PEVD-Based Speech Enhancement in Wireless Acoustic Sensor Networks

The ubiquity of smart devices equipped with microphones in modern environments has opened the door to performing audio processing tasks such as speech enhancement over networks of microphones instead of traditional compact arrays. However, wireless acoustic sensor networks (WASNs) also introduce geometry estimation, clock synchronisation, and latency compensation as often necessary steps prior to traditional enhancement techniques. Instead, this paper investigates the use of blind polynomial eigenvalue decomposition (PEVD)-based enhancement in such networks with an emphasis on its robustness to communication latency between nodes. Simulations and experiments on real data show that the network-based PEVD enhancement consistently gives equivalent or better signal-to-noise ratio (SNR) than the best node in the network, even when experiencing communication latency. In-depth analysis of the network geometry and the PEVD filterbanks provides insights into the strengths and shortcomings of PEVD-based enhancement in networks, and shows it is robust to the issue of reference microphone selection. Finally, evaluation of PEVD as a pre-processing step to automatic speech recognition (ASR) shows an improvement of 7 points in word error rate (WER) over multichannel Wiener filtering for very noisy real conversational data.