Bbkblo/DhVaani-0.5-ONNX

DhVaani-0.5 — ONNX (torch-free)

ONNX export of ARTPARK-IISc/DhVaani-0.5 (ZipVoice fine-tune, 27 Indic languages, zero-shot voice cloning).

Runtime: onnxruntime + numpy + scipy + soundfile — no PyTorch.

Files

file	size	what
text_encoder_int8.onnx	6 MB	character encoder + duration expand
fm_decoder_int8.onnx	119 MB	flow-matching velocity (CFG baked in)
vocoder_backbone.onnx	50 MB	Vocos ConvNeXt backbone
vocos_head.npz	2 MB	ISTFT head weights (numpy)
mel_fb.npz	—	torchaudio-HTK mel filterbank
tokens.txt	—	1058-char Indic vocab
text_encoder.onnx / fm_decoder.onnx	18 / 456 MB	fp32 (optional)

I/O matches k2-fsa ZipVoice ONNX:

• text encoder: tokens, prompt_tokens, prompt_features_len, speed → text_condition (1,T,100)
• fm decoder (one Euler step): t, x, text_condition, speech_condition, guidance_scale → v (1,T,100)

Quick start

pip install onnxruntime numpy scipy soundfile
python dhvaani_torchfree.py \
  --prompt-wav ref.wav --prompt-text "transcript of ref" \
  --text "आज भारत विज्ञान और तकनीक क्षेत्र में बहुत तेज़ी से आगे बढ़ रहा है।" \
  --onnx-dir . --num-step 16 --out out.wav

CPU (2 cores): ~4× RTF at 16 steps int8; ~3× RTF at 8 steps.

Credits

Model: ARTPARK-IISc / ZipVoice (k2-fsa). Apache-2.0. Vocoder: charactr/vocos-mel-24khz.