DSP Algorithms for Ad-Hoc Wireless Transducer Array Networks

The quality of a speech signal recorded through multiple microphones can be compromised by various undesired signals, including noise and reverberation, leading to diminished intelligibility. Some of these undesired signals may be generated by loudspeakers within the same acoustic environment, potentially generating acoustic echoes and/or feedback. Multi-microphone speech enhancement systems aim at suppressing or canceling these interference signals while maintaining the desired speech signal undistorted. The growing tendency to incorporate multiple microphones and loudspeakers across various devices has significantly boosted the capabilities of audio and speech processing technologies. Intensive digital signal processing (DSP) tasks can be performed, almost effortlessly, in small, portable devices. When multiple devices of this kind are connected, new possibilities emerge in terms of signal processing algorithms, where cooperation between the devices plays a key role. When multiple devices or nodes with microphones and loudspeakers are connected via wireless links a so-called wireless acoustic sensor and actuator network (WASAN) is formed. WASANs then have access to more spatially distributed microphones and loudspeakers, and nodes can benefit from exchanging parameters or signals between them to improve their performance in a specific DSP task.

This thesis will focus on the development and evaluation of acoustic signal processing algorithms in WASANs. Specifically, as opposed to collecting all sensor (microphone) and actuator (loudspeaker) signals in one common place, e.g., in a fusion center, novel distributed DSP algorithms are proposed. In these kinds of algorithms, nodes compress or fuse their local sensor and actuator signals, and exchange them with the other nodes in the network to share the computational burden of centralized processing using in-network processing.

In this thesis, first, a comparison between multi-channel Wiener filter (MWF) implementations for noise reduction (NR) using overlap-save (OLS) and weighted overlap-add (WOLA) filter banks is provided. It serves as an introduction to centralized MWF and justifies the typical use of WOLA filter banks for NR in 

speech applications. Then, distributed algorithms for combined acoustic echo cancellation (AEC) and NR in WASANs are presented, where each node may have multiple microphones and multiple loudspeakers, and where the desired signal is a speech signal. Centralized cascade and integrated algorithms are presented to solve the combined AEC and NR. The distributed cascade and integrated algorithms for combined AEC and NR are then obtained from their corresponding centralized implementations. The MWF and its distributed implementation, i.e., the generalized eigenvalue decomposition (GEVD)-based distributed adaptive node-specific signal estimation (DANSE) algorithm, serve as a common thread through the DSP tasks in this and the remaining chapters. The centralized and distributed solutions are compared to stand-alone solutions, namely a node working only with its local microphone and loudspeaker signals.

Second, centralized and distributed algorithms for combined acoustic feedback cancellation (AFC) and NR are presented. In this part, the signal processing tasks are performed within a closed-loop system. Three cascade algorithms are presented for combined AFC and NR, where one of them is based on an extended data model that includes the loudspeaker signal in the MWF formulation. This allows to design a cascade algorithm consisting of a NR stage first followed by a single-channel AFC stage in a multi-microphone single-loudspeaker scenario. Based on this centralized implementation a distributed algorithm for combined AFC and NR is designed to obtain node-specific estimates in a WASAN.

Finally, the last chapter concludes the thesis, summarizes the contributions of this thesis and provides possible future research directions.