Google has quietly shipped two new text-to-speech models under the Gemini banner, and the timing is interesting given how quickly the TTS landscape has been moving. The models — gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts — landed on September 23, 2026, and they bring a genuinely large voice library to the table: over 2,000 options. But the detail that caught my eye is not the count. It is the ability to clone a voice from a 30-second audio sample, provided you either own that voice or have rights to use it.
What the new models actually offer
The Flash and Flash-Lite naming follows Google's existing Gemini pattern, where Flash handles higher-quality generation and Flash-Lite is the cheaper, faster alternative. In practice, the difference matters for production workloads where cost per character adds up quickly. The Flash model is the one to reach for when audio quality or expressive range matters; Flash-Lite covers cases where throughput and budget are the priority.
The custom voice feature is worth understanding separately. A 30-second sample is a short window, but it is enough for the model to extract timbre, cadence, and spectral characteristics. For developers building branded experiences — think interactive voice response systems, podcast intro readouts, or character-driven apps — this removes a previous barrier. You no longer need to contract a voice actor or navigate licensing for a pre-recorded sample.
The conversation feature is the real differentiator
What makes the Gemini TTS API stand out from competing offerings is how it handles multi-character dialogue. The API lets you define a full conversation where each speaker has their own voice and their own style instructions. That is not a novel idea in concept — text-to-speech pipelines have supported dialogue for years — but the way this is exposed through the API makes it unusually straightforward to wire up.
Simon Willison demonstrated this with a short clip featuring two pelicans debating whether to relocate to the Pacifica Pier. He used Claude 4.5 Opus to write the script and generate a render URL, then passed the output through the Gemini TTS tool. The result is a quick illustration of how multiple distinct voices can be produced in a single request without complex orchestration on the developer side.
Performance and cost in practice
Willison generated 1 minute and 18 seconds of audio using the Flash model, and the process took roughly 20 seconds. The cost came in at 2.74 cents. That is a useful benchmark for anyone evaluating whether to integrate this into a real product. For context, generating a minute of audio at under 3 cents means even high-volume applications — daily podcasts, educational content narrations, accessibility tools — remain economically feasible at scale.
The playground interface itself was built as a bring-your-own-key tool using GPT-6 Astra, taking advantage of the Gemini API's open CORS policy. That openness is not trivial. Many API providers restrict cross-origin requests, which forces developers into proxy layers or backend services just to make a client-side demo work. Google's decision to leave the CORS policy open lowers the friction for experimentation, which is exactly where tools like this playground live.
What this means for developers
The practical takeaway is that TTS is becoming faster, cheaper, and more configurable without requiring specialized infrastructure. The ability to define character-specific voices in a single API call opens up workflows that previously needed multiple passes or third-party middleware. A developer building an audiobook narrator, a game dialogue system, or a multilingual customer support tool can now prototype all of that directly through the Gemini API.
The 2,000-voice library also signals where Google is headed with this product. Volume matters because voice selection affects how well a synthesized speaker matches the tone and context of the content. A library of this size suggests significant investment in cataloguing and categorizing voices across languages, accents, and stylistic ranges.
As generative audio continues to compress on cost and latency, the differentiator shifts from "can it generate speech" to "can it generate the right speech for the context." The multi-character conversation support and voice cloning pipeline point squarely at that second question.