How many songs do I need to record?

At least ten. More material — and more variety in dynamics and register — generally produces a more convincing voice adapter. You can mix karaoke-style recordings made in the app with files you already have.

Do my voice recordings get uploaded anywhere?

No. Recording, training and the finished voice adapter all live on your own machine. The model of your voice exists nowhere but your disk — that's the point of doing it locally.

Do I need a GPU to train?

Training runs on Apple Silicon's GPU cores and takes a coffee break, not a weekend. The live console shows you exactly where the 500 epochs stand.

ANALOGALCHEMY / GUIDES / YOUR AI VOICE

Train an AI to sing in your voice — locally.

GUIDE · UPDATED JUNE 2026 · ONE EVENING, START TO FINISH

Most "AI voice" tools want your recordings on their servers. This guide does the opposite: by the end of it you'll have a personal voice adapter — a small AI model of your singing voice — trained on your own Mac, stored in your own music folder, and used to render any song you generate from then on. Nothing about your voice ever leaves your disk.

One rule before we start: this is for your voice. Cloning someone else without consent is off-limits — legally, ethically, and by design here.

STEP 1

Pick the styles you actually sing

Open Voice and choose your genres — arabesk, pop, rock, whatever is honestly yours. The app builds the session playlist from this, so the songs you'll record sit in territory your voice already knows.

Genre picker for the voice training session playlist — STEP 1 · GENRES — THE PLAYLIST BUILDS ITSELF

STEP 2

Set up the session

Tell the wizard your comfortable starting note and vocal type. This keeps the karaoke material inside your range — straining for notes you don't have teaches the model a voice you don't want.

New voice session setup with starting note and vocal type — STEP 2 · SETUP — STAY IN YOUR RANGE

STEP 3

Record — or upload — ten or more songs

Two roads, same destination: sing karaoke-style into the app, or drop in recordings you already have. Ten songs is the floor; variety is what raises the ceiling. Mix quiet verses with belted choruses, different tempos, different moods. A decent microphone in a quiet room beats a great microphone in a noisy one.

STEP 4

Review the dataset

Before training, the app shows every captured sample with its caption and lyrics. Listen through and cut the weak takes — one mumbled bridge won't ruin the model, but consistently sloppy material trains a consistently sloppy voice. Quality in, quality out.

Review dataset screen with captions and lyrics for each sample — STEP 4 · REVIEW — CUT THE WEAK TAKES

STEP 5

Train — and watch it learn

Hit train and the lab takes over: a live console, a falling loss curve, 500 epochs counting up — all running on your Mac's GPU cores. It's a coffee break, not a weekend. There is no upload step because there is nowhere to upload to.

Voice adapter training console with live loss curve and epoch counter — STEP 5 · TRAINING — LIVE CONSOLE, 500 EPOCHS, ALL LOCAL

STEP 6

Sing everything

Done. From now on, any song you generate — tonight's arabesk, next week's synth-pop experiment — can render in your voice. The adapter lives in your music folder like any other file: yours to back up, yours to delete, nobody else's to hold.

Questions people ask

How many songs do I need?

Ten minimum. Variety in dynamics, tempo and register matters as much as count.

Is anything uploaded?

No — recording, training and the finished voice model all stay on your machine.

Can I train someone else's voice?

No — it's designed for your own voice, with your own recordings, by deliberate choice.

Your voice is the instrument.

Voice training is fully unlocked in the free 14-day trial. No account, nothing uploaded.

Download the free trial SEE PRICING — $49 EARLY-BIRD

MACOS 14+ · APPLE SILICON · WINDOWS COMING SOON

MORE GUIDES: STEM SEPARATION · AI REPAINT · TEXT-TO-SONG SETTINGS