Projects / Generation & pattern formation

Neural audio codec experiments

A SoundStream/EnCodec-style VQ-VAE with original additions: wavelet upsampling layers, self-organizing codebooks, and an energy-transformer bottleneck that outperformed RVQ.

  • Wavelet layers (original): upsampling layers that implicitly learn a wavelet decomposition, since sound is made of waves
  • Causal multiresolution convolutions
  • Self-organizing-map codebooks, via quantization-maps
  • An energy-transformer bottleneck, via energy_transformer
  • Discriminator changes that prevent collapse, and training one discriminator at a time, weighted by difficulty