JusperLee/Conv-TasNet

Based on the README, here's a technical summary: Uses dilated convolutions with temporal convolutional blocks (TCN) and global layer normalization to directly separate speech waveforms end-to-end, bypassing traditional time-frequency masking approaches. Implements configurable encoder-decoder architecture with skip connections, supporting both causal and non-causal modes, trained on WSJ0 speaker mixtures using PyTorch's DataLoader with multi-GPU distributed training. Provides pre-trained models on Hugging Face and includes batch inference pipelines for both audio file collections (via .scp files) and single WAV inputs.

535 stars. No commits in the last 6 months.