sunnymeadsound/SunnyRVC-models

SunnyRVC runtime models

The voice-independent networks Sunnymead Sound's SunnyRVC plugin runs: the content encoder and the pitch detectors, exported for its C++ engine. The plugin downloads these files itself the first time it opens, so there is nothing here to install by hand. training/ holds what RVC Studio, the companion training app, downloads the same way.

They are not drop-in replacements for the upstream checkpoints. The graphs are cut and shaped for the plugin (for example, hubert.onnx emits both the v1 and v2 features, and RMVPE's mel front end is left to the engine), and each build of the plugin pins these files by commit and SHA-256.

File	Network	Upstream	Licence
hubert.onnx, hubert.mlpackage	ContentVec feature extractor (HuBERT base)	ContentVec, as distributed with Applio	MIT
rmvpe.onnx, rmvpe_trunk.mlpackage, rmvpe_head.onnx	RMVPE pitch estimator (whole; and split for Core ML)	RMVPE, rmvpe.pt as distributed with RVC and Applio	Apache-2.0
crepe_full.onnx, crepe_full.mlpackage	CREPE (full)	torchcrepe	MIT
crepe_tiny.onnx, crepe_tiny.mlpackage	CREPE (tiny)	torchcrepe	MIT
fcpe.onnx	FCPE pitch estimator	FCPE, fcpe.pt as distributed with Applio	MIT

RVC Studio (training/)

The same networks for the training app's MLX engine, and the pretrained bases training starts from. RVC Studio fetches the networks and the 48 kHz base on first use, and another rate's base when someone trains at that rate.

File	Contents	Upstream	Licence
training/mlx/contentvec.safetensors	ContentVec feature extractor	hubert_base.pt as distributed with RVC	MIT
training/mlx/rmvpe.safetensors	RMVPE pitch estimator	rmvpe.pt as distributed with RVC	Apache-2.0
training/mlx/crepe_full.safetensors	CREPE (full)	torchcrepe	MIT
training/pretrained_v2/f0G{32k,40k,48k}.safetensors, f0D…	RVC v2 pretrained generator and discriminator	RVC's pretrained_v2	MIT

Each network keeps its upstream licence; LICENSES.md has every licence in full and the changes made in converting them.