RVC model store (rvc-models)
Model artifacts for the RVC v2 pipeline: deployable QNN graphs for the Snapdragon 8 Elite NPU, PyTorch training checkpoints, and retrieval indexes. Compiled graphs are treated as reproducible build artifacts; the checkpoints plus the stage scripts are the source of truth.
The code lives in separate repositories:
- rvc-android β RVC-only Android app (QNN/HTP)
- beatrice-android β unified Beatrice rc.0 + RVC v2 Android app
- beatrice-web β browser engine (Beatrice today, RVC ONNX path planned)
Layout
models/
packs/shared/ shared deployable graphs (HuBERT c0/r/enc, RMVPE, conv0_gn)
packs/voices/<name>/ per-voice encflow/flowreverse/decnsf + retrieval.index
+ voice.json + QNN graph manifests
source/ PyTorch checkpoints (.pth) + conv0_gn.npz
manifests/ build-only QNN manifests (hubert/, rmvpe/, synthesizer/)
scripts/
pipeline/ reproducible stage1 -> stage5 path + host runner
setup/ pretrained-weight download
tools/ probes and profiling helpers
archive/ superseded variants + one-off exploration scripts
voices/
<name>/ per-voice conversion notes and scripts
scratch/ ignored local data (index build, reference clone)
The JSON files next to each graph are QNN graph manifests (input/output shapes, tensor names); they are small and versioned in git.
Runtime arch
Context binaries are SoC-specific. Verified with
qnn-context-binary-utility (QAIRT 2.51.40):
| artifact | socModel | HTP |
|---|---|---|
models/packs/shared/* (HuBERT c0/r/enc, RMVPE) |
43 | v73 |
models/packs/voices/* |
43 | v73 |
Every pack is now v73, so one pack set covers 8 Gen 2/3 and runs on 8 Elite
through forward compatibility (v79 hardware executes v73 graphs; the
reverse is not possible). The current APK ships only the V79 runtime
(libQnnHtpV79Skel/Stub); an 8 Gen 2/3 target needs the matching V73/V75
Skel/Stub from QAIRT. v69 (8 Gen 1) needs a full recompile.
One pack was originally linked for v79. Switching it to v73 is a re-link of
the same DLCs (hub.submit_link_job) β no re-export needed β and both
builds produce bit-identical audio for the same input.
Compiled .bin graphs, ONNX exports, probes and test renders are not
kept locally: they are reproducible from the .pth checkpoints with the
stage scripts (QNN SDK required). hubert_base.pt is re-fetched by
scripts/setup/download_pretrained.py.
What is tracked
Git tracks the build scripts, JSON manifests, voice.json and
SHA256SUMS. The large artifacts (.bin / .pth / .pt / .index) are
mirrored to this HuggingFace repository instead; SHA256SUMS records their
checksums.
Build scripts
Run from the repository root. See scripts/README.md for the full table
and scripts/archive/README.md for superseded variants.
scripts/pipeline/stage1_*.pyβ export HuBERT / RMVPE / synthesizer ONNX.scripts/pipeline/stage2_sine_matmul.pyβ NSF sine/gate precompute.scripts/pipeline/stage3_hubert.py,stage4_rmvpe.py,stage5_synth.py,stage5b_decnsf_external.pyβ QNN compile stages.scripts/pipeline/pull_bins.shβ collect compiled binaries from qai-hub.scripts/pipeline/run_rvc_npu.pyβ end-to-end host runner for A/B checks.scripts/setup/download_pretrained.pyβ fetch the upstream weights.
License
Build scripts and documentation: GPL-3.0-only (see LICENSE). Model weights
and voice data retain their upstream terms; see NOTICE.