yuhogun0908 / MISOnet

Unofficial Multi-microphone complex spectral mapping for utterance-wise and continuous speech separation(MISO-BF-MISO)

MISOnet

Unofficial Pytorch Multi-microphone complex spectral mapping for utterance-wise and continuous speech separation(MISO-BF-MISO) https://arxiv.org/abs/2010.01703

Todo

• MISO1 implementation (seperation Network)
• Speaker Alignment System
• MVDR implementation
• MISO3 implementatino (enhancement Network)
• Speaker counting Network
• SMS-WSJ Dataset generation
• LibriCSS Dataset generation

Requirements

• Python>=3.8.0
• Pytorch>=1.10.0
• (optional) virtualenv

Training

0. (Optional) Setup Virtualenv

sudo pip3 install virtualenv
virtualenv -p python3 venv
source venv/bin/activate

1. Setup python packages environments

pip install -r requirements.txt

2. Run (todo)

python run.py --config=./config

3. Spectrogram # Example of 3_441c040w_445c040o_0.wav amoung test_eval92 (sms_wsj)

• Obervation

• Clean Source 1 & 2

• MISO1 Model Output Source 1 & 2

• MVDR Beamformer Output Source 1 & 2

• MISO3 Model Output Source 1 & 2

Reference

https://github.com/kaituoxu/Conv-TasNet https://github.com/fgnt/sms_wsj https://github.com/chenzhuo1011/libri_css