McNet

The official repo: McNet: Fuse Multiple Cues for Multichannel Speech Enhancement accepted by ICASSP 2023 (https://arxiv.org/pdf/2211.08872.pdf). Examples can be found at https://audio.westlake.edu.cn/Research/McNet.htm.

Table 1. Performance of offline speech enhancement.* means scores are quoted from the original papers.

Method	NB-PESQ	WB-PESQ	STOI	SDR
Noisy	1.82	1.27	87.0	7.5
MNMF Beamforming * [20]	-	-	94.0	16.2
Oracle MVDR	2.49	1.94	97.0	17.3
CA Dense U-net * [12]	-	2.44	-	18.6
Narrow-band Net [11]	2.74	2.13	95.0	16.6
FT-JNF [14]	3.17	2.48	96.2	17.7
McNet (prop.)	3.38	2.73	97.6	19.6

Table 2. Performance of online speech enhancement.

Method	NB-PESQ	WB-PESQ	STOI	SDR
Noisy	1.82	1.27	87.0	7.5
Narrow-band Net [11]	2.70	2.15	94.7	16.0
FT-JNF [14]	2.80	2.23	95.4	16.9
McNet (prop.)	3.29	2.67	97.2	19.0

Train & Test

Reminder: This project is built on the pytorch-lightning package, in particular its command line interface (CLI). To understand the commands below and config file, you need to have some basic knowledge about the CLI in lightning.

Train:

python McNetCLI.py fit --config config\mc_net_online.yaml

Test:

python McNetCLI.py test --config config\mc_net_online.yaml

If you want to use our pretrained model,

python McNetCLI.py test --config config/mc_net_offline.yaml  --trainer.gpus 0,1  --ckpt_path model_checkpoints/offline/epoch494_criteria18.78_sdr18.78.ckpt

Update

3.24 Add predict module

python McNetCLI.py predict --config config/mc_net_offline.yaml  --trainer.gpus 0,1  --ckpt_path model_checkpoints/offline/epoch494_criteria18.78_sdr18.78.ckpt