BSS-MSR of MONC Circular Array Multi-Party Recordings

This page is the setup for subjective evaluation of our proposed blind source separation algorithm via model-based sparse recovery (BSS-MSR). The experiments are conducted on multi-channel original numbers corpus recorded at Idiap using an 8 channel circular microphone array in a moderately reverberant meeting room set-up. To compare with the alternative techniques for convolutive demixing, beamforming results has been also provided. The details of this study has been published in:

Multi-party Speech Recovery Exploiting Structured Sparsity Models, Afsaneh Asaei, Mohammad J. Taghizadeh, Hervé Bourlard and Volkan Cevher, International Speech Commuinication Association, INTERSPEECH 2011.

Techniques Sample1 Sample2 Sample3 Sample4 Sample5 Sample6
Reference Distant Microphone
Reference Close Microphone
Reference Lapel Microphone

Demixing Techniques
Superdirective Beamforming
Near-field Superdirective Beamforming


We further conducted speech recognition on MONC data base using Super-Directive (SD) beamformer, SD followed by McCowan post-filtering (SD-M), Speech dereverberation by Room Impulse Response (RIR) estimated and fitted Image Model in Least Squares sense (RIR-LS) as well as Room Acoustic Modeling via joint Sparse Recovery (RAM-SR). Please find the results as follow