How kimang18/implementing-encodec-khmer works
This repository trains an EnCodec neural audio-compression model for speech tokenization. The main workflow is to configure a dataset and model, load and segment audio, encode and quantize it, then train against reconstruction and discriminator losses. The model also exposes tokenization and audio reconstruction capabilities. Edges reflect the README and shown imports/calls; some Lightning-managed training handoffs are not visible in the excerpts and are left unasserted.
Main entry points: train.py Training entry · encodec.py EnCodec model · seanet.py SEANet encoder and decoder · dataset.py Audio data module
Components 16
Training workflow
- Training config
configs/config.yaml - Training entry
train.py - Audio data module
src/dataset.py - Audio file helpers
src/utils.py
Codec model
- EnCodec model
src/modules/encodec.py - Residual vector quantizer
src/modules/quantizer.py - Standard RVQ variant
src/modules/standard_rvq.py
Network components
- SEANet encoder and decoder
src/modules/seanet.py - Convolution layers
src/modules/conv.py - Sequence LSTM
src/modules/lstm.py - STFT discriminator
src/modules/discriminator.py
Training support
- Training module
train.py - Loss balancer
src/balancer.py - Training logs and checkpoints
train.py
Other components
- Researcher
- Hugging Face dataset
Connections 15
- Researcher → Training entry: starts training
README.md - Researcher → Training config: sets options
README.md - Training entry → Training config: loads config
train.py - Training entry → Audio data module: creates data module
train.py - Audio data module → Hugging Face dataset: loads audio splits
src/dataset.py - Training entry → EnCodec model: creates generator
train.py - Training entry → STFT discriminator: creates discriminator
train.py - Training entry → Training module: assembles training
train.py - Training module → Loss balancer: balances losses
train.py - Training module → Training logs and checkpoints: logs and saves
train.py - EnCodec model → SEANet encoder and decoder: constructs encoder and decoder
src/modules/encodec.py - EnCodec model → Residual vector quantizer: quantizes latents
src/modules/encodec.py - SEANet encoder and decoder → Convolution layers: uses convolutions
src/modules/seanet.py - SEANet encoder and decoder → Sequence LSTM: uses sequence layers
src/modules/seanet.py - STFT discriminator → Convolution layers: uses 2D convolutions
src/modules/discriminator.py