Google: Gemini 3.1 Flash TTS Preview

Google: Gemini 3.1 Flash TTS Preview is a text to speech model from Google. It sits in the speech category and is a preview release, which means the behaviour and voice set may still change before a stable version lands. Google describes it as a substantial generational step up from Gemini 2.5 Flash TTS, so if you have used that model, this is the newer generation to evaluate.

The model takes text as its only input and returns speech as its only output. It supports a context window of 32,768 tokens, with a maximum output of 16,384 tokens, which is generous for narration and long form spoken content. Thirty voices are available: Zephyr, Puck, Charon, Kore, Fenrir, Leda, Orus, Aoede, Callirrhoe, Autonoe, Enceladus, Iapetus, Umbriel, Algieba, Despina, Erinome, Algenib, Rasalgethi, Laomedeia, Achernar, Alnilam, Schedar, Gacrux, Pulcherrima, Achird, Zubenelgenubi, Vindemiatrix, Sadachbia, Sadaltager and Sulafat. The range of names suggests a spread of vocal characters rather than a single default voice, so you can match tone to content.

Practical uses follow directly from a text in, audio out model with a large context window and many voices. You can generate narration for videos, podcasts or audiobooks, produce spoken versions of articles and documentation, build voice prompts for assistants and IVR flows, or create audio for accessibility features. Because the input limit is measured in tokens rather than characters, long scripts can be submitted in a single request, and the 16,384 token output ceiling gives room for extended spoken passages. The voice list makes it straightforward to assign different voices to different speakers or content types.

You call the model through api14's OpenAI compatible API using the model id "google/gemini-3.1-flash-tts-preview". Because the API follows the OpenAI shape, you point your existing client at api14's base URL and send a request in the usual format, with the model id above and your text as the input. The model lists no supported parameters beyond the basics, so plan on sending the model id and the text, and select a voice from the list above where the request format allows it. If your client expects a parameters block, keep it minimal rather than assuming tuning options that are not documented.

Two things are worth weighing before you commit. First, this is a preview model, so output and voice behaviour can shift, and it is not the right choice for a pipeline that needs frozen, reproducible audio. Second, it only accepts text and only returns speech, so it cannot transcribe audio, translate speech or handle images and documents. If your task is speech recognition, multimodal input or general text generation, a different model in the catalogue will fit better.