Systematic Integration of Acoustic Echo Canceller and Noise Reduction Modules for Voice Communication Systems

Transcription

1 INTERSPEECH 2015 Systematic Integration of Acoustic Echo Canceller and Noise Reduction Modules for Voice Communication Systems Hyeonjoo Kang 1, JeeSo Lee 1, Soonho Bae 2, and Hong-Goo Kang 1 1 Dept. of Electrical and Electronic Engineering, Yonsei University, Seoul, Korea 2 Samsung Electronics, Suwon, Korea volleruhe@dsp.yonsei.ac.r Abstract This paper proposes an efficient way of integrating acoustic echo canceller (AEC) and bacground noise reduction (NR) modules for voice communication systems. The main strategy for designing a standalone AEC or NR module is straightforward, but it is not easy to integrate two modules in a single system because of the undesired effect caused by the nonlinear nature of each module s output. The proposed algorithm directly estimates noise and echo components from the observed signal, then they are utilized in the process of updating AEC module. Since the estimation step is independent of the actual processing of NR module, the nonlinear effect caused by coupling the NR module with the AEC module can be minimized. Experimental results show that the proposed algorithm achieves the performance of standalone AEC in terms of echoreturn-loss-enhancement (ERLE) metric while maintaining that of standalone NR module in the spectral distortion aspect. Index Terms: Acoustic echo canceller (AEC), Speech enhancement, Integrated system, Noise reduction (NR) 1. Introduction In a hands-free voice communication scenario such as conference calls or in the car driving situations, it is very important to enhance intelligibility by removing acoustic echo and bacground noise. The acoustic echo canceller (AEC) typically utilizes an adaptive filter to estimate the room impulse response (RIR) of user s communication environment, then artificially generates the echo signal (i.e. signal from the far-end speaer to be removed). Several methods such as least mean square (LMS), affine projection (AP), and recursive least square (RLS) are well-nown adaptive filter algorithms [1]. The performance of AEC algorithms drops significantly in practical situations because most adaptive filters are sensitive to environmental changes such as variations of echo path, bacground noise, and nonlinear effects due to the characteristics of speaers or microphones [2]. Note that some of the problems can be relieved by efficiently updating the coefficients of adpative filters (ADF) [3, 4, 5]. However, it is virtually impossible to appropriately adjust filter coefficients in bacground noise environments, thus the AEC module also requires an additional noise reduction module. Several systems that include both AEC and NR modules have been proposed to address both problems in a unified structure [6, 7]. However, most of the previous literatures simply 2 This wor was done when Soonho Bae was in Yonsei University. concatenate two modules, where the standalone AEC module follows the NR module or vice versa. In the former case, the temporal variation of the noise suppression filter in the NR module distorts the characteristic of the echo signal, and the subsequent ADF does not converge correctly [8]. In the latter case, it is not easy to update the coefficients of the ADF because of the environmental noise of the near-end speaer, which results in residual echo or undesired alteration of noise power spectral density (PSD) [9]. Both problems mentioned above mainly come from the fact that each module operates independently, regardless of the processing order. Although the problems can be somewhat relieved by adjusting the parameters of the AEC and NR modules using ad-hoc rules [7, 8], they are not efficient if the characteristics of environment change rapidly, because the step size of ADF is set to be a small value. Hence, a more systematic integration approach where the two modules closely interact with each other is required. This paper proposes an integration of AEC and NR modules within a single processing framewor. In other words, parameters needed for both AEC and NR modules are collaborately estimated not to provide any severe impact to each other, which enhances the quality of the echo and noise-removed speech