Multi-channel Narrow-band Deep Speech Separation with Full-band Permutation Invariant Training

A narrow-band speech separation method. Detailed introduction (with images and examples) about this work can be found in English or in Chinese.

Results

Speech Separation Performance Comparision with SOTA Speech Separation Methods for 8-Channel 2-Speaker Mixtures

Model	SDR	SI-SDR	NB-PESQ	WB-PESQ
Mixture	0.18	0.00	2.05	1.6
Oracle MVDR [1]	8.15	4.58	3.20	2.62
FaSNet-TAC [2]	12.81	12.26	2.92	2.49
prop.	13.89	13.26	3.31	2.87

[1] https://github.com/Enny1991/beamformers

[2] Yi Luo, Zhuo Chen, Nima Mesgarani, and Takuya Yoshioka. End-to-end Microphone Permutation and Number Invariant Multi-channel Speech Separation. In ICASSP 2020.

Requirements

# TorchMetrics
pip install git+https://github.com/quancs/metrics.git@personal

# PytorchLightning (version 1.5)
pip install https://github.com/PyTorchLightning/pytorch-lightning/archive/master.zip

# jsonargparse
pip install jsonargparse[signatures]

Generate rirs

Generate rirs using configs/rir_cfg_3.json, and the generated rirs are placed in dataset/rir_cfg_3.

python generate_rirs.py

Train

Train NBSS on the 0-th GPU with config file configs/ifp/fit-WavForm.yaml (replace the rir & clean speech dir before training).

python cli_ifp.py --config configs/ifp/fit-WavForm.yaml fit --trainer.gpus 0, --seed_everything 2 --model.exp_name "train"

Test

Test on the 0-th GPU. Different seeds for dataset will generate different wavs.

python cli_ifp.py --config logs/NBSS_ifp/version_66/config-test.yaml test --model logs/NBSS_ifp/version_66/hparams.yaml --ckpt_path logs/NBSS_ifp/version_66/checkpoints/epoch707_neg_si_sdr-13.7777.ckpt --trainer.gpus=0, --data.seeds="{'train':null,'val':2,'test':5}" --model.exp_name="test"

Citation

If you like this work and want to cite us, please use:

@article{quan_multi-channel_2021,
	title = {Multi-channel {Narrow}-{Band} {Deep} {Speech} {Separation} with {Full}-band {Permutation} {Invariant} {Training}},
	journal = {arXiv preprint arXiv:2110.05966},
	author = {Quan, Changsheng and Li, Xiaofei},
	year = {2021},
}