VieNeu-TTS v3 Turbo, ONNX build for vieneu_tts
The files vieneu_tts downloads at
first run. This is a repackaging, not a new model: ONNX exports of
pnnbao-ump/VieNeu-TTS-v3-Turbo, the tied embedding tables pulled out as
flat float32 so they can be scored from Dart, and the sea-g2p Vietnamese
dictionary.
Both the model and sea-g2p are Apache 2.0 and both are by pnnbao. Credit for the model belongs there; only the packaging is mine.
What is here
| Path | What it is |
|---|---|
vieneu_prefill.onnx, vieneu_decode_step.onnx |
Qwen3 backbone, prompt pass and per-frame step |
vieneu_backbone_shared.data |
The backbone's weights, shared by both graphs |
vieneu_acoustic_cached.onnx |
Acoustic head, one call per codebook |
heads/ |
Tied embedding tables as raw float32, plus the speaker projection |
codec/ |
MOSS audio tokeniser decoder — codes to 48kHz audio |
g2p/sea_g2p.bin |
Vietnamese pronunciation dictionary |
voices.json |
Twenty preset speaker embeddings |
About 273MB in total. The engine memory-maps what it can, but expect a working set of roughly 650MB while generating.
Using it
await const ModelDownload(
baseUrl: 'https://huggingface.co/p424p424/vieneu-tts-v3-turbo-onnx/resolve/main',
).fetch(dir);
A warning about versions
g2p/sea_g2p.bin must match the sea-g2p engine version compiled into the
package. Mixing a dictionary from one version with an engine from another does
not fail — it produces fluent audio saying the wrong words. This dictionary is
from 0.9.0.
- Downloads last month
- 16