ViiTorVoice-NAR → Qualcomm Hexagon NPU (SM8750)
QNN context binaries (.bin) for running the ViiTorVoice-NAR text-to-speech / voice-cloning
chain entirely on a Snapdragon 8 Elite (SM8750) Hexagon NPU: w2v-BERT semantic encoder →
dualcodec encode-core → 32-step masked-diffusion NAR → dualcodec decoder.
Compiled with QAIRT 2.50 / qai-hub (onnx → dlc → qairt-dlc-prepare → .bin), tested on
Android 16 with a com.viitorvoice.demo app that loads these through QnnContext_createFromBinary.
Files
| group | files | notes |
|---|---|---|
| semantic frontend | enc_fe.bin (1.50 GB), fe_mel.bin |
w2v-BERT, fp16 single context, chunked 32×6k frames at inference |
| encode-core | enc_e1..e3, enc_front, enc_e4b, enc_e5, enc_e6 (104 MB total) |
7 linked segments; enc_front runs 32× per encode |
| NAR backbone | llm.bin (894 MB), head.bin (57 MB) |
Qwen3-0.6B, multi-shape shared weight link (weights stored once) |
| embeddings | text_emb.bin, sem_emb.bin, aco_emb.bin (165 MB) |
text embedding table + 12 codebook embeddings, gathered on CPU |
| decoder | dec_p1..p5.bin (138 MB) |
5-graph chain; p1 must stay fp16, p2–p4 w8a16, p5 uses chunk=48 |
| config | nar_steps |
NAR diffusion steps, 32 |
Hub: https://huggingface.co/mistmlm/viitorvoice-nar-sm8750 (private) — push_hf.sh syncs this dir.
See MANIFEST.md for sizes + sha256, install.sh to copy the set into an Android app's
private dir (/data/data/<pkg>/files/models) with correct owner/SELinux label.
Precision policy
- No fp32 anywhere in a compiled graph — QNN rejects it.
- Weights w8a16 (int8 weight / int16 activation) except the two places where it degrades quality audibly: the w2v frontend and decoder segment 1, both fp16.
- Post-training quantization is done offline (QDQ export + folding weight-side Q with numpy),
not via compile-time PTQ — the latter trips a
PostLayout::addTensorbug on these graphs.
Usage
Everything runs offline on device; no HF access needed at runtime:
./install.sh com.viitorvoice.demo # copies into the app's private dir (needs root)
The app calls QnnContext_createFromBinary once per file and chains them with its own
vv_encode / vv_nar / vv_pipeline C++ code.
Provenance
- Source models: ViiTorVoice (Apache-2.0), Qwen3-0.6B backbone (Apache-2.0), dualcodec.
- Pipeline scripts:
~/projects/viitorvoice/pipeline(qai-hub submit + link + download). - Derived artifacts only; upstream weights are not redistributed here beyond what the QNN contexts embed.