Google drops Gemini 3.1 Flash TTS with audio director tags

Getting an AI voice to sound genuinely dramatic has always been an uphill battle. Most text-to-speech engines can nail basic narration, but the moment you need a character to whisper a secret or sound out of breath mid-sentence, the illusion completely shatters.

Google is tackling this with Gemini 3.1 Flash TTS, its newest speech model designed specifically for high expressivity and granular control. Instead of relying purely on fixed voice profiles, the model introduces natural language audio tags that let you direct the performance right inside your prompt.

Starting today, the model is available in preview across Google AI Studio and the Gemini API for developers, on Vertex AI for enterprise teams, and inside Google Vids for Workspace users.

Sitting in the director chair

The marquee feature here is inline audio tagging. Developers can now drop plain-English instructions directly into a script to dictate vocal cadence, delivery, and emotional shifts, even changing tone halfway through a sentence.

Google has built out the developer experience in Google AI Studio to lean into this filmmaking metaphor. You can set the scene by defining environmental context, which keeps multiple characters grounded in their roles across extended multi-turn exchanges.

You can also assign unique audio profiles to individual speakers, add director notes to establish accents and overall pacing, and insert tags for moment-to-moment reactions. Once the scene feels right, Google AI Studio lets you export the entire configuration as Gemini API code to keep voices consistent across production environments.

On the benchmarking front, Gemini 3.1 Flash TTS scored an Elo of 1,211 on the Artificial Analysis TTS leaderboard, landing in what the firm calls the most attractive quadrant for pairing top-tier quality with low latency and cost. It supports native multi-speaker dialogue across more than 70 languages, and every clip comes embedded with imperceptible SynthID watermarking to verify synthetic origin.

Why it matters

If you build narrative games, interactive audio agents, or automated video workflows, this model solves a massive production bottleneck. Splicing multiple audio takes just to give a virtual character a hint of sarcasm is a nightmare; doing it with a single inline prompt changes the math on audio production speed.

For the broader voice industry, this is a clear shot across the bow of specialized startups like ElevenLabs. Big tech TTS used to feel sterile, leaving expressive, character-driven audio to niche platforms. By bringing dynamic emotion control, multi-speaker support, and 70-plus languages to a Flash-tier cost model, Google is making professional-grade voice generation a commodity API feature.

We think the move to make voice directing fully exportable into production-ready API calls is brilliant product design. The real test, however, will be consistency: natural language prompts can sometimes lead to unpredictable vocal performances, especially across lesser-resourced languages. Furthermore, as synthetic voices become this persuasive and easy to steer, the real-world pressure on SynthID detection will be higher than ever.