Signal Compaction Using PEVD for Spherical Array Processing with Applications

This website will demonstrate the PEVD Signal Compaction Framework for Spherical Arrays developed by Vincent W. Neo, Christine Evers, Stephan Weiss and Patrick A. Naylor. This is a joint work between Imperial College London, University of Southampton, and University of Strathclyde.

The framework is demonstrated for 2 example applications: Blind Speech Enhancement and Informed Source Separation.

About This Work

Clean speech signals are taken from the TIMIT corpus. Room impulse response (RIR) and noise signals are taken from the 32-channel Eigenmike recordings from the ACE corpus. Simulated RIRs are generated using the SMIRGen tool. Noisy reverberant speech is generated by convolving the anechoic speech with the RIR before adding diffuse noise at the required SNR for each microphone.

The speech enhancement approaches used in the comparison include eigenbeams (SHT), PEVD on the microphone signals (PEVD), KLT applied to the eigenbeams (SHT+KLT), 2 version of the the proposed SHT+PEVD using order 1 (PEVD L1) and order 2 (PEVD L2) approximation.

The source separation approaches used in the comparison include Maximum Directivity Index Beamformer (MaxDir), auxiliary function-based independent vector analysis (AuxIVA), independent low-rank matrix analysis (ILRMA), fast multi-channel non-negative matrix factorization (FastMNMF), our proposed PEVD using spherical harmonics and modal beamformer outputs.