Springer Topics in Signal Processing

Transcription

1 Springer Topics in Signal Processing Volume 3 Series Editors J. Benesty, Montreal, Québec, Canada W. Kellermann, Erlangen, Germany

2 Springer Topics in Signal Processing Edited by J. Benesty and W. Kellermann Vol. 1: Benesty, J.; Chen, J.; Huang, Y. Microphone Array Signal Processing 250 p [ ] Vol. 2: Benesty, J.; Chen, J.; Huang, Y.; Cohen, I. Noise Reduction in Speech Processing 240 p [ ] Vol. 3: Cohen, I.; Benesty, J.; Gannot, S. (Eds.) Speech Processing in Modern Communication 360 p [ ]

3 Israel Cohen Jacob Benesty Sharon Gannot (Eds.) Speech Processing in Modern Communication Challenges and Perspectives ABC

4 Prof. Israel Cohen Technion - Israel Institute of Technology Dept. Electrical Engineering Haifa Technion City Israel icohen@ee.technion.ac.il Dr. Sharon Gannot Bar-Ilan University School of Engineering Ramat-Gan Bdg Israel gannot@eng.biu.ac.il Prof. Dr. Jacob Benesty Université de Quebec Inst. National de la Recherche Scientifique (INRS) 800 de la Gauchetiere Ouest Montreal QC H5A 1K6 Canada benesty@emt.inrs.ca ISBN e-isbn DOI / Springer Topics in Signal Processing ISSN Library of Congress Control Number: c 2010 Springer-Verlag Berlin Heidelberg e-issn This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Cover Design: WMXDesign GmbH, Heidelberg Printed in acid-free paper springer.com

5 Preface More and more devices for human-to-human and human-to-machine communications, where sound pickup and rendering is necessary, require some sophisticated algorithms. This is due to the fact that the acoustic environment in which we live in and communicate is extremely challenging. The difficult problems encountered in this environment are very well known and they are mainly acoustic echo cancellation, interference and noise suppression, and dereverberation. More than ever, these fundamental problems need to be tackled rigorously. This is the objective of this edited book, which contains twelve chapters that are briefly summarized below. Chapter 1 addresses the problem of linear system identification in the short-time Fourier transform (STFT) domain. Identification of linear systems is of major importance in diverse applications of signal processing, including acoustic echo cancellation, relative transfer function (RTF) identification, dereverberation, blind source separation, and beamforming in reverberant environments. In this chapter, the authors introduce three models for linear system identification and investigate the influence of model order on the estimation accuracy. The three models are based on either crossband filters between subbands, multiplicative transfer functions, or cross-multiplicative transfer functions. It is shown both analytically and experimentally that the estimation accuracy does not necessarily improve by increasing the model order. The problem of RTF identification between sensors is addressed in Chapter 2. This transfer function represents the coupling between two sensors with respect to a desired or interfering source. The authors describe an alternative representation of time domain convolution with convolutive transfer functions in the STFT domain, and show improved results compared to existing RTF identification methods. In low-cost hands-free telecommunication systems the loudspeaker signal may contain a certain level of nonlinear distortions, which necessitate nonlinear modeling of the acoustic echo path. Chapter 3 describes a novel approach for nonlinear system identification in the STFT domain. It introv

