How to clone your singing voice with AI: a recording guide.

Two recordings. First, one full song sung over a guided beat, usually two to three minutes, so the model hears you across a range of notes, vowels and volumes. Second, one short spoken line, a sentence the app gives you, so it can confirm that the singer and the speaker are the same person before it trains. That second step exists to stop people cloning voices that are not theirs; it is also the step most people rush.

What the model does not need: a studio, a condenser microphone, or a great performance. It needs a clean signal of your voice with as little else in it as possible. Consistency beats quality: the same mic, the same distance and the same room for both recordings matter more than either one being polished.

Headphones with a microphone are the single biggest upgrade, because the beat plays in your ear instead of into the mic. The beat is then absent from the take, and the model learns your voice rather than your voice plus a drum loop. Wired earbuds with an inline mic are fine; Bluetooth is fine if it is already paired.

Quiet room means no fan, no open window, no television in the next room. Background sound is filtered, but filtering has a cost: every removal takes a little of your voice with it. Silence is free.

Distance: a hand-width from the mouth, slightly off to the side so plosives do not pop. If the recording screen shows a level meter, sing your loudest line once before you start and set the distance so the meter stays out of the red. Then leave the phone where it is. Moving it mid-take changes the tone, and the clone will average the two.

Octa shows a colour band under the lyrics while you sing. A dot follows the note you are producing; it slides toward blue when you go lower and toward red when you go higher, with the note name next to it and a small number showing how far you are from the nearest note in cents. In tune means the dot sits on a line and the number stays inside roughly plus or minus twenty.

The band is not a judge. It is there so you can hear yourself with your eyes, which matters because the ear forgives its own mistakes. Two habits help: glance at the band at the start of each line, not on every word; and when the dot sits consistently above the line, do not push harder, relax the jaw and let the note settle. A voice that is a touch flat on every line trains as a flat voice, and the clone will inherit it.

Every so often a short hint appears: a touch sharp, ease down; hold the note steady; slow down, let the words breathe. These come from simple rules about your last twenty seconds, such as how far your notes drift from the nearest semitone and how many syllables you are landing per second against the lyric line. They are deliberately rare. If you see one, act on it for the next line and forget it.

After the song, the app shows a sentence and asks you to record it. This is speech matching, not a performance. Read it at a normal speaking volume, clearly, once, in the same room with the same mic. Singing it, whispering it, or reading it with the beat still playing are the three most common ways to fail this step.

If the sentence arrives in English and you sang in another language, that is expected: the matching step supports a fixed set of languages and falls back to English. Read it slowly and say every word. If the line is longer than about twenty seconds or shorter than two, the app will ask you to try again before anything is sent, which saves you a wait.

Rejections fall into four buckets, and the app now tells you which one you hit rather than a generic error.

Too quiet or too short. The take never reaches a usable level, or the spoken line is under two seconds. Fix: hold the mic closer, speak up, record the line again. The song does not need to be re-sung.

The line did not match. The speaker check could not connect the spoken sentence to the singing. Almost always one of: the line was sung, background music was running, or a different mic or room was used. Fix: get a fresh line and say it plainly. Again, the song stays.

A problem on our side. Sometimes the training step fails for reasons that have nothing to do with you. The app keeps your recording and asks you to try again in a few minutes.

Still working. Training can take longer than the app expected. It will say so and let you check again instead of pretending it failed.

The pattern in all four: the full song is the expensive part, and the app is built so you rarely have to redo it.

A good clone sings any song you generate afterwards in your voice, in the styles and languages the engine supports, at a consistent quality that does not depend on how you felt that morning. It keeps your timbre and your general way of shaping vowels. It does not keep your phrasing choices, your breath sounds or the way you bend a note into a chorus; those belong to a performance, not a voice.

Ownership is simple: the voice is bound to your account, used only for songs you generate, and deletable in one tap. Clone only your own voice. The spoken-line check exists to enforce that, and a clone made from someone else’s recording is both against the rules and, in most places, against the law.

When the result disappoints, the cause is almost always in the take, not in the model. Re-record with headphones, in a quieter room, watching the band. Most people get a clone they recognise on the second attempt.