MAI-Voice instant cloning: gated access and a 5-60 second clip

Can you clone a voice with MAI-Voice-2.1 today? Only after approval: Microsoft's Learn page says instant voice cloning is gated, needs an application through the Custom Neural Voice and Custom Avatar Limited Access Review, and works from a short clip, recommended at 5 to 60 seconds, with no training.

The page adds that only authorized, licensed voices can be synthesized in production and that no unlicensed voice cloning is possible. So the gate is part of the product, not a formality you can plan around.

The four steps Microsoft lists

The Learn page gives a fixed sequence for custom voices. Each step is a dependency in your schedule.

• Apply through the Limited Access Review, which is a separate approval from getting a Speech resource.
• After approval, use the personal voice APIs in the Speech SDK custom-voice samples.
• Upload an audio consent statement and the prompt audio to create a personal voice.
• Synthesize with a speaker profile ID inside an mstts:ttsembedding element, using the MAI-Voice model name.

How it compares with two other routes

Cloning appears across the voice tools in this week's launches, but the access terms differ. The Pocket TTS repository accepts plain WAV files as the voice input and carries an MIT license, so there is no approval step in the software, and the README prohibits cloning without explicit and lawful consent, which leaves checking consent and rights to your own process. Sume's route uses workspace avatars or the Voices library: you pass an avatar reference or a voice id, and Sume resolves the voice at submit time.

What to keep on file whichever route you take

A gate from the vendor does not replace your own record of who agreed to what. Keep the speaker's written consent, the scope of use, the date, and the clip you used. The consent-record checklist lists the fields.

Plan for the review time as a dependency. If a campaign voice must be cloned by a date, apply before you write the script, because the approval, not the audio generation, is the slow step.

If you only need a consistent branded narrator and not the exact voice of a person, a prebuilt voice avoids the gate entirely. MAI-Voice-2.1 ships licensed curated voices across its supported languages for that purpose.

Finally, check the language you need. The cloned voice carries across the supported languages according to Microsoft, but your approval, consent statement and licence terms should name every language and every channel where the cloned voice will be heard.