Google Gemini 3.1 Flash TTS

Audio tags are the headline feature. You embed natural-language commands directly in the text, like [very fast] or [sarcastically, one slow word at a time], and the model performs accordingly. There's no fixed list, so you can write tags in plain words and the model does its best to interpret them. You can also add non-verbal sounds such as [cough], [sighs] or [gasp]. Short tags, big changes.

Gemini 3.1 Flash TTS covers more than 70 languages and offers regional accents instead of one flat register, so the same script can sound genuinely local rather than translated into a single generic voice. English, for example, includes accents like Southern US, California Valley, Brixton and standard British. The model detects the language of your input automatically, so you don't tag it by hand.

The model generates several voices in a single output and keeps each one consistent across turns. You assign each character an Audio Profile, then add Director's Notes to toggle pace, tone and accent per speaker. Long conversations don't drift into the same voice partway through.

Scene Direction sets the context for a whole recording. You describe the setting and the dialogue instructions. The model uses that world-building to keep characters in character. A scene framed as a tense interrogation room reads differently from a late-night convenience store, and the model reacts to those cues. Same script, different room. The delivery shifts.

On the Artificial Analysis TTS leaderboard, a benchmark built from thousands of blind human preferences, the model scored an Elo of 1,211, according to Google. The company calls it its most natural and expressive speech model to date. Artificial Analysis also placed it in its "most attractive quadrant" for the balance of quality and cost. Numbers only go so far, but that score puts it near the top.

Every clip generated by the model carries a SynthID watermark, Google's invisible and machine-verifiable label that rides along with the audio and can be confirmed by tools Google provides. It lets platforms and auditors confirm audio came from the model. For publishers worried about voice clones and misinformation, that's a built-in provenance check rather than an add-on.

Once you dial in a voice in AI Studio, you can export the whole configuration as Gemini API code. That keeps a character's sound consistent across projects, so you don't rebuild a voice from scratch every time you switch apps. Want the same narrator in two different products? Export once. Reuse it anywhere.