Helwani K. Adaptive Identification of Acoustic Multichannel Systems Using Sparse Representations

Springer, 2015. — 120 p.
Adaptive signal processing is a key component of high-quality hands-free full-duplex acoustic communication. Prominent examples are algorithms for the suppression of unwanted noise sources and interferers, and for acoustic echo cancelation (AEC). Many of the well-established solutions to these problems involve the use of adaptive finite impulse response filters. High resolution spatial sound field reproduction synthesizes a desired sound field within a listening area. Closely spaced arrays of a large number (tens to hundreds) of individually driven loudspeakers are required for a physically accurate synthesis. Well-known techniques are Wave Field Synthesis (WFS) and Higher Order Ambisonics (HOA). Similar considerations also hold for an accurate analysis of sound fields. Typically, spherical microphone arrays with a high number of microphones are used due to their favorable properties. Advanced audio communication systems are increasingly using virtual auditory environments to improve the immersion of human-to-human telecommunication. An audio communication system utilizing these techniques thus implies the use of multiple-input and multiple-output (MIMO) adaptive filters. Similar problems arise also in other array-based applications such as medical imaging, seismics, etc. The limitations of conventional adaptive MIMO filters become apparent for systems with a high number of inputs and outputs (massive multichannel systems) and create the need for improving the known adaptation techniques. Typical challenges that have to be tackled in this context are nonuniqueness, ill-posedness and numerical complexity of the underlying system identification problem. The present dissertation treats the topic of extending the adaptive filtering theory in the context of massive multichannel systems by taking into account a priori knowledge of the underlying system or signal. The starting point of this book is exploiting the sparseness in acoustic multichannel systems in order to solve the nonuniqueness problem with an efficient algorithm for adaptive filtering that does not require any modification of the loudspeaker signals. The derivation of general sparse representations of acoustic MIMO systems is discussed in detail. In the present book, it is also shown that such sparse representations can be derived based on prior knowledge about the system or the signal. Efficient adaptive filtering algorithms that exploit the sparsity in the transform domains are presented and the relation between the signal- and the system-based sparse representations is emphasized. Furthermore, a novel approach to spatially preprocess the loudspeaker signals in a full-duplex communication system is presented. The idea of the preprocessing is to prevent the echoes from being captured by the microphone array in order to support the AEC system. The preprocessing stage is presented as an application of a novel unified framework for the synthesis of sound figures. Finally, a multichannel system for acoustic echo suppression is presented that can be used as a postprocessing stage for removing residual echoes. The presented approach in this book to acoustic echo suppression copes as a first of its kind with highly correlated loudspeaker signals of multichannel reproduction systems, does not require a double-talk detector, and constrains near-end signal distortion.