IndexTTS-2-MLX

IndexTTS-2-MLX

IndexTeam/IndexTTS-2 converted to MLX safetensors (fp32) for native Apple Silicon inference, with mlx-indextts.

IndexTTS-2 is a zero-shot voice-cloning TTS system with emotion control. It clones a voice from a single short reference clip (5–15 s) and synthesizes Chinese and English speech with optional explicit emotion vectors.

Model files

Auxiliary encoders (w2v-BERT 2.0, MaskGCT semantic codec, CAMPPlus) are loaded from their original upstream checkpoints at runtime; see the inference repo for details.

Use with mlx-indextts

git clone https://github.com/Jup33Q/mlx-indextts
cd mlx-indextts && uv sync

hf download Jup33QE/IndexTTS-2-MLX --local-dir models/IndexTTS-2-MLX

HF_HUB_OFFLINE=1 uv run python -m mlx_indextts.generate_v2 \
  --model-dir models/IndexTTS-2-MLX \
  --reference path/to/reference.wav \
  --text "你好，这是一段 MLX 原生推理的中文语音。" \
  --output out.wav

Conversion notes

• Converted from the official PyTorch checkpoints with mlx_indextts/convert_v2.py, fully offline from local weights.
• Layer-wise alignment vs PyTorch (MPS): end modules match within fp32 tolerance (BigVGAN full-model MAE 6.8e-3, emotion pipeline ~5e-4); a few intermediate layers show device-level numerical outliers that do not affect end-to-end output.
• Verified end-to-end on Apple M5 Max: Mandarin synthesis RTF ≈ 1.9 (fp32, batch).

License

IndexTTS-2 weights are released under the upstream IndexTTS license — see LICENSE.txt (English) and LICENSE_ZH.txt (Chinese), which are included in this repo and apply to the converted weights as well. The MLX conversion code lives in mlx-indextts under its own license.