signal. We demonstrate, in simple concatenated structures, how the anterior module affects the posterior module and degrades overall performance of the integrated system. Then we show how such coupling effect can be minimized in the proposed algorithm by employing the noise PSD estimated directly from the microphone input signal. By employing both raw echo and estimated noise PSD information simultaneously, the echo signal can be successfully removed without being affected by the bacground noise component. This paper is organized as follows. Section 2 describes the problem statement of this paper and explains the fundamental concepts of AEC and NR algorithms. Section 3 presents the detailed processing step of the proposed method. Experimental results and conclusions are followed in section 4 and 5, respectively. 2. Bacground Theory This section describes a basic building structure of a typical voice communication scenario that requires echo removal and noise reduction modules. Based on the functional description of conventional AEC modules and NR module, several drawbacs of combining two modules are explained. Copyright 2015 ISCA 1770 September 6-10, 2015, Dresden, Germany

2 (a) Figure 1: A simplified voice communication scenario 2.1. Problem definition Figure 1 depicts a simple voice communication scenario that we consider in this paper. The signal acquired at the near-end microphone z(t) is modeled by a summation of acoustic echo y(t), near-end speech s(t), and bacground noise n(t), z(t) = y(t) + s(t) + n(t), (1) where the acoustic echo is assumed to be generated by a convolution between the far-end signal x(t) and the RIR w(t), y(t) = x(t) w(t). (2) The frequency domain representation of microphone input becomes Z(l, ) = Y (l, ) + S(l, ) + N(l, ), (3) where l and denote frame and frequency bin indices, respectively. The ADF estimates W (l, ) to generate a synthesized echo Ŷ (l, ) to be subtracted from the input signal. Note that the ultimate goal of the proposed system is to remove acoustic echo and bacground noise while preserving the near-end speech as natural as possible. In a simple form, it can be achieved by multiplying a gain function to input signal, Ŝ(l, ) = G AECNR(l, )Z(l, ) = f(g AEC (l, ), G NR (l, ))Z(l, ). The gain term, G AECNR (l, ) includes two sub-gain terms, G AEC(l, ) and G NR(l, ) that need to be estimated in both AEC and NR modules Conventional integration method Bloc diagrams in Figure 2 depict two simple concatenative structures to remove both acoustic echo and bacground noise. The configuration in Figure 2-(a) represents an NR module followed by an AEC module and Figure 2-(b) represents a configuration in reverse order, where each system will be referred to as system 1 and 2, repectively. Both systems 1 and 2 share the same NR and AEC modules. Since it is well nown that speech quality is less affected by phase components, both AEC and NR modules employ an amplitude based enhancement technique [10]. A noise reduction module typically taes a spectral amplitude estimator based algorithm such as: (4) Z NR (l, ) = G NR (l, )Z(l, ), (5) where G NR (l, ) is a frequency dependent gain function that varies depending on the signal-to-noise ratio (SNR) of each frequency bin. A minimum mean-square-error (MMSE) criterion, (b) Figure 2: Bloc diagrams for (a) conventional system 1 and (b) conventional system 2 either in linear or log domain, is used as a cost function to estimate G NR(l, ), where it is determined as a function of a priori and a posteriori SNRs [11, 12]. To accurately estimate the SNR values, a noise PSD, σ 2 N (l, ) = E[ N(l, ) 2 ] should be estimated correctly. Minimum statistics noise estimation (MSNE) based algorithms are employed to estimate ˆσ 2 N [9, 13]. A spectral subtraction based AEC algorithm is nown for many advantages over conventional ones because of its robustness to minor echo path changes, fewer control parameters, and lower computational complexity [14]. The echo removed signal is obtained as Z AEC(l, ) = Ŷ (l, ) = G AEC(l, ), where G AEC (l, ) denotes a frequency dependent gain function to reduce the echo term defined as: { [ Ŷ (l, ) ] } G AEC (l, ) = max 1, δ, (7) where δ is a flooring constant. Normalized absolute LMS (NABSLMS) adaptation is employed to update the RIR, Ŵ (l + 1, ) = Ŵ (l, ) + µ E(l, ) X(l, ), (8) where is the PSD of the far-end signal computed by a first-order recursive averaging method, and E(l, ) = Ŷ (l, ). A double-tal detector (DTD) that uses a coherence function is employed to determine whether the filter coefficients of ADF are updated or not in the current processing frame. Since the coherence function represents the similarity between far-end and near-end input, it is possible to decide whether single speaer is present (single-tal) or two speaers are present (double-tal) [15]. 