Source Separation

Followed by the first speech separation and recognition challenge in 2006, the more realistic scenario was released as SiSEC and CHiME in 2011 where additional realistic background noise were added to GRID sentences recorded in a reverberant environment. In the following contributions, we investigated the trade-off between a speech enhancement and noise estimation and speaker-dependent source separation algorithm. We further extended the idea to binaural scenario by combining a GMM-based model-driven speech enhancement as a postfilter after the beamforming and noise suppression stages. For a recent overview on the phase-aware single-channel source separation we refer to "Ch. 5 in the book".

Multisource reverberant environment at SNR = 3 (dB):

Clean Speech Noisy Speech
Enhanced Target Speech Separated Noise