Clone your voice once. Use it across scripts.

Cloning is zero-shot: nothing is trained or fine-tuned on your voice. The model reads your reference recording at generation time and conditions its output on it. That is why it takes seconds rather than hours, and why the reference clip has more influence on the result than anything else you do.

1. Record or upload a reference clip

Record straight in the browser, or upload a file. Mayantar shows you a script to read that is timed to fit the limit for your language at a natural pace — the scripts are on each language page if you want to rehearse first.

2. The clip is checked before anything is generated

Every reference passes a quality gate. Clips that are too short, too quiet, clipped, distorted or unreadable are rejected outrightwith a message telling you what to fix — “record in a quieter room”, “speak closer to the mic”. Mild background noise produces a warning and proceeds. This is why results are reproducible: a bad clip fails fast instead of quietly producing a bad voice you then blame the model for.

3. Consent is confirmed and recorded

Two checkboxes, both required, both stored against your account with a timestamp: that you have the right to use this recording, and that you will not use the result to impersonate anyone. The server re-checks consent independently — the interface cannot skip this step, and neither can an API call.

4. Generate as many scripts as you like

The voice is saved and reusable. Type or paste a script, pick the language and the voice, adjust speaking tempo if you want it slower or faster, and generate. You can crop the result, download the WAV and share it.

Being specific about this is more useful than claiming it is indistinguishable from you, which no honest vendor can promise.

It works well for: scripted narration, explainer and faceless video, course and training audio, podcast intros and inserts, product walkthroughs, long-form reading where consistency across a hundred clips matters more than any one take, and re-recording a line you got wrong without setting the mic back up.

It is weaker at: high-emotion performance, comic timing, character acting, singing, and any delivery where the point is an unusual choice rather than a natural read. It also inherits whatever is in your reference clip — record tired and you get a tired-sounding voice for every script afterwards.

It cannot invent what it never heard. A reference clip in one language is a worse starting point for output in another; the mismatch is audible. Record your reference in the language you intend to publish in.

Voice cloning is the one AI capability where the misuse case is obvious and the victims are identifiable. We treat that as a product constraint rather than a policy page:

• No consent, no clone. Consent is required at submission and re-validated server-side, with the confirming account and timestamp stored on the voice record.
• Submitted clips are screened. Reference recordings are checked against known voices; matches are held for human review rather than auto-approved.
• Generated audio stays watermarked. We keep the model watermark rather than removing it.
• There is a real abuse route. A monitored grievance address is in the footer of every page, and takedown is a documented process, not an inbox that nobody reads.

The full detail — retention, deletion, what we screen for — is on the safety and consent page.

The free tier is a real trial of the generation quality, not a locked demo: 3 minutes and 3,000 characters a month, using preset voices, with watermarked output. Cloning your own voice starts on Starter at ₹299/month, which includes 60 minutes and 5 voice slots.

Full limits per plan are on the pricing page. Billing is in rupees through Razorpay including UPI, or in dollars through PayPal.