Skip to main content
Gemini TTS is Google’s speech generation family, and the broadest-reaching voice engine in the Ollang catalog. Two things distinguish it: 99 supported languages, more than any other engine here, and multi-speaker synthesis — two distinct speaker voices rendered into a single generation, each preserving its own tone, pitch, and style across the exchange. Delivery is controlled with natural-language prompts covering style, accent, pace, tone, and emotion, rather than with parameter sets.
Available in Ollang Workflows as geminiElevenlabs — AI Dubbing. See the Text-to-Speech catalog.

​
Model versions

Google documents three Gemini TTS models on the Gemini API, all currently in preview: On Google Cloud Text-to-Speech the same family is published as gemini-3.1-flash-tts-preview, gemini-2.5-pro-tts, gemini-2.5-flash-tts, and a single-speaker-only gemini-2.5-flash-lite-preview-tts.

​
Coverage and limits

​
Key capabilities

  • Multi-speaker in one pass — a two-hander is generated as a conversation rather than as separate lines stitched together, which is what keeps turn-taking and interruption sounding natural.
  • Prompt-directed delivery — style, accent, pace, tone, and emotion are written in natural language, closer to how a voice director works than to how a TTS parameter set works.
  • Widest language reach in the catalog at 99 languages, covering many locales that Cartesia, ElevenLabs, and Deepgram do not serve at all.
  • Code-switching within sentences, with intonation appropriate to each language.
  • Streaming on the 3.1 Flash version, for pipelines that consume audio before generation completes.

​
Where it fits in a workflow

Two situations make Gemini TTS the right call:
  1. Language reach. If a program has to ship in locales that the 29-language and 44-language engines do not cover, this is the option that makes it possible without falling back to a second provider.
  2. Performed dialogue. Scripted two-character scenes, dramatized narrative, and immersive content where the exchange between speakers is the point.
For single-speaker informational content in major languages, Cartesia or OpenAI TTS will serve just as well with less setup.

​
Trade-offs

  • All TTS versions are in preview, so behavior may change between releases. Re-benchmark after a version change rather than assuming parity.
  • Two speakers maximum per generation, and a 32k-token session context — long multi-character scenes have to be segmented.
  • Expressiveness depends on prompting. Undirected output sits much closer to the field than the ceiling suggests, so budget effort for direction.

​
Reference