted as shown in the following subsections. 16/31

Austrian Patent Office AT 509 570 B1 2011-12-15

A. SIMPLE GAMMATON FILTER

The simple gammaton filter (GF; RF Lyon, &quot; The All-Pole Gammatone Filter and Auditory Models &quot;, Proc. Forum Acusticum, Antwerp 1996) must be evaluated from the time-continuous impulse response using Laplace and impulse variance transformation (AV Oppenheim, RW Schäfer, JR Buck, &quot; Discrete-Time Signal Processing &quot;, Prentice Hall, 1999): (23) h (t) = tm 1e Bfc t cos (27r / fct), [00127] which is the unknown Polynominal Hnum, k (z) determined in (21). Because of its shape and computational effort, its use is not recommended.

B. ALL-POL GAMMATON FILTER

An all-pole gamma-tone filter (APGF) is obtained when the polynomial in (21) disappears Hnum, k (z) = 1. It is the most efficient gammaton filter (RF Lyon, &quot; The All-Pole Gammatone Filter and Auditory Models ", Proc. Forum Acusticum, Antwerp 1996).

C. ONE-ZERO GAMMATONE FILTER

Placing Hnum, kZ) = (1-z'1) in (21) results in a so-called One-Zero Gam-maton filter (RF Lyon, &quot; The All-Pole Gammatone Filter and Auditory Models &quot; Proc. Forum Acusticum, Antwerp 1996). The One-Zero Gammaton Filter (OZGF) can be efficiently derived from a &quot; One-Zero &quot; for all channels k before disassembling in k all-pole gammaton filters are assembled.

D. THREE-ZERO GAMMATON FILTERS _ | _Λ [00130] If a pair of complex-conjugate zeros z = rz-e ~ z * with the digital frequency 0z, k at 1 Bark above the center frequency 0k with a radius rz «0.98 and an additional zero at z = 1 are added, one obtains Hnwnk {z) = (\ - 2rzcos {ezk) zi + r2z ~ 2) - (lz ~ 1) for the Three Zero Gammaton Filter (TZGF) an improved form (L. Lin, E. Ambikairajah, WH Holmes, &quot; Auditory Filterbank Design Using Masking Curves &quot;, Proc. EUROSPEECH Scandinavia, 7th European Conference on Speech Communication and Technology, 2001). The arithmetic effort of the one-zero gammaton filter of order m + 1 is equal to the effort of the three-zero gammaton filter of order m, in case a single &quot; one-zero &quot; is used for all channels k. Suitable transforms and digital frequency calculations 0z, k follow from (19), (20) and (22).

X. REPRODUCTION

The re-composition of a wideband signal from the audible band signals may be implemented as addition of all signal bands. Unfortunately, destructive signal cancellation can occur in the overlapping regions of adjacent signal channels. Therefore, we derive a simple criterion that shows the need for a sign change for every other channel before the summation:

(24) Using this formula, the frequency response of the superposition of all signals is in the range of CdB + 3 [dB] and 0 [dB]. The omission of a necessary sign may result in destructive signal cancellation at the band edges of adjacent filters.

XI. (RATE-REDUCED) LEVEL DETECTION

Blanking effects modeled by the auditory filter bank can not be exploited as long as the amplitude of the filter bank channel is not determined. Suitable ways of level detection are proposed in the following subsections.

We propose the first simple approach for high frequency channels and the runtime reduced approach for the low frequency bands.

A. EASY LEVEL DETECTION WITH PRE-MASKING

Normally, nonlinearities such as e.g. Absolute magnitude, square, half-wave rectification, used to transform the signal amplitude into baseband at about 0 Hz. Furthermore, a smoothing filter removes higher frequency components, and ultimately the desired amplitude signal is found. Fig. 11 shows an example which also takes into account the form factor F.

Usally used approaches of amplitude detection are computationally efficient, smoothing filters include group delay times in the signal path, which are to be compensated. We recommend to describe the recursive smoothing parameter α by a time constant Tavg in [s] a = e_7T7 *. (25) Suitable time constants are consistent with the pre-holographic time constant, and is approximately Tavg ~ 2 [ms] (G. Stoll, JG Beerends, R. Bitto, K. Brandenburg, C. Colomes, B. Feiten, M Keyhl, C. Schmidmer, T. Sporer, T. Thiede, WC Treurniet, &quot; PEAQ - The New ITU Standard for the Objective Measurement of Perceived Audio Quality &quot;, RTM - Rundfunktechnische Mitteilungen, the trade journal for radio and television technology, 43rd grade , ISSN 0035-9890 (81-120), Mensing GmbH + Co. KG, Publishing Department, Sept 1999).

