Dataset
RVC dataset preparation: how much audio do you need?
The most common cause of a disappointing model is not too few epochs. It is a dataset that was never going to work. Duration matters, but consistency matters more, and the two are easy to confuse.
Duration targets
Around ten minutes of clean, varied speech from one speaker is a realistic floor for a usable model. Ten to thirty minutes is the range most guidance settles on. Beyond roughly an hour the returns flatten sharply, and past about two hours extra audio changes very little.
| Amount | What to expect |
|---|---|
| Under 5 minutes | Usually thin and inconsistent; recognisable at best |
| 5–10 minutes | Workable if the audio is very clean and varied |
| 10–30 minutes | The practical sweet spot for most voices |
| 30–60 minutes | Better coverage of range and emotion, if quality holds |
| Over 2 hours | Diminishing returns; preparation quality dominates |
The NiceVois beta accepts up to 15 minutes. That sits inside the usual sweet spot, but it does mean the preparation advice below matters more here than it would with an unlimited local run. Fifteen well-chosen minutes will outperform fifteen careless ones by a wide margin.
A well-prepared twenty minutes beats a messy hour
The model learns whatever is in the file. If a second voice appears, it learns that too. If there is room echo throughout, it learns the room. Length cannot compensate for contamination — it multiplies it.
What to remove
- Other voices. Interviews, duets, and crowd noise are the single most damaging inclusion.
- Background music. Even quiet backing bleeds into the learned timbre.
- Long silences. They add no information and inflate the duration figure misleadingly.
- Clipping and distortion. Overloaded peaks cannot be repaired by training.
- Heavy reverb. Room character is learned as part of the voice and is hard to remove afterwards.
What to include
Variety within one speaker is what you are after: different sentences, a natural range of pitch and volume, and more than one emotional register. A dataset of one sentence repeated is long but narrow, and the model will sound narrow too. If the model is intended for singing, include sung material rather than only speech.
A practical sequence
- Gather the cleanest available recordings of a single speaker.
- Separate the vocal from any music, and discard anything that will not separate cleanly.
- Cut every passage containing another voice.
- Trim long silences and any distorted sections.
- Check the total: aim for at least ten minutes of material you would be happy to listen to.
Sanity check: listen to the finished dataset end to end once. Anything that makes you wince is something the model is about to learn.
Then choose epochs
Dataset size and training length interact — a short dataset trained too far tends to overfit rather than improve. The epoch guide covers choosing a sensible number, and the audio requirements page covers accepted formats.
Upload it and download the finished .pth, .index, and complete ZIP.
Start Training