6 vi duces Volterra filters in the STFT domain and considers the identification of quadratically nonlinear systems. It shows that a significant reduction in computational cost as well as substantial improvement in estimation accuracy can be achieved over a time-domain Volterra model, particularly when long-memory systems are considered. Chapter 4 presents a family of non-parametric variable step-size (VSS) algorithms, which are particularly suitable for realistic acoustic echo cancellation (AEC) scenarios. The VSS algorithms are developed based on another objective of AEC application, i.e., to recover the near-end signal from the error signal of the adaptive filter. As a consequence, these algorithms are equipped with good robustness features against near-end signal variations, like double-talk. Speech enhancement in transient noise environments is addressed in Chapter 5. An estimate of the desired signal is obtained under signal presence uncertainty using a simultaneous detection and estimation approach. This method facilitates suppression of transient noise with a controlled level of speech distortion. Cost parameters control the tradeoff between speech distortion, caused by missed detection of speech components, and residual musical noise resulting from false-detection. Chapter 6 describes a model-based approach for combined dereverberation and denoising of speech signals. This approach is developed by using a multichannel autoregressive model of room acoustics and a time-varying power spectrum model of clean speech signals. Chapter 7 investigates separation of speech and music signals from singlesensor audio mixtures. It describes codebook approaches and a Bayesian probabilistic framework for source modeling and source estimation. The source models include Gaussian scaled mixture models, codebooks of auto regressive models, and Bayesian non negative matrix factorization (BNMF). Microphone arrays are becoming increasingly more common in the acquisition and denoising of acoustic signals. Additional microphones allow us to apply spatiotemporal filtering methods, which are significantly more powerful than conventional temporal filtering techniques. Chapter 8 is concerned with beamformer designs tailored to the specific nature of microphone array environments, i.e., broadband signals and reverberant channels. A distinction is made between wideband and narrowband metrics, and the relationships between broadband performance measures and the corresponding component narrowband measures are analyzed. Chapter 9 presents some new insights into the minimum variance distortionless response (MVDR) beamformer. It analyzes the tradeoff between dereverberation and noise reduction achieved by using the MVDR beamformer, and discusses relations between the MVDR and other optimal beamformers. Chapter 10 addresses the problem of extracting several desired speech signals from multi-microphone measurements, which are contaminated by nonstationary and stationary interfering signals. A linearly constrained minimum variance (LCMV) beamformer is designed with two sets of linear constraints:

7 one for maintaining the desired signals and one for mitigating both the stationary and non-stationary interferences. Spherical microphone arrays have been recently studied for spatial sound recording, speech communication, and sound field analysis for room acoustics and noise control. Complementary studies presented progress in beamforming methods. Chapter 11 reviews beamforming methods recently developed for spherical arrays, from the widely used delay-and-sum and Dolph-Chebyshev, to the more advanced optimal methods, typically performed in the spherical harmonics domain. Finally, Chapter 12 presents a family of broadband source localization algorithms based on parameterized spatiotemporal correlation, including the popular and robust steered response power (SRP) algorithm. It develops source localization methods based on minimum information entropy and temporally constrained minimum variance. This book has been edited for engineers, researchers, and graduate students who work on speech processing for communication applications. We hope that the readers will find many new and interesting concepts that are presented in this text useful and inspiring. We deeply appreciate the efforts, willingness, and enthusiasm of all the contributing authors. Without their commitment, this book would not have been possible. We would like to take this opportunity to thank again Christoph Baumann, Carmen Wolf, and Petra Jantzen from Springer (Germany) for their wonderful help in the preparation and publication of this manuscript. Working with them is always a pleasure and a wonderful experience. Finally, we would like to dedicate this edited book to our parents. vii Haifa/ Montreal/ Ramat-Gan Nov Israel Cohen Jacob Benesty Sharon Gannot

8 Contents 1 Linear System Identification in the Short-Time Fourier Transform Domain... 1 Yekutiel Avargel and Israel Cohen 1.1 Introduction Problem Formulation System Identification Using Crossband Filters Crossband Filters Representation Batch Estimation of Crossband Filters Selecting the Optimal Number of Crossband Filters System Identification Using the MTF Approximation The MTF Approximation Optimal Window Length The Cross-MTF Approximation Adaptive Estimation of Cross-Terms Adaptive Control Algorithm ExperimentalResults Crossband Filters Estimation Comparison of the Crossband Filters and MTF Approaches CMTF Adaptation for Acoustic Echo Cancellation Conclusions Appendix References Identification of the Relative Transfer Function between Sensors in the Short-Time Fourier Transform Domain Ronen Talmon, Israel Cohen, and Sharon Gannot 2.1 Introduction Identification of the RTF Using Multiplicative Transfer Function Approximation xi

