ds2svc 0.1.0

Restore SVC models disguised as OpenUTAU/DiffSinger voicebanks

ds2svc

Restore SVC models that have been disguised as OpenUTAU-compatible DiffSinger voicebanks, and run inference on the recovered model.

Some packed "DiffSinger" voicebanks in the wild are not DiffSinger at all — they are existing SVC models (RVC, DDSP-SVC) repackaged to run on a DiffSinger ONNX API. This harms the open community's trust and safety. ds2svc identifies such banks, extracts the genuine SVC model back out, and provides inference.

What it does

1. Detect — load a voicebank ZIP, read its yaml manifest, and inspect the ONNX graph internals (not just config) to decide whether it is a legitimate DiffSinger bank or a disguised SVC model.
2. Extract (in progress) — recover the SVC ONNX with clean inputs (cvec_units + f0).
3. Infer (planned) — Streamlit WebUI + realtime inference.

Supported disguises:

Source	Based on	Disguise	Genuine model recovered
UTAI	RVC	RVC packed as a "vocoder"	(units[768], f0) → waveform
DDSP-SVS	DDSP-SVC	DDSP-SVC packed inside the "acoustic" model	(units[768], f0, volume) → mel → pc-nsf-hifigan

Detection signatures

Detection trusts the ONNX graph, not the config (config fields are easy to fake, and unusual-but-legit banks exist):

• UTAI / RVC — the payload graph contains a normalizing flow (net_g/flow) beside net_g/enc_p and an NSF decoder net_g/dec: the RVC SynthesizerTrn.
• DDSP-SVS — the acoustic graph contains a real STFT op (legit DiffSinger acoustics have none) plus ddsp_model/unit2ctrl, istft_method, melext, and a backend_reflow ReFlow stage.
• DSP ops alone (Sin/Cos/CumSum/RandomNormalLike) are not used — they also appear in genuine DiffSinger diffusion backbones.

Usage

pip install -e .[infer,realtime]
ds2svc detect path/to/bank.zip
ds2svc extract path/to/bank.zip --out artifacts/<id>
ds2svc infer artifacts/<id> --in vocal.wav --out converted.wav --f0 fcpe
ds2svc realtime artifacts/<id>

See MANUAL_DOWNLOADS.md for offline weight installation.

Status

• Voicebank ZIP reader + manifest resolver (paths resolved relative to yaml)
• ONNX graph probe + detection (RVC / DDSP-SVS / legit)
• UTAI/RVC extraction — graph-surgery excision of the empty-index TopK chain, input rename mel→cvec_units. Output: (cvec_units[1,T,768], f0[1,T]) → waveform[1,N] at 44.1 kHz, net_g (encoder + flow + NSF decoder) preserved byte-for-byte.
• DDSP-SVS extraction — bit-lossless weight-port into a rebuilt Unit2Wav (CombSubSuperFast + RectifiedFlow/LYNXNet2), re-exported as a single ONNX (cvec_units[1,T,768], f0[1,T], volume[1,T]) → mel[1,T,128] (ReFlow loop baked in). Pair with pc-nsf-hifigan for audio; the mel→wav stage can later be merged into this graph for a one-shot audio-output model.
• Inference frontend — ContentVec (lengyue233/content-vec-best), f0 (fcpe/rmvpe), volume, pc-nsf-hifigan vocoder, and a unified load_model(ModelDir) → infer(audio, sr, key_shift, f0_method) that auto-selects RVC or DDSP path from ds2svc_meta.json.
• Realtime GUI — FreeSimpleGUI + sounddevice stream with sliding buffer, SOLA crossfade, correlation-based overlap alignment, f0=fcpe default. ds2svc realtime <model_dir>.
• Streamlit WebUI — drag-drop vocal WAV → convert → preview + download. Auto-caches model, persists output across re-renders. streamlit run ds2svc/webui/app.py
• Manual download guide for offline weight placement, with per-file URLs and directory layout.

See PLAN.md for the full design.

Development

Tests reference private voicebanks by anonymized IDs. Copy tests/fixtures/local_banks.example.json → local_banks.json (gitignored) and point the IDs at your local archives, then:

pytest

License

MIT. This project is for community safety and lawful recovery of open or self-owned models. Respect the licenses of upstream RVC / DDSP-SVC / DiffSinger.