How to clone your voice
The whole process takes about a minute of work and one short recording. Most of the advice below is about the recording, because that is the part that decides the result.
6 min read · Updated
Before you record
Three decisions, all of which are harder to change later than to get right now.
- Pick the language you will actually publish in. The clone inherits the language of the reference clip, and a mismatch is audible. If you publish in Hinglish, record in Hinglish — not in careful English.
- Use the microphone you will keep using. Matching the reference to your normal setup means generated audio sits inside your existing edits instead of on top of them.
- Record when you sound like yourself. Not last thing at night, not straight after shouting at a match. Everything in the clip transfers.
Recording the reference clip
Record in the browser or upload a file. Mayantar shows a script to read, timed to fit your language's limit at a natural pace — the same scripts are published on each language page if you want to rehearse.
The scripts are not tongue-twisters. They deliberately cover three things cloning models otherwise miss: relaxed conversational prosody, spoken digits and dates, and a question with rising intonation. Read one at normal speed, in a normal voice, about a hand's width from the mic.
The quality gate
Every reference clip is checked before it is used for anything. This is deliberately strict, because a marginal clip produces a marginal voice that you then blame on the model.
Hard failures — the clip is rejected and nothing is generated: too short, too quiet, clipped or distorted, or unreadable as an audio file. Soft warnings — generation proceeds but you are told: mild background noise or a low signal-to-noise ratio.
Rejection messages tell you what to change rather than what failed: record in a quieter room, speak closer to the mic, the clip is too short. If you get one, fix that one thing and re-record — it takes fifteen seconds. Full detail is in recording a good reference clip.
Confirming consent
Two checkboxes, both required: that you have the right to use this recording, and that you will not use the result to impersonate anyone. Both are stored against your account with a timestamp, and the server checks them independently of the interface.
If you are cloning someone else's voice — a colleague for a company video, a narrator you have hired — get their permission in writing first and keep it. The checkbox records that you asserted permission; it does not create permission.
Generating your first script
- Paste or type your script. Write it the way you speak: short sentences, contractions, the code-mixing you actually use.
- Pick the language. Devanagari text is Hindi; Roman-script Hindi is Hinglish. This is the most common cause of odd output.
- Pick your voice from the ones you have cloned.
- Adjust the tempo if the script needs to be slower for a tutorial or faster for a Short.
- Generate, then play it back. Crop the tail if you want, download the WAV, and use it.
If the result does not sound like you
- Re-record the reference before changing anything else. Nine times out of ten this is it.
- Check the language selector. English-path Hinglish is unmistakable once you know the sound.
- Read your script aloud. If it is awkward to say, it will be awkward to hear. Written-for-reading prose is the second most common cause.
- Shorten your sentences. Long clauses are where generated prosody drifts.
How many voices you can have
Voice slots depend on your plan — 1 on free, 5 on Starter, 20 on Premium and 50 on Premium Plus. Cloning your own voice requires a paid plan; the free tier generates with preset voices. Plan limits are on the pricing page.