Gemini TTS: What the API Offers and What This Repo Uses

Everything below was verified against the live API (2026-07) rather than read off the docs page, because the docs page and the REST endpoint disagree — see The docs describe a different surface. When a fact here stops holding, re-probe; don't reason from this file's memory of the API any more than from your own.

Consumers in this repo: packages/robocar/unified/main/gemini_tts.c (the request), voice_persona.c (the table), dialogue_style.c (per-line variation), speech_tags.c (the tag allow-list), audio_player.c (playback).

Always use the streaming endpoint

:generateContent synthesises the entire utterance before sending a byte. Measured on one Finnish line, same body, same model: