Latency-Agnostic Speech Enhancement for Wireless Acoustic Sensor Networks Using Polynomial Eigenvalue Decomposition

Wireless acoustic sensor networks (WASNs) allow for ad-hoc placement of microphones in rooms such that some nodes may capture conversations with higher signal-to-noise ratio (SNR) than conventional compact arrays. While this makes WASNs useful tools for speech enhancement, microphone positions are generally unknown and the wireless link between nodes can introduce random latency in the network. These issues are particularly relevant for traditional enhancement techniques such as beamforming as they typically rely on having calibrated arrays with known geometries. This work studies the use of polynomial eigenvalue decomposition (PEVD) as a blind, latency-agnostic speech enhancement method in networks. PEVD has been used previously with various compact arrays and does not require knowledge of the network geometry to perform enhancement. Simulation results in reverberant shoe-box rooms show that PEVD is able to improve predicted intelligibility and quality of speech, even in the presence of random signal transmission latency.