3. Proposed System In the previous section we briefly reviewed the conventional systems. In this section, we first analyze the limitations of conventional systems. Then we propose a systematic integration of AEC and NR modules by exploiting such characteristics Analysis of conventional systems Due to the frame-wise update process of the ADF, the system 1 underestimates RIR whereas the system 2 overestimates. Since (6) 1771

3 the update rate in Eq. (8) is same across all frequency bins, and the update process is performed to the frames where echo signal is grater than noise signal on average. However, because of the sparseness characteristics of speech signal, we may not assure that Y (l, ) N(l, ) over all frequency bins. Hence, when the echo-to-noise ratio is high, i.e. Y (l, ) N(l, ), it is safe to assume that Ŵ (l, ) W (l, ) in the conventional systems given in Figure 2. As the SNR decreases, the ADF coefficients in the systems show opposite behavior. The update of ADF in AEC is given as: Ŵ (l, ) = Ŵ (l + 1, ) Ŵ (l, ). (9) From Eq. (8) and Eq. (9), Ŵ (l, ) in the system 1 is Ŵsys1(l, ) = µ ENR(l, ) X(l, ), (10) S XX(l, ) where E NR (l, ) = Z NR (l, ) Ŵ (l, ) X(l, ). Note that the system 1, the noise reduction gain G NR(l, ) is close to 0 in low SNR regions, consequently Z NR(l, ) Y (l, ). Then E NR(l, ) Ŵ (l, ) X(l, ) and Eq. (10) can be simplified as Ŵ (l, ) X(l, ) 2 Ŵsys1(l, ) µ (11) µŵ (l, ), if it is assumed that X(l, ) 2 S XX(l, ). This suggests that the estimator, i.e. Ŵ (l, ) W (l, ) because the update process consequantly reduces the filter coefficients in low SNR regions. In turn, this leads to residual echo in high SNR regions, since Ŷ (l, ) Y (l, ) Z NR(l, ). On the other hand, the update of ADF in the system 2 becomes E(l, ) X(l, ) Ŵsys2(l, ) = µ. (12) In low SNR conditions, N(l, ) and Ŷ (l, ) Y (l, ) N(l, ) are satisfied. Hence, Eq. (12) becomes { N(l, ) Ŷ (l, ) } X(l, ) Ŵsys2(l, ) µ 0, (13) which results in Ŵ (l, ) W (l, ). The overestimation of RIR does not affect to the echo cancellation performance directly, but brings problems in the NR module. The overestimated RIR causes the overestimation of the echo signal in high SNR regions and brings undesired subtraction of the bacground noise. The reduced bacground noise level also brings the underestimation of noise PSD to the NR module due to the sensitiveness of a minimum statistics based NR module. This underestimation process propagates in subsequent frames, which also drops the amount of estimated noise PSD continuously, thus it brings residual noise to output signal. Hence, in both conventional systems, the anterior module distorts the characteristic of posterior module, thus the output quality degrades Systematic integration As mentioned in the previous subsection, it is inevitable to overestimate the RIR because it is difficult to perfectly cancel out the echo signal. This phenomenon results in the underestimation of Figure 3: Bloc diagram of the proposed system noise PSD. However, underestimation of noise PSD leads to a malfunction of spectral gain estimation function which causes perceptually annoying residual signal [9]. Therefore, a systematic integration approach is required in order to overcome the problem. This paper proposes a structure that alleviates the problem by collaborately manipulating the information from each module. As depicted in Figure 3, the proposed system estimates noise PSD directly from the near-end input. Thus its estimation accuracy is very high. The estimated noise PSD ˆσ N 2 (l, ) is utilized to modify the AEC gain function in eq.