Voxtral Voice Clone fills encoder gap

Community reverse-engineering like this is what turns an open-weight demo into something builders can actually use. The catch is that this is still research-heavy plumbing, not a turnkey voice-cloning product.

• –The repo targets the exact blocker in Voxtral's TTS stack: no codec encoder weights means no custom speaker embeddings from reference audio.
• –It retrains the encoder and LoRA-distills the LLM, which is the right shape if you want embeddings that still land in Voxtral's expected voice space.
• –The 80GB GPU requirement and two-phase training pipeline make this a serious research project, not a casual local install.
• –The repo is CC BY-NC 4.0, and the derivative weights still inherit Mistral's license constraints, so commercial teams need to read the fine print.
• –Sources: [Reddit](https://www.reddit.com/r/LocalLLaMA/comments/1s6rmoi/the_missing_piece_of_voxtral_tts_to_enable_voice/), [GitHub](https://github.com/Al0olo/voxtral-voice-clone)