Google ships Gemini 3.8 Flash TTS and Flash-Lite TTS with prompt-designed voices and 30-second cloning

Prompt-written voices shift TTS from preset selection to directorial control

The headline change in the release is that voice design moves from picking a preset to writing a description. According to one outlet, a text prompt can define a voice's role, accent, and vocal traits; another reports that the models let you direct AI voices like actual actors. Coverage from a third source describes the same capability as crafting entirely new voices from natural language prompts, with the developer able to replicate them with consent. For practitioners, the practical consequence is that a single prompt — not an asset library — now produces a custom narrator, and that the same prompt can be re-written to retune the voice without swapping IDs.

Multi-language reach and a 30-second cloning path, gated by consent

The new models cover more than 100 languages and dialects, according to one report; another adds that Flash TTS creates new voices from prompts describing role, accent and voice characteristics across that language set. Cloning is positioned as low-friction but gated: one outlet says the models clone a voice from 30 seconds of audio after a consent check. A separate write-up frames the same workflow as replicate them with consent. The consent gate is the key compliance point for editors working with real talent, and the 30-second threshold is short enough that a single recorded read-aloud is sufficient reference material.

Line-level direction with 200+ expressive audio tags

Beyond voice identity, the models add explicit performance control. One report cites 200+ expressive audio tags and multi-language narration; another frames the feature as directing line-level emotion. Combined with the prompt-based design, this means a script can now carry stage directions inline: the same passage can be narrated with a different mood by changing tags, not by re-generating with a new sample. For audiobook and e-learning producers, that collapses what used to be a per-chapter re-record into a per-tag edit.

Where the models ship and what surfaces are covered

Distribution is broad inside Google's stack. One listing enumerates Google AI Studio, the Gemini API, Gemini Enterprise, Gemini Notebook, and Google Vids as launch surfaces; a third outlet describes the same rollout as shipping to AI Studio and the Gemini API for developers building voice apps. For a writing team, the relevant entry point is AI Studio or the Gemini API; for a video pipeline, Google Vids adds direct in-product narration without an export step. The 2000+ voice library and the prompt-based design are available on the same surfaces, so teams do not need a second product to switch between bespoke and stock voices.

What to verify before putting the voices into a workflow

Three checks matter before production use. First, confirm the consent flow on the cloning path — every outlet that mentions cloning also names a consent step, so the gate is built in but how it is enforced on the API is worth testing. Second, validate the language and dialect list against your deliverable locales; the published figure is more than 100, and the exact roster should be checked against the API reference rather than inferred. Third, bench the expressive tags on a representative script, since tag-driven delivery behaves differently from prompted identity and is the part most likely to need prompt-tuning.

Follow-up for practitioners

The release is live as of September 24, 2026. The next concrete steps for a writing or localisation team are: prototype one voice prompt per locale in AI Studio, attach tags to a stress-test passage, and trial the 30-second cloning path with a recorded consent line. A browser-based Text To Speech utility can serve as a quick scratchpad for comparing prompt designs before API integration; a Reading Time Calculator helps size scripts that the new models will narrate, and a Unicode Encoder / Decoder is the practical companion when scripts cross scripts that the more than 100 language set covers.

A pending release or decision date is not named in the evidence and so is not stated here.