About This Work

Clean speech signals are taken from the TIMIT corpus. Room impulse response (RIR) and noise signals are taken from the 3-channel mobile phone recordings from the ACE corpus. Additional diffuse noise are generated using signals from NOISEX database and International Sound Effects Library. 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 Log-MMSE, Single-channel subspace for coloured noise (COLSUB), Multi-channel subspace method (MCSUB), 2 versions of Multi-channel Wiener filter including a practical (MWF) and Oracle version (OMWF), Generalized Weighted Prediction Error (GWPE), Weighted Power minimisation Distortionless (WPD), the Integrated Sidelobe Canceller and Linear Prediction (ISCLP) Kalman filter, and the proposed (PEVD)-based algorithm [1].

[1] V. W. Neo, C. Evers, and P. A. Naylor, "Enhancement of noisy reverberant speech using polynomial matrix eigenvalue decomposition," IEEE/ACM Trans. Audio, Speech and Lang. Process., vol. 28, 2021. doi: 10.1109/TASLP.2021.3120630

Code

Code is now available on GitHub

Projects

1. PEVD of Spherical Harmonics for Speech Enhancement
2. PEVD-based Source Separation Using Informed Spherical Microphone Arrays
3. PEVD Preprocessing for Multichannel Voice Activity Detection
4. PEVD-based Target Speaker Voice Activity Detection
5. Fixed Beamformer Design Using PEVD