A voice model is 90% about the dataset. Garbage in, garbage out — the model can never sound better than the audio you train it on. This guide shows how to prepare it right, step by step, with our tools.
| 5–8 min | our sweet spot — fits the limit and gives a quality model. |
| ~9–10 min | upper edge for WAV — see the limit below. |
| <3 min | too little — the voice will be flat. |
⚠️ Upload limit — 50 MB. A 48 kHz mono WAV is ~5–6 MB per minute, so 50 MB holds about 8–9 minutes. A 14-minute dataset comes out to ~80 MB — it simply won’t upload. So aim for 5–10 minutes (ideally 5–8) of dense clean voice without pauses (Vocal Stitch strips them). 5–8 varied minutes is plenty for a good model.
Varied 5–8 minutes beat the same minutes of monotone, single-note reading. Make sure the dataset covers:
RVC’s weak spot is sibilants, “th” and the “r”. If they’re scarce or slurred in the data, the model mangles them. So deliberately add clear, unhurried “th”, “r”, “s / sh / ch” sounds — the easiest way is to record a few tongue-twisters, clearly and slowly:
Read them clearly and slowly — that gives the model clean samples of the hard sounds so it stops “swallowing” them. (Cloning a non-English voice? Use tongue-twisters in that language.)
Gather recordings of the target voice so that after cleanup you have 5–10 minutes of clean vocals: songs, podcasts, voice notes. The cleaner the source, the less cleanup later.
If the voice is already clean (a cappella or a mic recording) — skip this. If it’s a song — split out the vocals with any vocal remover and keep only the vocal stem, ideally with reverb removed.
🔗 vocalremover.org (free, in-browser)Drop the vocal files into our Vocal Stitch — it cuts the silence between phrases and joins everything into one continuous 48 kHz mono WAV. That’s your ready dataset.
🎚 Open Vocal Stitch →You can run it through the Track Analyzer — check duration and loudness, make sure there’s no clipping or dropouts.
🎛 Open Track Analyzer →Upload your finished dataset to Train a voice — that’s it. Training is fully automatic: every parameter (sample rate, method, training length) is tuned for you, nothing to choose. In a few minutes you get a model (available 6 hours — download it or use it right away in Change Timbre).
Quality depends only on the dataset — that’s why steps 1–4 are everything.
In Change Timbre: upload the vocal you want to re-voice, pick your model and set the Pitch — “No change” if both voices are the same gender, or “Male → female” / “Female → male” when converting across genders. Nothing else to tweak — the rest is automatic.
If the voice mangles sounds, that’s not a setting — it’s the dataset: go back to clean vocals and the Ж/З/Р sounds (see the block above).
⚖️ Use only your own voice or one you have the rights/permission for. Cloning someone’s voice without consent is prohibited.