Modelpublic

yaanfpv/svara-tts-voiceclone-beta-mlx-bf16

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes41downloads

Model Card

Svara TTS VoiceClone Beta MLX BF16

This is an unquantized BF16 MLX conversion of `kenpath/svara-tts-voiceclone-beta` for Apple Silicon local inference with mlx-audio and the `ComfyUI-Svara-TTS` helper backend.

The source model is an experimental Svara-TTS beta model for Indic reference-audio voice cloning. It is based on `kenpath/svara-tts-v1`.

Conversion

—Source model: kenpath/svara-tts-voiceclone-beta
—Conversion tool: mlx_lm.convert
—Saved dtype: bfloat16
—Quantization: none
—Intended runtime: mlx-audio on Apple Silicon
—Tested integration: ComfyUI-Svara-TTS MLX Audio helper

Example conversion command:

bash

python3 -m mlx_lm convert \
  --hf-path /path/to/kenpath--svara-tts-voiceclone-beta \
  --mlx-path ./svara-tts-voiceclone-beta-mlx-bf16 \
  --dtype bfloat16

Use With ComfyUI-Svara-TTS

In ComfyUI-Svara-TTS, choose the VoiceClone Beta model bundle and the MLX Audio local helper on macOS. The helper uses this converted model for Apple Silicon inference.

The model still requires the Svara/SNAC-style audio token pipeline provided by mlx-audio; this repository is the converted language model weights, not a standalone ComfyUI node pack.

Responsible Use

Use only reference audio that you have the right and consent to use. This is a beta voice-cloning model, not a guaranteed exact speaker replica. Results can vary by language, recording quality, reference clip, text length, and sampling settings.

Do not use this model for impersonation, deception, fraud, harassment, or other harmful uses.

Attribution

Original model and model card by Kenpath:

—Source model: https://huggingface.co/kenpath/svara-tts-voiceclone-beta
—Base model: https://huggingface.co/kenpath/svara-tts-v1
—Inference repo: https://github.com/Kenpath/svara-tts-inference

The original model is released under Apache-2.0.