Speech enhancement has been the backbone of audio processing since the early days of research. Whether used as pre-processing for automatic speech recognition and human-machine interactions, or directly in telecommunications and hearing devices, enhancement is essential to all speech processing systems. State-of-the-art methods have always followed advances in technology: with better processing power, algorithms evolved from spectral subtraction to large neural networks; and with the advent of MEMS microphones, from single-channel to multi-channel systems. Looking towards a future of internet-of- things, the next paradigm shift is likely from compact microphone arrays to wireless acoustic sensor networks (WASNs).

Modern WASNs require careful algorithmic design for enhancement due to issues of clock synchronisation, geometry estimation, and latency compensation. This thesis investigates a polynomial eigenvalue decomposition (PEVD)-based method as a practical solution to these challenges. Relying on an orthogonality assumption between target and noise subspaces, PEVD decorrelates microphone signals in space, time, and frequency. This thesis demonstrates the method’s potential in WASNs, and proposes a hybrid beamformer-PEVD approach to reduce computational complexity. The shift-invariant property of PEVD-based enhancement is also exploited to show robustness to latency in wireless networks. A drawback of the current PEVD-based implementation is its reliance on long-term statistics of noisy speech, typically computed over periods of stationary acoustic environments. This is uncommon in realistic scenarios with non-stationary sources or moving microphones, as in WASNs. Therefore, this thesis presents an adaptive frame-based implementation of the PEVD-based algorithm.

Many results in this thesis are obtained from simulated acoustic data. To assess practical viability, the methods are also evaluated on real recordings of a challenging group conversation in a restaurant. The data collection procedure is detailed, and enhancement performance is discussed. This thesis contributes to the practical implementation of PEVD-based speech enhancement in dynamic real-world conditions and outlines directions for future work.