(7) with an adaptive flooring gain { [ Ŷ (l, ) ] } G AEC AF (l, ) = max 1, G min (l, ), (14) where the adaptive flooring gain G min(l, ) is represented as 2 ˆσ G min (l, ) = N (l, ). (15) While the fixed gain δ in eq. (7) brings spectral distortion to Z AEC (l, ) that interrupts precise nosie PSD estimation, the adaptive gain G min (l, ) in eq. (15) compensates the oversubtracted spectral components. Indeed, the system preserves the bacground noise component even after the echo is cancelled, thus the performance of echo cancellation can be enhanced. In addition, since the NR module operates independently, it is not affected by the AEC module. In other words, the performance of the noise PSD estimation module is ept to be stable due to the usage of original microphone signal. In summary, the NR gain estimation module plays a role in both bacground noise removal and residual echo suppression steps. The total gain function of the proposed system is obtained by multiplying an adaptive flooring function G AECAF (l, ) with the NR gain function based AEC gain, i.e. G NR(l, ), G AECNR(l, ) = G NR(l, )G AECAF (l, ). (16) Note that the near-end input signal Z(l, ) passes the AEC module at first, and the echo-removed signal Z AEC(l, ) passes through the NR gain function. 4. Experiment In order to evaluate the performance of proposed system, experiments with objective measures under various types of noises and SNRs are conducted. The computational complexity is compared with other ADF algorithms. The input speech signal which used for the experiment starts with far-end speech which is followed by near-end speech, and ends with double-tal situation. The sampling frequency of input data is set to 16 Hz, the frame length is set to 32 ms, and 1772

4 Table 1 Mean ERLE under various SNR and noise conditions white, volvo, babble noise with 0, 10, 20 db SNR SNR System 1 System 2 System 3 White Volvo Babble 0dB dB dB dB dB dB dB dB dB Table 2 Mean LSD under various SNR and noise conditions white, volvo, babble noise with 0, 10, 20 db SNR SNR System 1 System 2 System 3 White Volvo Babble 0dB dB dB dB dB dB dB dB dB Table 3 Computational complexity Frequency bloc NLMS Unconstrained FBNLMS Proposed approach Number of FFT Total multiplication 10Mlog 2 (2M) + 16M 6Mlog 2 (2M) + 16M 2Mlog 2 (M) + 5M each frame is shifted with 16 ms. The RIR is generated using image-source method (ISM) [16], where the size of the room is set to 4x4x3 (m 3 ), with 0.45 second of reverberation time (RT) 60. Three types of noise (i.e. white, volvo, and babble) from NoiseX-92 database [17] are mixed with input speech using an additive noise assumption with three SNR (i.e. 0, 10, and 20 db) conditions. This paper uses the improved minima control recursive averaging (IMCRA) method for the noise PSD estimation submodule [13] and the optimally modified log-spectral amplitude estimator (OM-LSA) for spectral gain estimation submodule [12]. Two objective measurements are performed: mean echo return loss enhancement (ERLE) for measuring echo attenuation; log-spectral distance (LSD) for measuring the distortion of the estimated near-end speech, which are defined as: ERLE(dB) = 1 10log 10 L E l l Y (l, ) 2, Y (l, ) Ŷ (l, ) 2 (17) LSD(dB) = 1 ( ) 2 S(l, ) 20log 10, (18) L Ŝ(l, ) where L is the total number of frames. For the comparison with the proposed system, this paper refers to integration methods described in [7] for conventional systems. Systems 1, 2, and 3 (proposed one) are depicted in Figure 2-(a), Figure 2-(b), and Figure 3, respectively. Table 1 displays the mean ERLE result of each system under various types of noise and SNR conditions. In table 1, ERLE increases as input SNR increases for all types of noises and systems. The system 2 and the proposed system show that the echo is effectively removed in this system, but the system 1 shows the worst performance to all types of noise and input SNR conditions. The reason can be found from the underestimation of RIR, W (l, ) caused by the NR gain G NR (l, ). Table 2 shows the mean LSD results. The system 1 and the proposed system show similar performance, while the system 2 has about 40% higher distortion which results from wrong estimation of the noise PSD due to the oversubtraction of echo. From Table 1 and 2, it is concluded that the proposed system shows balanced results in terms of both ERLE and speech distortion by resolving the problems caused by simply concatenating the AEC module and the NR module. The computational complexity of the three systems is described in Table 3: the frequency bloc NLMS (FBNLMS) filter with overlap-and-save (OLS) approach, the unconstrained FBNLMS approach as a simplified one, and the proposed approach [18]. In Table 3, M refers the length of the adaptive filter. Unlie to the high computational complexity of FBNLMS due to the OLS and linear convolution, the proposed algorithm shows about 5 times lower computational complexity than conventional algorithm. 