VieNeu-TTS v3 Turbo β browser pack
Files repackaged so that VieNeu-TTS v3 Turbo can run entirely in a web browser with onnxruntime-web (WebAssembly). Nothing here was retrained; the weights are the upstream weights.
Sources
| Part | Upstream | Licence |
|---|---|---|
| Acoustic model, tokenizer, preset voices | pnnbao-ump/VieNeu-TTS-v3-Turbo, VieNeu-TTS | Apache-2.0 |
| Audio codec decoder | OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano-ONNX | Apache-2.0 |
| Text normalisation and grapheme-to-phoneme | sea-g2p | Apache-2.0 |
What was changed
turbo/*.f16: the external weight files stored as float16, half the download. The browser expands them back to float32 before creating the ONNX sessions. The backbone and the embedding table were already float16 values upstream, so that part is lossless; the codec decoder is rounded once.turbo/vieneu_acoustic_heads.onnx: upstreamvieneu_acoustic_cached.onnxwith five nodes appended that compute the logits of one codebook (Gatherthe last position,MatMulwith the codebook's embedding table). In Python that product is a NumPy call; in a browser it has to live inside the graph to be fast. The embedding table is external data (vieneu_acoustic_heads.f16).turbo/text_emb.f32,turbo/xvec.f32: arrays taken from upstreamvieneu_v3_heads.npz.turbo/voices.json: upstreamvoices_v3_turbo.json, same values, flat list.g2p/sea_g2p.wasm: sea-g2p 0.10.0 built forwasm32-unknown-unknownthrough its C ABI, with a small patch so the dictionary can be handed over as bytes instead of a memory-mapped file.g2p/sea_g2p.dictz: upstreamsea_g2p.bin, gzip-compressed.
v1/manifest.json lists every file with its size and SHA-256.