Google launches Gemini 3.8 Flash TTS models with custom voice design and 100+ language support

Gemini 3.8 Flash TTS and Flash-Lite TTS roll out across Google's developer and enterprise surfaces

Google introduced two new text-to-speech models, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, on September 23, positioning Flash TTS for creative voice design and Flash-Lite for high-volume audio generation. The launch extends the Gemini 3.8 family with tools that let users build original voices through natural-language prompts, direct delivery line by line, and stage two-speaker dialogues from a single script. Flash TTS also adds the ability to replicate a voice from a 30-second audio sample, subject to a verbal consent check by the voice owner.

Voice creation, cloning guardrails and language breadth

Google says Flash TTS lets developers describe a voice in plain language, with examples including a high-energy Melbourne DJ and a Japanese dragon, then refine the result through additional instructions. Custom voices can be saved for reuse across projects, and Google plans a voice remixing feature that adjusts timbre, pitch, pace and accent from a library voice. The models support more than 100 languages and dialects, including regional varieties such as Mexican Spanish, Quebec French and Scots English, and ship with what Google describes as more than 2,000 production-ready voices. Voice replication requires a matching verbal consent recording from the voice owner, and Google says every generated Gemini Audio clip carries a SynthID watermark, with replicated voices additionally carrying C2PA content credentials. Replicated voices are unavailable in Illinois, Texas, the European Economic Area, the United Kingdom, Switzerland and India.

Benchmark claims and how Google framed them

Google reports that Flash TTS took the top overall position on Hume AI's Voice Design Benchmark with a score of 71.4 and led the benchmark's accent-modeling category at 60.8. Google also says Flash TTS and Flash-Lite ranked first and second respectively on Hume AI's Overall Quality Index. A separate source reports Flash TTS scored 89.5% on the Artificial Analysis Pronunciation Robustness Benchmark, above the 88.2% recorded for Gemini 3.1 Flash TTS. On Voice Arena blind human preference tests, Google says both models placed near the top among competitors in Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi. Those are Google's characterizations of third-party benchmark results, not independently verified findings, and the announcement does not disclose pricing for either model.

Distribution and early partners

Google is distributing the new TTS models through existing products rather than a separate voice service. Flash TTS is available through the Gemini API and Google AI Studio for developers, with planned rollout to Gemini Enterprise via API and to Gemini Notebook for general users. Flash-Lite is targeted at high-volume use cases such as dubbing, audio production and voice agents, with developer access through Google AI Studio and broader creative use through Google Vids. Named integration partners include Figma, HeyGen, Wondercraft, Linguana, 99.co, Ollang, Agora, LiveKit, Pipecat and Vercel, putting Google's new voice models in direct contact with established voice and dubbing specialists including ElevenLabs and Murf AI, and with Adobe's Firefly speech generation, which Adobe added in August.

Other developments

Separately, former Hume AI founder Alan Cowen, now a Google DeepMind Director of Research Science, is credited alongside Leland Rechis on the launch, with his prior work on measuring and modeling emotional behavior cited as a throughline for the announcement's emphasis on character design and performance direction. Google did not publish pricing or a cost comparison between Flash TTS and Flash-Lite, leaving developers unable to judge the economic difference between the two models from the launch materials alone. Availability is rolling out from September 23, with enterprise and notebook access to follow.

What remains uncertain

Google has not disclosed pricing, latency, or general availability versus preview status for the two models, and has not detailed procedures for handling disputed consent or rights claims after a voice has been created. The benchmark scores cited by Google come from Hume AI and Artificial Analysis evaluations as reported by the company, and have not been independently verified.