SYS/02 · NEURAL AUDIO
KrisCodec
Music-specialized neural audio codec, built from first principles, producing a custom .kris format.
GitHub
PyTorchRVQAudio DSP
- TESTS
- 55 passing
- DECODE
- ~7.4 ms
- QUANT
- RVQ
- ACTIVATION
- Snake
KrisCodec is a neural audio codec specialized for music, written from first principles to understand codec design end to end, from the DSP front end through vector quantization to the decoder.
KEY FACTS
- Snake activations and a Residual Vector Quantizer (RVQ) at the codec core.
- 55-test suite covering the model and training path.
- End-to-end training pipeline from raw audio to encoded format.
- Benchmarked at ~7.4 ms decode latency.
Codec architecture
The codec uses Snake activations, periodic activations suited to audio's oscillatory structure, and a Residual Vector Quantizer (RVQ) to discretize the latent into a compact, layered code. Output is a custom .kris format.
Training and validation
A full end-to-end training pipeline takes raw audio through the encoder, quantizer, and decoder. A 55-test suite covers the model components and the training path. Decode is benchmarked at roughly 7.4 ms.