the reason FastMNMF2 is not good ---- remove sqrt()

KyleZhang1118 / Voice-Separation-and-Enhancement

A framework for quick testing and comparing multi-channel speech enhancement and separation methods, such as DSB, MVDR, LCMV, GEVD beamforming and ICA, FastICA, IVA, AuxIVA, OverIVA, ILRMA, FastMNMF.

141 stars 34 forks source link

the reason FastMNMF2 is not good ---- remove sqrt() #3

Open qiansichong opened 3 years ago

qiansichong commented 3 years ago

https://github.com/KyleZhang1118/Voice-Separation-and-Enhancement/blob/4e98a2a6cb7d860cabe0b4cb2c2c8c122a11bb52/Process_FastMNMF2.m#L104

KyleZhang1118 commented 3 years ago

Sorry I just noticed your issue and really thanks for your correction. It's indeed a mistake. The idea of jointly-diagonaliztion seems good to model the reverberation, but the performance of FastMNMF is not ideal compared to the results in published paper. It is really confused.