Google: Gemini 3.5 Transcribe

Google: Gemini 3.5 Transcribe is a speech-to-text model from Google. It sits in the transcription category, and its provider release date is 25 September 2026. If you are building anything that turns recorded or streamed speech into written text, this is the model family Google positions for that job, and it is available on api14 under the model id google/gemini-3.5-transcribe.

The model takes audio as its only input modality and returns transcription as its output. It has a context length of 98,304 tokens and a maximum output of 32,768 tokens. Google describes it as suited for synchronous transcription that needs word-level timestamps or speaker diarization, with support for up to eight speakers. Those two details shape what you can expect from the output: not just a flat block of text, but timing information aligned to words and labels that separate voices when more than one person is talking.

That combination points to a fairly specific set of use cases. Meeting and call transcripts benefit from diarization, since you can attribute each line to a participant rather than guessing who spoke. Word-level timestamps are useful when you need to align text to a media timeline, for example generating captions or subtitles, building searchable indexes that jump to the exact moment a phrase was said, or highlighting the spoken portion of a recording during playback. Synchronous transcription suggests interactive scenarios too, where text is produced while audio is still arriving, such as live captioning or note taking during a session. The eight speaker limit fits panel discussions, interviews and standups more than large conference rooms with many rotating voices.

Calling it through api14 works the same way as any other model on the platform, because the API is OpenAI-compatible. You send your request to the chat completions endpoint with the model id google/gemini-3.5-transcribe and your audio as the input. One thing to check before you start: the supported parameters list for this model is empty, so you should not assume that sampling controls such as temperature, top_p or max_tokens will be accepted or will change behaviour. Build your integration around the defaults and treat the audio payload and the model id as the parts you control.

The main constraints to keep in mind are the input and output types and the size limits. Only audio goes in, and only transcription comes out, so this is not the model to reach for if you need reasoning over the transcript, summarisation, translation or any other text generation in the same call. For that you would pass the transcript to a general purpose text model. The 98,304 token context and 32,768 token output ceiling also mean very long recordings may need to be split into chunks, and if your audio has more than eight distinct speakers, diarization may not cover everyone.