Hebrew AI voice generator and voice cloning

• Reference recordings for this language are capped at 12 seconds, the same limit used throughout Mayantar.
• Unpointed text is read from context. Adding niqqud to a specific word forces a specific reading, which is worth doing for names and for homographs the sentence does not disambiguate.
• Final-form letters (ך ם ן ף ץ) are ordinary Hebrew and need no special handling; write normal text.
• Mixed Hebrew and Latin script is common in Israeli copy and is read as written, with the Latin material treated as foreign.
• Cloning is cross-lingual: a voice cloned from a clip in one language can generate in another on this path. It works, but it is not free — accent from the reference language carries over. Record in the language you intend to publish in unless a foreign accent is the point.

Mayantar ships written prompts for English and the Indian languages; for Hebrew the recorder offers the English prompts as a starting point, and you are better off reading Hebrew of your own. Anything works as long as it is you talking normally — introduce yourself, describe what you are working on, read a paragraph of your own script.

Aim to fill most of the 12-second limit, keep it in one take, and include a question if you can: rising intonation is the thing a clone most often lacks when the reference was a flat, careful read.

The checks your clip has to pass — and the specific reasons clips get rejected — are in our guide to recording a reference clip.