5. Conclusions This paper proposed a systematic integration method for an acoustic echo canceller and a noise reduction module. The proposed system utilized information from a noise PSD estimation module to an acoustic echo cancellation module in order to achieve stable operation of spectral gain estimation module. From the objective measurements, the proposed system showed similar LSD results to system 1 and ERLE results to system 2, while conventional systems had problems in one of two metrics. 6. Acnowledgement The authors would lie to than Naver Corporation for funding this research through integrated acoustic echo canceller and noise reduction system. We than Dr. Min-Ki Lee in Naver Corporation and Dr. Dongwon Lee in Line Plus Corporation providing speech signals. We would also lie to than our colleague, Ho-Seon Shin for her valuable advices related to analyzing the conventional systems. 1773

5 7. References [1] S. Hayin, Adaptive Filter Theory, Prentice Hall, [2] H. Huang, J. Benesty, J. Chen, K. Helwani, and H. Buchner, A study of the mvdr filter for acoustic echo suppression, in Proc. International conference of acoustics, speech, signal processing (ICASSP), pp , May [3] Y. Zhou and X. Li, A variable step-size for frequencydomian acoustic echo cancellation, IEEE Worshop on Applications of Signal Processing to Audio and Acoustics, pp , October [4] P. Heitaemper, An adaptation control for acoustic echo cancellers, IEEE SIGNAL PROCESSING LETTERS, vol. 4, no. 6, pp , June [5] K. Shi and X. Ma, A frequency domain step-size control method for lms algorithms, IEEE SIGNAL PRO- CESSING LETTERS, vol. 17, no. 2, pp , February [6] S. Gustafsson, R. Martin, P. Jax, and P. vary, A psychoacoustic approach to combined acoustic echo cancellation and noise reduction, IEEE Trans. on Speech and audio processing, vol. 10, no. 5, pp , July [7] R. Martin and P. Vary, Combined acoustic echo control and noise reduction for hands-free telephony state of the art and perspectives, in Proc. European signal processing conference, pp , Setptember [8] R. L. B. Jeannes, P. Scalart, G. Faucon, and C. Beaugeant, Combined noise and echo reduction in hands-free systems: A survey, IEEE Trans. on Speech and audio processing, vol. 9, no. 8, pp , November [9] R. Martin, Noise power spectral density estimation based on optimal smoothing and minimum statistics, IEEE Trans. on Speech and audio processing, vol. 9, no. 5, pp , July [10] J. S. Lim and A. V. Oppenheim, The unimportance of phase in speech enhancement, IEEE Trans. on ASSP, vol. ASSP-30, no. 4, pp , August [11] Y. Ephraim and D. Malah, Speech enhancement using a minimum mean-square error log-spectral amplitude estimator, IEEE Trans. on Acoustics, speech and signal processing, vol. ASSP-32, no. 6, pp , December [12] I. Cohen, Optimal speech enhancement under signal presence uncertainty using log-spectral amplitude estimator, IEEE Signal processing letter, vol. 9, no. 4, pp , April [13] I. Cohen, Noise spectrum estimation in adverse environments: improved minima controlled recursive averaging, IEEE. Trans on Speech and audio processing, vol. 11, no. 5, pp , September [14] C. Faller and J. Chen, Suppressing acoustic echo in spectral envelop space, IEEE Trans. on Speech and audio processing, vol. 13, no. 5, pp , September [15] I. Tashev, Coherence based double-tal detector with adaptive threshold, in Proc. XX Scientific conference ELECTRONICS ET 2011, September [16] J. Allen and D. Berley, Image method for efficiently simulating small-room acoustics, Journal of Acoustical Society in America, vol. 65, pp , April [17] DSP group of Rice University, Noisex-92 database,. [18] J. J. Shyn, Frequency-domain and multirate adaptive filtering, IEEE Signal processing magazine, vol. 9, pp , January