Add native MLX RVC voice-training initialization

9218cca verified 28 days ago

970 Bytes

license: mit
pipeline_tag: audio-to-audio
tags:
  - mlx
  - voice-conversion
  - rvc

MLX-RVC Serena e70

Native MLX RVC v2/48 kHz composite checkpoint for AI2Apps. It contains the MLX-trained e70 synthetic Serena test voice, its exact retrieval index, the shared ContentVec and RMVPE components, and the converted RVC v2/48 kHz generator/discriminator initialization used by the native MLX trainer. The checkpoint contains only safetensors and JSON; it does not require Torch, torchaudio, FAISS, or pickle at runtime.

The voice training corpus was generated with the serena preset of Qwen3-TTS and is intended as a built-in functional voice, not as the identity of a real person. Users must have permission to process the source audio they submit.

The implementation is derived from RVC at source revision 81eed5e8f68b6bed1789f682fe78cdd324495afc. The selected voice is the native MLX safe-adaptation epoch-70 checkpoint chosen by listening comparison.