Model formats
RVC .pth to .onnx: what conversion actually does
ONNX is a portable format for running a trained model outside the framework it was built in. Converting an RVC model to it is a deployment step, not an upgrade — and for most people it is unnecessary.
NiceVois does not export ONNX. Training here returns a .pth, a matching .index, and a ZIP. If you need ONNX, the .pth is the input a separate converter takes, so nothing here blocks you — but you will run that step elsewhere.
What each format is for
| Format | What it is | Use it when |
|---|---|---|
.pth | PyTorch weights, the native RVC training output | Normal use in any RVC interface |
.onnx | A portable exported graph for cross-runtime inference | Embedding in an app or runtime without PyTorch |
.index | The retrieval feature index | Alongside either of the above, where supported |
When conversion is worth it
- You are shipping voice conversion inside an application and do not want a PyTorch dependency.
- You are targeting a runtime with its own ONNX acceleration path.
- You need the model to run somewhere a full Python environment is impractical.
When it is not
If you are making covers, converting vocals, or using any standard RVC interface, you do not need ONNX. Those tools load the .pth directly, and converting first adds a step that can only lose information.
What you give up
An ONNX export is a frozen inference graph. It is not a training checkpoint, so you cannot resume or fine-tune from it, and support for the retrieval index varies by runtime — some ONNX paths drop index retrieval entirely, which changes how the output sounds. Export also fixes certain assumptions at conversion time, so a graph exported for one runtime is not guaranteed to behave identically in another.
Keep the original either way. Conversion is one-directional in practice: you cannot reconstruct a training-ready .pth from an .onnx. Archive the .pth and .index pair before exporting anything.
If you decide to convert
Conversion is done with a separate tool — several RVC distributions include an export option, and standalone converters exist. Whichever you use, verify the result by converting the same short audio clip with both the original .pth and the exported .onnx and comparing them. If the exported version sounds noticeably flatter, index retrieval is the usual reason.
Related
.pth vs .index explains the two files a training run produces, and the after-training guide covers loading them in common interfaces.
Train from your permitted audio and download the model, index, and complete ZIP.
Start Training