IndexTTS-2.5-onnx

IndexTTS-2.5 zero-shot voice cloning, exported to ONNX fp32 — bit-exact against the PyTorch CPU reference. Torch-free inference via the companion package index-tts-2.5-onnx:

uvx index-tts-2.5-onnx synth --ref voice.wav \
    --text "大家好, this is IndexTTS." --out out.wav

Runs anywhere ONNX Runtime runs (Linux / Windows / macOS, x86 / ARM). CPU by default; --device cuda with onnxruntime-gpu for NVIDIA GPUs.

Verification (vs PyTorch CPU reference)

• every stage cosine = 1.0000; greedy acoustic tokens exactly reproduced (fx0 73/73, fx1 83/83)
• vocoder output mel-SNR ≈ 85 dB (inaudible)
• no int8: dynamic QInt8 flips GPT greedy argmax (26% tokens match); MatMulNBits weight-only int8 is both less accurate (cos 0.985) and slower on ORT CPU — fp32 only, by measurement

Layout

campplus.onnx                    speaker embedding
emo_vec.onnx                     emotion vector
semantic_model/                  w2v-bert semantic encoder (+ .onnx.data)
semantic_codec_decode.onnx       acoustic codes -> latent
gpt_prefill/  gpt_step/          autoregressive GPT (+ .onnx.data)
length_regulator.onnx
cfm_estimator/                   flow-matching DiT
bigvgan.onnx                     vocoder -> 22050 Hz wav
multilingual_zh_ja_yue_char_del.tiktoken   tokenizer
spk_proj.npz                     speaker projection
hf_cache/w2v-bert-2.0/           feature-extractor configs only

Sibling packages: index-tts-2.5-mlx (Apple Silicon GPU, faster than real-time), index-tts-2.5-mnn (fast CPU).

Model weights are subject to the original Bilibili IndexTTS license. Use voice cloning responsibly and only with consent from the voice owner.