A new layer of control for synthetic speech
Google introduced Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS on September 23, 2026. The company says developers can describe a character voice using natural language, including an accent, role and vocal qualities. It also describes controls for dialogue scenes with multiple speakers and natural turn-taking.
Google says the models support more than 100 languages and dialects and are available across AI Studio, the Gemini API, Gemini Enterprise, Notebook and Google Vids. Availability and specific features may vary by product and account.
Why it matters
Directing a voice with a prompt could shorten the iteration cycle for narrated media, accessibility tools and conversational software. For teams that previously selected from a fixed catalog of voices, the more interesting change is creative control rather than speech generation alone.
Google positions Flash TTS for detailed character design and line-by-line direction, while Flash-Lite TTS is aimed at higher-volume production. That distinction matters for a publisher choosing between painstaking creative control and scalable dubbing or narration. The announcement also describes long-form generation and two-speaker staging, both especially relevant to podcasts and audio drama.
Voice replication deserves extra scrutiny. Google says a person must provide a matching verbal consent recording before their voice can be replicated. It also says generated audio carries SynthID watermarking and that the system uses C2PA credentials. These are safeguards described by the provider, not a guarantee against every misuse. The company notes geographic restrictions for voice replication in AI Studio.
Availability is uneven. Google says Flash TTS is rolling out through the Gemini API and AI Studio, with Gemini Notebook access for users; Flash-Lite TTS is also rolling out to developers and is listed for Google Vids. Enterprise API availability is described as coming soon. Product names on a launch page therefore should not be read as universal access on day one.
The examples and quality claims come from Google's own announcement. Google cites voice-design and preference evaluations, but a team buying the technology should run its own tests with the scripts, accents and output requirements that matter to its audience. Latency, consistency, consent flows, commercial terms and performance in less represented languages all deserve scrutiny. NEURA has not independently benchmarked these models.
The practical question
Better controls make compelling demos; reliable output in real workflows is the actual test. We will update this story if independent benchmarks or important availability details emerge.




