Google unveils Gemini 3.1 Flash TTS: a speech model with fine-grained control, dialogue and 70+ languages

Google has put Gemini 3.1 Flash TTS into preview, a new speech synthesis model that bets not simply on reading text aloud but on directing the voice. The main difference in this release is that a developer can set intonation, pace, accent and even a change of emotion right in the text prompt, instead of tuning the result blindly through a set of fixed parameters. For the voice AI market this is a notable shift: text-to-speech looks less and less like a black box and more like a directing tool.

The release is rolling out in preview for developers through the Gemini API and Google AI Studio, for enterprise customers through Vertex AI, and for Workspace users through Google Vids.

According to Google, Gemini 3.1 Flash TTS scored 1211 Elo on the Artificial Analysis TTS leaderboard, which is built on blind user comparisons of speech quality. The company calls it the most natural and expressive model in its TTS lineup. Google also stresses the combination of high quality and relatively low cost, meaning the model targets mass-market product use cases, not just demos.

The key feature of the update is audio tags: built-in text commands that control exactly how a phrase is spoken. A developer can describe a scene, give a character a voice profile, add director's notes on tone and pace, and then refine individual lines, or even parts of a single line, with inline tags in square brackets. In plain terms, the same phrase can sound calm, irritated, whispered or hurried without switching to a different pipeline. Google AI Studio gets configurable controls for this, and ready-made settings can be exported to Gemini API code to keep the sound consistent across projects and platforms.

The second major emphasis is global reach. Gemini 3.1 Flash TTS supports more than 70 languages and aims beyond a formal conversion of text to audio, covering local speech features: accents, dialect nuances and pacing. For product teams this matters most in localization, dubbing, voice assistants, podcasts, training videos and audiobooks.

Another notable difference is the native multi-speaker mode. The model can generate a dialogue between two speakers in a single request, without splitting the conversation into separate API calls. This should give a more natural rhythm and more coherent exchanges than the classic approach, where each voice is synthesized separately and then stitched together on the application side.

Google has also embedded SynthID watermarks in the entire generated audio stream. They should be imperceptible to the listener but allow reliable detection that a recording was created by AI. As synthetic speech gets better, this is no longer an optional extra but a basic safety element: the more convincing the voice, the more important it is to verify its origin by machine.

The model is still in preview, though, and it has limitations. In its documentation, Google states that TTS here does not support streaming, that long outputs of several minutes can lose stability and quality, and that in rare cases the service returns text tokens instead of audio, so a request may end with a 500 error. There is also a prompt caveat: if an instruction is vague, the model may reject the request or literally read aloud the director's notes meant as instructions.

The takeaway is simple: Google is trying to turn speech synthesis from a narrow API tool into part of the full multimodal Gemini platform. Gemini 3.1 Flash TTS is interesting not only because it sounds better than earlier versions, but because it gives developers a clearer, more controllable interface for working with voice. If the company quickly stabilizes long generations and keeps the balance of price and quality, it has a good chance of establishing itself not only in the infrastructure layer but also in creative voice products, where specialized TTS services have dominated until now.