Voice-over (text to speech) — estimate

A word for the director on the delivery: "motivational, builds to a shout", "calm explainer", "funny, deadpan"

"auto" (the default) is always the newest model, eleven_v4 — the most emotive, speaks Hebrew, follows audio tags like [excited]. Name an older id only for a call that depends on it; Hebrew text always runs on eleven_v4, whatever is named

Hebrew: the voice director reads the text first (on by default) — niqqud on every Hebrew word (English words, brands and acronyms stay in Latin letters), an English delivery tag per sentence ([softly, smiling], [firmly], [yelling furiously, out of breath]…), a stability that fits; the words never change. Text that already carries your own tags or niqqud is reviewed, not skipped: your tags stay where they are and the words get their niqqud (a word you vocalised keeps your points when they fit the sentence). false = read exactly what you send

Kept for older calls and ignored: every voice-over runs on the newest model ("auto")

GET /v1/ai/voices — a stock voice, or a voice you cloned (POST /v1/ai/voices/clone)

The text that comes right after — the voice ends this one leading into it

Also return every word with its start/end (seconds) — to cut the take at its pauses, place it exactly, and time captions without a transcript

Force a language, e.g. he

The text read just before this one (the previous passage) — the voice continues its flow instead of starting fresh