Using your model
How to make an AI cover with your own RVC model
A trained model converts a voice, not a song. Almost every disappointing first attempt comes from feeding in a finished mix instead of an isolated vocal. The sequence below is the one that works.
Before anything else: only publish or share covers you have the rights to. A trained voice model does not grant permission to use someone's voice, and the underlying song has its own separate rights. See our acceptable use policy.
Step 1 — Separate the vocal
Split the source track into an isolated vocal and an instrumental. Keep both: the vocal is what you convert, and the instrumental is what you mix back under it afterwards. Work from the highest-quality source you have, because separation artefacts carry straight through the conversion.
Step 2 — Convert the vocal
Load your .pth and its matching .index into an RVC interface, select the isolated vocal as input, and convert. If the model does not appear in the list, see the after-training guide — it is nearly always a folder-depth problem.
Step 3 — Get the pitch right
This is the setting that decides whether the result sounds like a voice or a chipmunk. Pitch is measured in semitones, and twelve semitones is one octave.
| Situation | Try |
|---|---|
| Source and model in a similar range | 0 |
| Lower-voiced source, higher-voiced model | +12 |
| Higher-voiced source, lower-voiced model | -12 |
| Nearly right but slightly strained | Adjust a semitone or two from there |
The index ratio is the other lever. Higher values lean on the retrieval index and pull the result toward the training data; lower values track the input more literally. Start mid-range and adjust once you have heard it.
Step 4 — Mix it back
Place the converted vocal over the original instrumental and line the two up exactly — conversion should not shift timing, but always check. From there it is ordinary mixing: level, a little compression, and reverb to seat the voice in the track. A converted vocal almost always arrives drier than the original vocal was.
When the result is poor
- Robotic or garbled: the input was probably a full mix, not an isolated vocal.
- Wrong octave: pitch shift, per the table above.
- Doesn't resemble the voice: usually the dataset, not the conversion — see dataset preparation.
- Thin or artefact-heavy: check the
.indexis actually loaded, not just the.pth. - Breathy or unstable on held notes: the model may be undertrained; see the epoch guide.
What NiceVois does and does not do
We train the model and give you the files. We do not separate vocals, convert audio, or produce covers — those steps happen in whichever tool you prefer. The files we return are standard RVC v2 outputs and work in any compatible interface.
Upload permitted audio and download the .pth, .index, and complete ZIP.
Start Training