9 x Contents Problem Formulation and the Multiplicative Transfer Function Approximation RTF Identification Using Non-Stationarity RTF Identification Using Speech Signals Identification of the RTF Using Convolutive Transfer Function Approximation The Convolutive Transfer Function Approximation RTF Identification Using the Convolutive Transfer Function Approximation Relative Transfer Function Identification in Speech Enhancement Applications Blocking Matrix The Transfer Function Generalized Sidelobe Canceler Conclusions References Representation and Identification of Nonlinear Systems in the Short-Time Fourier Transform Domain Yekutiel Avargel and Israel Cohen 3.1 Introduction Volterra System Identification Representation of Volterra Filters in the STFT Domain Second-Order Volterra Filters High-Order Volterra Filters A New STFT Model For Nonlinear Systems Quadratically Nonlinear Model High-Order Nonlinear Models Quadratically Nonlinear System Identification Batch Estimation Scheme Adaptive Estimation Scheme ExperimentalResults Performance Evaluation for White Gaussian Inputs Nonlinear Undermodeling in Adaptive System Identification Nonlinear Acoustic Echo Cancellation Application Conclusions Appendix References Variable Step-Size Adaptive Filters for Echo Cancellation 89 Constantin Paleologu, Jacob Benesty, and Silviu Ciochină 4.1 Introduction Non-Parametric VSS-NLMS Algorithm VSS-NLMS Algorithms for Echo Cancellation VSS-APA for Echo Cancellation VFF-RLS for System Identification

10 Contents xi 4.6 Simulations VSS-NLMS Algorithms for AEC VSS-APA for AEC VFF-RLS for System Identification Conclusions References Simultaneous Detection and Estimation Approach for Speech Enhancement and Interference Suppression Ari Abramson and Israel Cohen 5.1 Introduction Classical Speech Enhancement in Nonstationary Noise Environments Simultaneous Detection and Estimation for Speech Enhancement Quadratic Distortion Measure Quadratic Spectral Amplitude Distortion Measure Spectral Estimation Under a Transient Noise Indication A Priori SNR Estimation ExperimentalResults Simultaneous Detection and Estimation Spectral Estimation Under a Transient Noise Indication Conclusions References Speech Dereverberation and Denoising Based on Time Varying Speech Model and Autoregressive Reverberation Model Takuya Yoshioka, Tomohiro Nakatani, Keisuke Kinoshita, and Masato Miyoshi 6.1 Introduction Goal Technological Background Minimum Mean-Squared Error Signal Estimation andmodel-basedapproach Dereverberation Method Heuristic Derivation of Weighted Prediction Error Method Reverberation Model Clean Speech Model Clean Speech Signal Estimator and Parameter Optimization Combined Dereverberation and Denoising Method Room Acoustics Model Clean Speech Model

11 xii Contents Clean Speech Signal Estimator Parameter Optimization Experiments Conclusions References Codebook Approaches for Single Sensor Speech/Music Separation Raphaël Blouet and Israel Cohen 7.1 Introduction Single Sensor Source Separation Problem Formulation GSMM-Based Source Separation AR-Based Source Separation Bayesian Non-Negative Matrix Factorization Learning the Codebook Multi-Window Source Separation General Description of the Algorithm Choice of a Confidence Measure Practical Choice of the Thresholds Estimation of the Expansion Coefficients Median Filter Smoothing Prior GMM Modeling of the Amplitude Coefficients ExperimentalStudy Evaluation Criteria Experimental Setup and Results Conclusions References Microphone Arrays: Fundamental Concepts Jacek P. Dmochowski and Jacob Benesty 8.1 Introduction SignalModel ArrayModel Signal-to-Noise Ratio ArrayGain Noise Rejection and Desired Signal Cancellation Beampattern Anechoic Plane Wave Model Directivity Superdirective Beamforming WhiteNoiseGain Spatial Aliasing Monochromatic Signal Broadband Signal

