GitHub - ZitengWang/nn_mask: multichannel linear filters based on mask estimation neural networks for CHiME4

ZitengWang / nn_mask Public

Notifications You must be signed in to change notification settings
Fork 18
Star 35

multichannel linear filters based on mask estimation neural networks for CHiME4

35 stars 18 forks Branches Tags Activity

Name		Name	Last commit message	Last commit date
Latest commit History 25 Commits
extention		extention
fgnt		fgnt
simulation		simulation
LICENSE		LICENSE
README.md		README.md
RESULTS		RESULTS
beamform.py		beamform.py
beamform.sh		beamform.sh
chime_data.py		chime_data.py
mic_error		mic_error
nn_models.py		nn_models.py
nn_models_sa.py		nn_models_sa.py
train.py		train.py
train_sa.py		train_sa.py

Repository files navigation

forked from https://github.com/fgnt/nn-gev , follow the instructions therein for installation

extentions: beamformers (MVDR, GEV, GEV-BAN, Variable Span, SDW-MWF, rank-1 MWF with different constraints ) used in Rank-1 Constrained Multichannel Wiener Filter for Speech Recognition in Noisy Environments and ON SDW-MWF AND VARIABLE SPAN LINEAR FILTER WITH APPLICATION TO SPEECH RECOGNITION IN NOISY ENVIRONMENTS

Recognition results, see RESULTS for all

	CHiME4 official backend		without LM rescore
	et-simu	et-real	et-simu	et-real
Noisy	-	-	14.15	23.52
WDAS	10.91	11.47	14.20	15.04
MVDR	5.93	7.30	8.70	10.31
GEV-BAN	6.29	7.25	9.17	10.48
GEV	6.98	7.14	10.01	10.53
MWF	10.81	12.92	12.54	16.16
VS-rank1	4.15	7.35	6.37	10.22
r1MWF-0	4.56	8.36	7.03	11.40
r1MWF-1	4.59	8.34	7.07	11.44
r1MWF-\mu_G	5.34	7.27	8.00	10.33
r1MWF-\mu_G-evd	5.17	6.51	7.63	9.25
r1MWF-\mu_G-gevd	4.68	6.03	6.87	8.74
gevd-SDW-MWF	4.39	7.38	-	-
gevd-GEV-BAN	4.21	6.17	-	-
gevd-GEV	4.65	5.93	-	-