Gemini 3.8 text-to-speech says hello
Short summary
Google launched two new text-to-speech models that let developers create custom voices from scratch and control delivery line-by-line, with the premium version ranking first on voice customization benchmarks. The models include safeguards like consent verification for voice replication and SynthID watermarking to detect AI-generated speech.
- Gemini 3.8 Flash TTS lets you generate entirely new voices using natural language prompts, customize 2,000+ existing voices, or replicate a voice from a 30-second sample with built-in consent checks.
- Both models rank #1 and #2 on Hume AI's Overall Quality Index and support 100+ languages with strong performance in Japanese, Brazilian Portuguese, Vietnamese, Arabic, Spanish, and Hindi.
- You can direct performance line-by-line with stage directions, maintain voice consistency across hours of audio, stage two-speaker conversations, and add realistic conversational sounds like laughs and sighs.
- Every generated audio is watermarked with SynthID to remain detectable as AI-generated, and the models are rolling out today to developers via Gemini API and Google AI Studio.
Full summary
Google just released two new text-to-speech models for Gemini that let you generate custom voices instead of picking from preset options. The Flash version is built for creative work—you can design entirely new character voices from scratch using natural language descriptions, then direct their performance line by line with control over acting cues, pacing, and accents. The Flash-Lite version prioritizes speed and cost for high-volume jobs like dubbing and voice agents. You get access to over 2,000 production-ready voices across 100+ languages, plus the ability to create custom voices from a 30-second audio sample if you have consent. Both models support long-form generation without voice drift, two-speaker conversations from a single script, and realistic conversational touches like laughs and sighs.
The technical specs matter here. On Hume AI's benchmarks, the Flash model scored 71.4 for voice customization and 60.8 for accent modeling—both top positions. In blind preference tests on Voice Arena, both models ranked at the top for languages including Japanese, Brazilian Portuguese, Vietnamese, Arabic, Mexican Spanish, and Hindi. The improvements over the previous 3.1 version are substantial, especially for long-form content and dual-speaker screenplay control. Google built in safeguards too: voice replication requires verbal consent from the person whose voice you're using, and every generated audio gets a SynthID watermark that makes AI-generated speech detectable.
Developers can start experimenting today through Google AI Studio, which works like a voice design workspace where you can prompt new voices and edit scripts with line-by-line control. The Flash model is available now in the Gemini API and Google AI Studio for developers, rolling out to Gemini Notebook for everyone, with enterprise access coming soon. Flash-Lite launches the same way but also goes into Google Vids. Companies like Figma, HeyGen, and Wondercraft are already integrating these models to handle dubbing, regional localization, and conversational AI at scale.