Gemini 3.8 TTS Tops a Benchmark Its Own Hire Built

Google released Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS on Sept. 23, two text-to-speech models it says rank first on a voice-quality benchmark built by Hume AI, the startup whose founder now works on the team that shipped them.

The two models are available today to developers through the Gemini API and Google AI Studio, according to Google's announcement. Enterprise access through the Gemini Enterprise API is coming later, and consumer access is limited to Flash TTS inside Gemini Notebook and Flash-Lite TTS inside Google Vids. Google did not list a price for either model in the post.

What ships today

Flash TTS is built for what Google calls deep creative direction and character design, generating an original voice from a text prompt; Flash-Lite TTS targets high-volume, cost-efficient use such as dubbing and voice agents, the company said. Both draw on a library of more than 2,000 production-ready voices across more than 100 languages and dialects, and both can clone a voice from a 30-second audio sample with what Google calls consent verification. A separate feature lets a developer direct a script line by line, marking where a voice should pause, change pace or add a conversational sound such as a laugh or a backchannel murmur, rather than leaving the whole passage to a single prompt.

Model	Built for	Consumer surface
Gemini 3.8 Flash TTS	Creative direction, character voices	Gemini Notebook
Gemini 3.8 Flash-Lite TTS	High-volume, cost-efficient scale	Google Vids

Every output carries a SynthID watermark, which Google describes as "an imperceptible watermark woven directly into the audio output." Voice replication is unavailable in Illinois, Texas, the European Economic Area, the United Kingdom, Switzerland and India, the announcement said.

A benchmark built by the man now running the team

Google credited the launch to Leland Rechis, a group product manager, and Alan Cowen, director of research science on the Gemini Audio team, and said the two models placed first on Hume AI's Voice Design Benchmark, scoring 71.4, and first or second on its Overall Quality Index.

Cowen founded Hume AI, the company that built both benchmarks, before joining Google DeepMind as director of research science in January, according to runtimewire's reporting on the launch. His research at Hume focused on measuring and predicting emotional expression, the same territory Google is now selling the new models on.

"Those results are reported by Google," runtimewire wrote; "they do not, by themselves, settle how the products compare for a particular use case or what they cost to run." Google's announcement does not name an evaluator independent of Hume AI, or say whether Cowen's team had a role in choosing the benchmark it now tops.

The release lands in the same week Anthropic revised pricing on Claude Opus 5.5 and Google's own open-source agent runtime, ax, passed 7,600 GitHub stars. Google has not said when Gemini Enterprise API access opens or published pricing for either TTS model.