vasyadeva/StyleTTS2-UA-Denys

StyleTTS2 Ukrainian (multispeaker) — ONNX export

An ONNX export of patriotyk/styletts2_ukrainian_multispeaker for on-device inference (ONNX Runtime, e.g. iOS). This repository adds no new training — it is a faithful export of patriotyk's PyTorch weights, plus a small graph fix so the model loads in ONNX Runtime.

All credit for the model, training data, and Ukrainian front-end belongs to @patriotyk. Please cite and support the original work.

Source

• Model: patriotyk/styletts2_ukrainian_multispeaker
• Inference / export code: patriotyk/styletts2-inference
• Front-end (stress + IPA): patriotyk/ukrainian-word-stress, patriotyk/ipa-uk
• Demo: patriotyk/styletts2-ukrainian

What was changed

1. Exported the PyTorch model to ONNX with torch.onnx.export (opset 19), using patriotyk's own export_onnx.py procedure.
2. Applied the one required graph fix: a Transpose node whose perm contained -1 (which ONNX Runtime rejects) was rewritten to 2. This is the same patch patriotyk's export script applies. The file here is already patched — no post-processing is needed to load it in ORT.

Files

file	description
model.onnx	the exported + patched model (330 MB)
config.yml	model config (vocab, params) — from the source repo
style_denys.pt / style_denys.bin	the "Денис Денисенко" speaker style vector (s_prev, 256 float32), as torch tensor and raw little-endian floats

Inputs / outputs

name	type	shape
tokens	int64	[seq_length]
speed	float32	scalar
s_prev	float32	[1, 256] (a speaker style vector)
output_wav	float32	[seq_length] — 24 kHz mono PCM

Tokens are produced by patriotyk's front-end (ukrainian-word-stress → ipa-uk → tokenizer over the config.yml vocab).

License

MIT, inherited from the source model. See LICENSE.