Multi-channel Narrow-band Deep Speech Separation

A multichannel speech separation method. The official repo of:
[1] Changsheng Quan, Xiaofei Li. Multi-channel Narrow-band Deep Speech Separation with Full-band Permutation Invariant Training. In ICASSP 2022.
[2] Changsheng Quan, Xiaofei Li. Multichannel Speech Separation with Narrow-band Conformer. In Interspeech 2022.
[3] Changsheng Quan, Xiaofei Li. NBC2: Multichannel Speech Separation with Revised Narrow-band Conformer. arXiv preprint arXiv:2212.02076.

Audio examples can be found at https://audio.westlake.edu.cn/Research/nbss.htm. More information about our group can be found at https://audio.westlake.edu.cn.

Requirements

pip install -r requirements.txt

# gpuRIR: check https://github.com/DavidDiazGuerra/gpuRIR

Generate rirs

Generate rirs using configs/rir_cfg_4.json, and the generated rirs are placed in dataset/rir_cfg_4.

python generate_rirs.py

Train & Test

This project is built on the pytorch-lightning package, in particular its command line interface (CLI). Thus we recommond you to have some knowledge about the CLI in lightning.

Train NBC2 on the 0-th GPU with config file configs/NBC2_small.yaml or configs/NBC2_large.yaml (replace the rir & clean speech dir before training).

python NBSSCLI.py fit --config=configs/NBC2_small.yaml \
 --data.batch_size=[2,2] \ # batch size for train and val
 --trainer.accumulate_grad_batches=1 \
 --trainer.gpus=0,

More gpus can be used by appending the gpu indexes to trainer.gpus, e.g. --trainer.gpus=0,1,2,3,.

Configs configs/NBC-fit.yaml and configs/NB-BLSTM-fit.yaml can be used to train and test NBC and NB-BLSTM in the same way respectively. But mind to change the number of utterances for training in one mini-batch. As we use ddp for distributed training, the number of utterances in one mini-batch = num of gpus used * the number of utterances for dataloader * accumulate_grad_batches. In the above command, we have 2 utterances in one mini-batch, i.e. 1 2 1.

Resume training from a checkpoint:

python NBSSCLI.py fit --config=logs/NBSS/version_x/config.yaml \
 --data.batch_size=[2,2] \
 --trainer.accumulate_grad_batches=1 \ 
 --trainer.gpus=0, \ 
 --ckpt_path=logs/NBSS/version_x/checkpoints/last.ckpt

where version_x should be replaced with the version you want to resume.

Test the model trained (Dataset with different seeds will generate different wavs):

python NBSSCLI.py test --config=logs/NBSS/version_x/config.yaml \ 
 --ckpt_path=logs/NBSS/version_x/checkpoints/epochY_neg_si_sdrZ.ckpt \ 
 --trainer.gpus=0, \ 
 --data.seeds="{'train':null,'val':2,'test':3}" \ 
 --data.audio_time_len="['headtail 4', 'headtail 4', 'headtail 4']"

where headtail is the speech overlap way and it can be mid, full, or startend (please refer to [3]).

Module Version

see models/arch/NBSS.py