Gemini 3.1 Flash TTS

This is a note from my public notes. View the canonical version: Gemini 3.1 Flash TTS.

Google's latest Text-to-Speech (TTS) model (April 2026), part of the Gemini family. Focused on controllability, expressivity, and quality for developers and enterprises building speech applications. Complements Gemini 3.1 Flash Live (real-time dialogue) with higher-quality, more controllable offline generation.

Key Capabilities

• Audio tags for style control: natural-language inline tags (e.g. [whispers], [laughs], [excited], [sighs], [sarcastic], [very fast]) give granular control over delivery, tone, pace, and non-verbal sounds. Tags can be combined and mixed mid-sentence; no exhaustive list, experimentation encouraged.
• Creative expressivity: supports stylistic directives like [like a cartoon dog] or [like dracula]; scene direction and speaker-level audio profiles.
• Multi-speaker dialogue: native support for multi-speaker conversations with distinct voices.
• Broad language coverage: 70+ languages. For non-English transcripts, English tags recommended.
• Quality: 1,211 Elo on the Artificial Analysis TTS leaderboard; positioned in the attractive quadrant for quality-vs-cost.
• SynthID watermarking: all generated audio is invisibly watermarked to flag AI-generated content.

Availability (April 15, 2026)

• Developers: preview via Gemini API and Google AI Studio (configurable controls with exportable API code).
• Enterprises: preview on Google Vertex AI.
• Consumers: available via Google Vids for Workspace users.

Notable Observations

• Simon Willison flagged the prompting guide as "surprising" — effective prompts can span hundreds of words, specifying accent, emotional shading, even "the grin in the audio".
• Accent control is prompt-driven; switching between UK regions (London, Newcastle, Exeter) in the same base prompt produces distinct regional deliveries.
• Willison built an interactive playground at tools.simonwillison.net for multi-speaker experimentation.

Why It Matters

Moves TTS from "read this text" to "perform this text". Unlocks voice-first agents, character-driven audio, multi-speaker narration, and podcast-style content without manual voice acting; also accelerates plausible audio impersonation (mitigated partially by SynthID).

References

• Announcement: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-flash-tts/
• Transcript tags docs: https://ai.google.dev/gemini-api/docs/speech-generation#transcript-tags
• Simon Willison's notes: https://simonwillison.net/2026/Apr/15/gemini-31-flash-tts/
• Playground: https://tools.simonwillison.net/

Related

• Gemini
• Gemini 3.1 Flash Live
• Text-to-Speech (TTS)
• Google AI Studio
• Voice Cloning
• Simon Willison

About Sébastien

Ready to get to the next level?