12 Contents xiii 8.11 Mean-Squared Error Wiener Filter Minimum Variance Distortionless Response Conclusions References The MVDR Beamformer for Speech Enhancement Emanuël A. P. Habets, Jacob Benesty, Sharon Gannot, and Israel Cohen 9.1 Introduction Problem Formulation From Speech Distortion Weighted Multichannel Wiener Filter to Minimum Variance Distortionless Response Filter Speech Distortion Weighted Multichannel Wiener Filter Minimum Variance Distortionless Response Filter Decomposition of the Speech Distortion Weighted Multichannel Wiener Filter Equivalence of MVDR and Maximum SNR Beamformer PerformanceMeasures PerformanceAnalysis On the Comparison of Different MVDR Beamformers Local Analyzes Global Analyzes Non-Coherent Noise Field Coherent plus Non-Coherent Noise Field Performance Evaluation Influence of the Number of Microphones Influence of the Reverberation Time Influence of the Noise Field Example Using Speech Signals Conclusions Appendix References Extraction of Desired Speech Signals in Multiple-Speaker Reverberant Noisy Environments Shmulik Markovich, Sharon Gannot, and Israel Cohen 10.1 Introduction Problem Formulation Proposed Method The LCMV and MVDR Beamformers The Constraints Set Equivalent Constraints Set Modified Constraints Set

13 xiv Contents 10.4 Estimation of the Constraints Matrix Interferences Subspace Estimation Desired Sources RTF Estimation Algorithm Summary Experimental Study The Test Scenario Simulated Environment Real Environment Conclusions References Spherical Microphone Array Beamforming Boaz Rafaely, Yotam Peled, Morag Agmon, Dima Khaykin, and Etan Fisher 11.1 Introduction Spherical Array Processing Regular Beam Pattern Delay-and-Sum Beam Pattern Dolph-Chebyshev Beam Pattern Optimal Beamforming Beam Pattern with Desired Multiple Nulls D Beam Pattern and its Steering Near-Field Beamforming Direction-of-Arrival Estimation Conclusions References Steered Beamforming Approaches for Acoustic Source Localization Jacek P. Dmochowski and Jacob Benesty 12.1 Introduction Signal Model Spatial and Spatiotemporal Filtering Parameterized Spatial Correlation Matrix (PSCM) Source Localization Using Parameterized Spatial Correlation Steered Response Power Minimum Variance Distortionless Response Maximum Eigenvalue Broadband MUSIC Minimum Entropy Sparse Representation of the PSCM Linearly Constrained Minimum Variance Autoregressive Modeling Challenges Conclusions References

14 Contents xv Index

15 List of Contributors Ari Abramson Technion Israel Institute of Technology, Israel Morag Agmon Ben-Gurion University of the Negev, Israel Yekutiel Avargel Technion Israel Institute of Technology, Israel Jacob Benesty INRS-EMT, QC, Canada Raphaël Blouet Audionamix, France Silviu Ciochină University Politehnica of Bucharest, Romania Israel Cohen Technion Israel Institute of Technology, Israel Jacek P. Dmochowski City College of New York, NY, USA xvii

16 xviii List of Contributors Etan Fisher Ben-Gurion University of the Negev, Israel Sharon Gannot Bar-Ilan University, Israel Emanuël A. P. Habets Imperial College, UK Dima Khaykin Ben-Gurion University of the Negev, Israel Keisuke Kinoshita NTT Communication Science Laboratories, Japan Shmulik Markovich Bar-Ilan University, Israel Masato Miyoshi NTT Communication Science Laboratories, Japan Tomohiro Nakatani NTT Communication Science Laboratories, Japan Constantin Paleologu University Politehnica of Bucharest, Romania Yotam Peled Ben-Gurion University of the Negev, Israel Boaz Rafaely Ben-Gurion University of the Negev, Israel Ronen Talmon Technion Israel Institute of Technology, Israel Takuya Yoshioka NTT Communication Science Laboratories, Japan