Voice Models and Covers

Train a voice model from singing you have the rights to, then use it — or a catalog voice — to sing another song: what training removes from your recordings, recording guidelines, cover settings, and what each slider changes.

Last updated October 13, 2026

A voice model learns a singing voice from recordings. A cover uses a voice model to sing another song, keeping the original melody and timing. The two steps live on separate pages: the AI Singing Voice Generator trains a model, and the AI Song Cover Generator makes covers.

Train a voice model

Enter a name for the voice, and optionally add an image so you can recognise it later.
Add training audio: upload files or record directly. Up to 5 recordings can be combined.
Confirm that you own or have permission to train with this voice and these recordings.
Start training. It usually takes under half an hour; you can leave the page and come back. When it is ready, the voice's card shows how much clean voice it was trained on and what was removed.

What makes good training audio

At least 1 minute, about 10 minutes recommended. More clean singing gives the model more of the voice's range.
Finished songs work. A voice model learns everything left in its training audio, so before training each recording is cleaned to the dry lead voice: where there is a backing track the vocals are extracted and the harmonies split off (harmonies left in the lead are the main cause of cracked or hoarse notes), then reverb, steady background noise, faint separation residue and breaths are removed. A clean a cappella take keeps its original audio.
One singer. A second voice on the lead line, as in a duet, cannot be split off and blurs the model.
Singing rather than speech. Singing teaches the model pitch range and vibrato.
Accepted formats: MP3, WAV, OGG, M4A, AAC or FLAC.

Make a cover

Choose a voice: one of your own (My Voice), or a catalog voice from the tabs.
Choose the song: upload it, or pick one from My Music.
Adjust the settings if needed, then generate and compare Before and After.

The cover settings

Voice octave (under Advanced): Auto compares the singer's pitch with the voice's own range and moves the new vocal by an octave when that fits better, so a deep voice can take a high melody; Octave down, Same octave and Octave up choose it yourself. The instrumental never moves.
Song pitch transposes the whole mix by semitones. Use it when the song sits outside the voice's range; transposing the full mix can slightly reduce quality.
Main vocals / Instrumental / Reverb balance the converted vocal against the backing track; 0 dB keeps the original level.
Index rate sets how strongly the trained voice's character pulls the timbre; higher follows the target voice more closely.
Protect keeps the original's consonants and breaths natural; 0.5 turns the protection off.

Harmonies and backing vocals keep the voices they were sung in. Only the lead vocal is converted: it is split from the harmonies and de-reverbed first, then gets fresh reverb and goes back in at the original lead's level. An uploaded song keeps its stereo image.

Rights and privacy

Your trained models are private to your account. Only train with your own voice or one you have written permission to use, and only cover songs you have the rights to release. A download is not a licence, and a licence never covers a voice or recording you upload — see Commercial license.