B. LENGTH-REDUCED LEVEL DETECTION

Our new method exploits the phase of a simple filter section. This method of level detection can also find application in other technical fields and is not limited to noise suppression alone.

With the Hilbert transform, the broadband signal can be consistently phase-shifted by 90 °. By summing the squares of the original and the shifted signals, the squares of the amplitudes (e.g., signal power) remain, and the sinusoidal components cancel each other out. But a causal implementation of the Hilbert transform does not exist.

In contrast to the ideal Hilbert transformer, we need the 90 ° phase shift only in the considered frequency interval, e.g. in the corresponding audible frequency group.

We propose to use the following filter types for a 90 ° phase shift at a frequency 0k: a simple first-order FIR portion, a simple first-order 11R all-pass (AP), and [00144] · a simple delay line with a Kl4 delay at 0k.

Each of the above methods provides 90 ° phase shift at a virtual arbitrary frequency 0k and is therefore suitable.

One can choose between the following properties: FIR: numerically unstable at 0k = [0, π / 2, π], offers the widest band with 90 ° phase shift.

AP: numerically not stable at 0k = [0, π / 2, π], the 90 ° phase frequency band is narrower and the computational effort is greater. 18/31 Austrian Patent Office AT 509 570 B1 2011-12-15 · λ / 4-delay: numerically stable, the narrowest frequency band with 90 ° phase shift, low computation, much memory required.

[00150] Fig. 12 shows an example of the FIR level detection method. A suitable parameter can be found via the phase equation for the corresponding system, e.g. A. V Oppenheim, R.W. Schäfer, J.R. Buck, Discrete-Time Signal Processing, Prentice Hall, 1999.

XII. AUDITOR POST MASKING

The use of the non-linear post-masking filters (e.g., recursive averaging responsive to falling edges) has several advantages: The Impulsive Noise Variance is slightly overestimated (over-subtraction) because of the fade-out.

Noise suppression algorithms can not attenuate signals until the

After-eclipse time has passed.

· Aliasing effects after the downsampling or the ripple in the amplitude signal are reduced due to the smoothing effect of the fade-out.

It is smoothed and the amplitudes of the important transient signals experience no additional group delay times.

We propose a structure that works on the signal power in each channel (see Fig. 13, L. Lin, E. Ambikairajah, WH Holmes, &quot; Perceptual Domain Based Speech and Audio Coder &quot;, Proc. Of the Third International Symposium DSPCS 2002, Sydney, Jan. 28-31, 2002).

The averaging parameter ak in channel k has to correspond to the human post-audible time constant for the corresponding frequencies fk. Therefore we use the following equation to derive the mean parameter α: et * = e G rk · ^. (26) A parameter G can be used to scale the fade out time constant.

The time constant for 1 [Bark] is approximately τ ~ 40 [ms], and for 20 [Bark] approximately τ ~ 4 [ms] (G. Stoll, JG Beerends, R. Bitto, K. Brandenburg, C. Colomes, B.

Feiten, M. Keyhl, C. Schmidmer, T. Sporer, T. Thiede, WC Treurniet, &quot; PEAQ - the new ITU standard for the objective measurement of perceived audio quality &quot;, RTM - Rundfunktechnische Mitteilungen, the trade journal for radio and television technology, Volume 43, ISSN 0035-9890 (81-120), Mensing GmbH + Co. KG, Publishing Department, Sept 1999). The following equation can be used to derive τκ:

T * / [ms] = -_ j___i_ 1 - W

Alternatively, the equation may be used in the cited references, but our formula provides suitable interpolation with longer time constants.

XIII. REKURSIVE MINIMUM STATISTICS

We can use the structure in Fig. 14 to estimate the noise level in each frequency band. Similar approaches can be found in R. Martin, "Noise Power Spectral Estimation Based on Optimal Smoothing and Minimum Statistics", IEEE Transactions on Speech and Audio Processing, no. 5, vol. 9, pp. 504-512, Jul. 2001 or WO 00/30264 (International application No. PCT / SG99 / 00119). 19/31

Austrian Patent Office AT 509 570 B1 2011-12-15 This method mainly uses three time constants for averaging the signal levels. Falling edges are easily averaged, while during increasing input edges the output is kept constant during the period of Nw sampling intervals (infinite time constant). If Nw sampling intervals have elapsed, the rising edge is averaged by a third time constant. The time constants can be converted to a recursive mean parameter, similar to (25) and (26).

