Google ships Gemini 3.1 Flash TTS with audio tags for voice control

TL;DR

• Google released Gemini 3.1 Flash TTS on 15 April, its latest text-to-speech model, scoring an Elo of 1,211 on the Artificial Analysis TTS leaderboard
• New audio tags embedded in text let developers direct vocal style, pace and delivery without complex API calls; multi-speaker dialogue and 70+ languages supported
• All output is watermarked with SynthID — a provenance feature that becomes relevant as UK insurers, broadcasters and broadcasters-regulators face rising synthetic-audio fraud

Google has released Gemini 3.1 Flash TTS, its latest text-to-speech model, targeting developers, enterprises and Workspace users with improved expressivity, multi-speaker dialogue and granular voice control through natural-language commands. The release is rolling out in preview through the Gemini API, Google AI Studio, Vertex AI and Google Vids from 15 April.

What the audio tags enable

The headline feature is “audio tags” — inline natural-language commands that steer vocal output without separate API calls. Developers can set scene direction (environment, dialogue context), cast characters via Audio Profiles with Director’s Notes, and adjust pace, tone and accent mid-sentence. The export flow translates these configurations into Gemini API code, preserving consistent character voices across projects and platforms.

Google frames the experience as placing developers in “the director’s chair.” Early testers have highlighted controllability and expressivity as the standout benefits, according to Google. On the Artificial Analysis TTS leaderboard — which measures thousands of blind human preferences — Gemini 3.1 Flash TTS scored an Elo of 1,211 and was positioned in the benchmark’s “most attractive quadrant” for combining high-quality speech with low cost.

The 70-language footprint and SynthID

The 70+ language support matters for UK enterprises building multilingual customer-facing applications — contact centres, localised advertising, accessibility tooling. Granular style and accent control across those languages is a meaningful step beyond earlier multilingual TTS models that often produced monotone or accentless output.

Every audio output is watermarked with SynthID, Google’s imperceptible audio watermark. This sits in pointed contrast to the synthetic-audio fraud landscape: the same week, Admiral reported a 71% rise in AI-assisted insurance fraud, and regulators continue to raise concerns about voice-cloning scams targeting UK consumers. A reliable watermark on new generations of synthetic audio is a precondition for regulators to require provenance-based detection in areas like financial services and political advertising.

Looking forward

For UK AI product builders, Gemini 3.1 Flash TTS joins ElevenLabs, OpenAI’s TTS models and Azure Speech Service as the options worth testing. The audio-tag interface is the differentiator — lowering the barrier to producing controllable, characterful voice output that currently requires significant engineering effort. Expect Google Vids adoption to accelerate in Workspace tenants as video narration becomes cheaper to produce at quality. The harder test will be SynthID detection adoption by third parties; a watermark only matters if tools and regulators actually check for it.