Google is rolling out two Gemini text-to-speech models that let developers create reusable synthetic voices and direct how individual lines are spoken. Gemini 3.8 Flash TTS is aimed at expressive narration and dialogue, while Flash-Lite TTS is designed for higher-volume, lower-latency speech generation. Both are rolling out through the Gemini API and Google AI Studio, Google said on September 23.
The release separates a speaker's identity from the delivery of a particular line. A developer can describe a voice's accent and vocal character in ordinary language, receive a stored voice ID, and reuse it in later speech requests. Google's documentation says this voice-design feature works with both models. It lists support for 130 languages in Flash TTS and 101 in Flash-Lite TTS.
Google says Flash TTS is suited to demanding narration and dialogue; Flash-Lite prioritizes throughput and cost for uses such as read-aloud features and voice agents. Access through the Gemini Enterprise API is planned but not yet available.
Developers must update scripts and audio handling
For a multi-speaker scene, the developer writes the dialogue and attaches a speaker name and delivery style to each turn. A style might call for a whisper or an excited response. Short tags such as <laugh> can mark a momentary sound within the spoken transcript; sustained directions and speaker labels belong in structured speech_metadata.
The distinction matters to existing Gemini speech applications. The new models treat ordinary input text as words to be spoken: an old instruction such as “Say cheerfully: Hello!” might come out in the audio. Google's migration guide also says a single-request response now defaults to a WAV file rather than the raw audio bytes returned by earlier models, so software that adds its own WAV header must change.
Voice replication requires consent
Developers can also replicate a voice from a 30-second sample, but Google says the owner must provide a matching verbal consent recording before a replicated voice is created. The company says generated audio carries a SynthID watermark. Voice remixing, which would let users alter an existing library voice, is listed as coming soon rather than part of the current rollout.
Sources
- Google: Gemini 3.8 text-to-speech launch, September 23, 2026
- Google AI for Developers: Gemini 3.8 Flash TTS model and migration guide
- Google AI for Developers: voice design documentation
- Illustrative image: microphone photograph by Dennis Hill, CC BY 2.0. The image does not depict a Google product.
