Dereverberation toolkit for the REVERB challenge

[-] This tool is for deep learning-based monaural speech dereverberation.
[-] We want to make an open source tool for the REVERB CHALLANGE dataset. 

[-] If you find any bugs in this tool, please contact me! Thank you!
    My e-mail: hshi.cca@gmail.com.

Models included in this tool

Frequency-domain

[x] DNN (Frames-level Loss, mapping)
[x] LSTM (Frames-level Loss, mapping)

Evaluation measures on test set

PESQ Results

System	Far room 1	Far room 2	Far room 3	Near room 1	Near room 2	Near room 3
Noisy	2.59	1.99	1.87	3.11	2.39	2.27
DNN (mapping)	2.80	2.40	2.29	3.09	2.72	2.63
Bi-LSTM (mapping)	2.89	2.47	2.36	3.17	2.82	2.70

SRMR Results

System	Far room 1	Far room 2	Far room 3	Near room 1	Near room 2	Near room 3	Real far	Real near
Noisy	4.59	2.97	2.73	4.50	3.74	3.57	3.19	3.17
DNN (mapping)	5.62	4.72	4.03	5.88	5.57	5.34	5.58	5.35
Bi-LSTM (mapping)	5.50	4.82	4.18	5.58	5.32	5.15	5.89	5.75

STOI Results

System	Far room 1	Far room 2	Far room 3	Near room 1	Near room 2	Near room 3
Noisy	84.69%	78.20%	71.31%	95.18%	92.32%	89.38%
DNN (mapping)	87.93%	84.72%	82.46%	92.23%	90.99%	89.48%
Bi-LSTM (mapping)	88.65%	85.00%	82.83%	93.07%	91.90%	90.14%

Todo

Frequency-domain

[ ] DNN (Frames-level Loss, masking)
[ ] DNN (Frames-level Loss, multi-target learning)
[ ] LSTM (Frames-level Loss, masking)
[ ] LSTM (Frames-level Loss, multi-target learning)
[ ] LSTM (Utterance-level Loss, mapping)
[ ] LSTM (Utterance-level Loss, masking)
[ ] LSTM (Utterance-level Loss, multi-target learning)
[ ] CNN (Frames-level Loss)
[ ] CNN (Utterance-level Loss)
[ ] U-net (Frames-level Loss)
[ ] U-net (Utterance-level Loss)

Time-domain

[ ] SEGAN
[ ] FCN
[ ] Conv-Tasnet

Data description

The data provided consists of a training set, a development test set, and a (final) evaluation test set.

SimData: utterances from the WSJCAM0 corpus [1], which are convolved by room impulse responses (RIRs) measured in different rooms. Recorded background noise is added to the reverberated test data at a fixed signal-to-noise ratio (SNR).
RealData: utterances from the MC-WSJ-AV corpus [2], which consists of utterances recorded in a noisy and reverberant room.

References:
[1] T. Robinson, J. Fransen, D. Pye and J. Foote and S. Renals, "Wsjcam0: A British English Speech Corpus For Large Vocabulary Continuous Speech Recognition", In Proc. ICASSP 95, pp.81--84, 1995
[2] M. Lincoln, I. McCowan, J. Vepa and H.K. Maganti, "The multi-channel Wall Street Journal audio visual corpus (MC-WSJ-AV): Specification and initial experiments", IEEE Workshop on Automatic Speech Recognition and Understanding, 2005

Acknowledge

I modified my codes based on below github:

1. https://github.com/yongxuUSTC/sednn
2. https://github.com/snsun/pit-speech-separation
3. https://github.com/kaituoxu/Conv-TasNet

hshi-speech / Dereverberation-toolkit-for-REVERB-challenge

readme