road2018 / upit-for-speech-separation

Speech separation with utterance-level PIT experiments

Speech Separation with uPIT

Speech separation with utterance-level PIT(Permutation Invariant Training)

Requirements

see requirements.txt

Usage

1. Generate dataset using create-speaker-mixtures.zip

2. Prepare cmvn, .scp and configure experiments in .yaml files

3. Training:

   ./run_pit.py --config $conf --num-epoches 100 > $checkpoint/train.log 2>&1 &

4. Inference:

   ./separate.py --dump-dir cache $mdl_dir/train.yaml $mdl_dir/epoch.40.pkl egs.scp
   

Experiments

Configure	Mask	Epoch	FM	FF	MM	FF/MM	AVG
config-1	AM-ReLU	75	10.41	6.73	7.35	7.19	8.82
config-2	AM-sigmoid	50	9.95	5.99	6.72	6.35	8.26
config-3	PSM-ReLU	73	10.29	6.54	7.28	7.09	8.71
config-4	PSM-ReLU	80	10.37	6.59	7.29	7.10	8.76
config-5	PSM-ReLU	62	10.58	7.00	7.55	7.40	9.01
config-6	PSM-ReLU	62	10.47	7.44	7.78	7.69	9.10
config-7	PSM-ReLU	61	10.43	7.17	7.41	7.34	8.91
-	IAM-oracle	-	12.49	12.73	11.58	11.88	12.19
-	IBM-oracle	-	12.94	13.20	12.04	12.35	12.65
-	IRM-oracle	-	12.86	13.14	11.96	12.27	12.57
-	PSM-oracle	-	15.79	16.03	14.90	15.20	15.50

Reference

• Kolbæk M, Yu D, Tan Z H, et al. Multitalker speech separation with utterance-level permutation invariant training of deep recurrent neural networks[J]. IEEE/ACM Transactions on Audio, Speech and Language Processing (TASLP), 2017, 25(10): 1901-1913.