A suitable counter limit Nw can be calculated by means of a continuous time interval Tw (28)

Nw = round (Tw · fs).

For utterances or words of human speech, this time interval may be appropriately selected, e.g. Tw ® 1.5s. The falling edge time constant may be a scaled version of the fade out time constant, or e.g. be constant 200 [ms].

The rising edge defining time constant β may be approximately 700 [ms], which corresponds to a speed of approximately 6 [dB] / [s]. In contrast to all other time constants, this is suggested as equal to all channels k.

The saturation effect in Fig. 14 can be given as follows: f (x) = &lt; 1 if x &gt; 0, (29) v 0 otherwise. XIV. EPHRAIM-MALAH NOISE REDUCTION RULES (EMSR) [00167] With the EMSR (Y. Ephraim and D. Malah, &quot; Speech Enhancement Using a Minimum Mean-Square Error Short-Time Spectral Amplitude Estimator &quot; IEEE Transactions on Acoustics, Speech and Signal Processing, No. 6, vol., ASSP-32, pp. 1109-1121, Dec. 1984, Y. Ephraim and D. Malah, &quot; Speech Enhancement Using a Minimum Mean- Square Error Log Spectral Amplitude Estimator ", IEEE Transactions on Acoustics, Speech, and Signal Processing, no.2, vol. ASSP-33, pp. 443-445, Apr. 1985), we can obtain the clear speech amplitude from the given noisy speech amplitude and estimate the noise variance. We can e.g. the definition of Wolfe and Godsill for the spectral weights (PJ Wolfe and SJ Godsill, &quot; Simple Alternatives to the Ephraim and Malah Suppression Rule for Speech Enhancement &quot;, Proc. 11th IEEE Signal Processing Workshop, pp. 496-499, 6- 8 Aug 2001) and a Modified Decision-driven Approach (F. Zotter, M. Noisternig, R. Höldrich, "Speech Enhancement Using the Ephraim and Malah Suppression Rule and Decision Directed Approach: A Hysteretic Process", to appear in IEEE Signal Processing Letters, 2005. First manuscript sub-mitted Jan 24, 2005)

(30) The following relationships are involved in the above equation:

£ fc [m] = α · min (7fc [ra] -1,0) + (31) 20/31 (32) Austrian Patent Office AT 509 570 B1 2011-12-15 P (1 - a) · 7k [m The noise variance a2dk [m] is given by the noise estimation algorithm, m and n are time indices, fs is the system sampling rate, and L is .sigma..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times..times is a downsampling factor.

According to Y. Ephraim and D. Malah, "Speech Enhancement Using a Minimum Mean-Square Error Short-Time Spectral Amplitude Estimator", IEEE Transactions on Acoustics, Speech, and Signal Processing, no. 6, vol. ASSP-32, pp. 1109-1121, Dec. 1984, istyk [m] is the a posteriori SNR and ξκ [ηι] is the a priori SNR. Gw, k [m] is the spectral weight of the Wiener filter, α the mean parameter defined by a mean time constant Tsnr, k, which is either approximately 2 [ms] (F. Zotter, M. Noisternig, R. Hoeldrich, &quot; Speech Enhancement Using the Ephraim and Malah Suppression Rule and Decision Directed Approach: A Hysteretic Process, to appear in the IEEE Signal Processing Ladder, 2005. First manuscript submitted Jan 24, 2005) or derived from the hearing-time constants.

The &quot; over subtraction factor &quot; p (see Zotter et at) can be chosen as p = 10'15 / 1 ° and the lower noise parameter ζ as ζ = 10'40 / 10.

XV. RUNTIME-REDUCED UPSAMPLING

Normal upsampling requires either a processing delay or a group delay because of the interpolation operation involved. When using the upsampling factor L, such delay times are approximately L sampling steps long.

We propose to use a special method for upsampling that does not add extra delay times. This can be accomplished by splitting the signal into buffers (preferably with a buffer size of the ADC and DACs).

If the last sample of the previous block is present in each signal block, it is possible to linearly interpolate the following samples. Therefore, the last sample in each block must match the sample time of the lower sample rate.

XVI. CONCLUSIONS

Frequency domain solutions using equivalent hearing models require delay times in the order of 10 milliseconds. The implementation of our system with 20 frequency bands and a third-order TZGF has a mean latency of 3.5 to 4 milliseconds. The required computational effort is about 8.9 MIPs at fs = 16 [kHz], which is a little more than what is needed for DFT solutions (7 MIPs). We also used a slightly modified Ephraim Malah suppression rule (EMSR) with the simplified Wolfe-Godsill formula and the modified decision-driven approach.

The disclosure of all cited publications is incorporated in its entirety in this specification. 21/31

Claims (15)

Austrian Patent Office AT 509 570 B1 2011-12-15 Claims 1. An interference suppression method for an input audio signal (y [n]) having a desired signal (x [n]) and a noise signal component, the method comprising the steps of: - Splitting the input audio signal (y [n]) into a plurality of subbands (yk [n]) by a band splitting analysis based on a gammaton filter bank (GFB), preferably a nonuniform gammaton filter bank, noise suppression in each subband (yk [nj ) by a plurality of noise reduction processors, - composition of the plurality of subbands (yk [n]) into an output signal (x [nj) through a synthesis filter, wherein all steps are performed in the time domain, characterized in that the gammaton filter bank (GFB) performs a phase shift on the subbands. 2. The method according to claim 1, characterized in that a preprocessor (Home) and a post-processor (Hiome) perform a non-linear filtering of the input audio signal (y [nj], thereby: a. a preprocessing filter which emulates the transfer behavior of the human outer and middle ears and is applied to the discrete-time noisy input audio signal (y [nj]), and b. a post-processing filter applied to the canceled / improved full-band signal to compensate for the effect of the pre-processed filter. 3. The method according to any one of claims 1 or 2, characterized in that the noise reduction processors each comprise a signal level detection (LD), a noise estimator (NE), an auditory blanking filter (PM) and a subtraction processor. Method according to claim 3, characterized in that the signal level detection (LD) uses that phase of the low-order subband to generate a quadrature signal and to evaluate an in-phase signal from the subband (yk [nj) and the squared amplitudes of these signals summed to the squared amplitude envelope. A method according to claim 3, characterized in that the noise estimator (NE) generates a sub-band noise sound value by smoothing based on the minimum statistics, in particular using a weighted averaging of the previous noise value and the current input value with three different time constants. Method according to claim 3, characterized in that the auditory blanking filter (PM) uses the detected signal power in each subband to generate a temporary blanking behavior based on human auditory sense, in particular a non-linear, weighted average of the previous subband input value and the current input value only is applied to falling edges as a function of the detected level in each subband. A method according to any one of claims 1 to 6, characterized in that the noise estimator (NE) depends on the current input value compared to time-dependent, level-dependent thresholds. 22/31 Austrian Patent Office AT 509 570 B1 2011-12-15 A method according to any one of claims 1 to 7, characterized in that the noise cancellation in each subband is performed by the Ephraim Malah Noise Suppression Rule (EMSR). 9. Method according to one of claims 1 to 7, characterized in that the noise suppression in each subband is performed by a decision-driven approach (DDA). An apparatus for noise suppression for an input audio signal (y [nj) comprising a desired signal (x [nj] and a noise signal component, the apparatus comprising: a band split analyzer for splitting the input audio signal (y [nj] into a plurality subbands (yk [nj], based on a gammaton filter bank (GFB), preferably a nonuniform gammaton filter bank, - a plurality of noise canceling processors for noise canceling in each subband (yk [nj), - a synthesis filter for composing the plurality of subbands (yk [nj] to an output signal (x [nj), wherein all components work in the time domain, characterized in that a preprocessor (Home) and a post-processor (Η, ομε) perform a non-linear filtering of the input audio signal, thereby: a. a preprocessing filter which emulates the transfer behavior of the human outer and middle ear and is applied to the discrete-time noisy input audio signal; and b. a post-processing filter applied to the improved full-band signal to compensate for the effect of the preprocessed filter. An apparatus according to claim 10, characterized in that the noise reduction processors each comprise a signal level detector (LD), a noise estimator (NE), an auditory blanking filter (PM) and a subtraction processor. 12. Apparatus according to claim 11, characterized in that the signal level detector (LD) uses the phase of the low-order filter section to generate a quadrature signal, evaluates an in-phase signal from the subband (yk [nj) and the sums up the quadratic amplitudes of these signals. 13. Apparatus according to claim 12, characterized in that the quadrature signal is generated by a in the signal level detector (LD) provided FIR portion of the first order. 14. Apparatus according to claim 12, characterized in that the quadrature signal is generated by a in the signal level detector (LD) provided FIR-AII pass (AP) first order. 15. Apparatus according to claim 12, characterized in that the quadrature signal from a delay line to create a λ / 4 delay at a digital frequency (0k) is generated. For this 8 sheets drawings 23/31