Can Gemini transcribe audio?

Yes. Gemini can turn speech in an audio file into text, in two ways. In an ordinary prompt, a general Gemini model can transcribe the file when you ask for a transcript. For developers, Google's Gemini API also has Gemini 3.5 Transcribe, a dedicated speech-to-text model, and a live model for streaming audio. The chat route is quick for one file. The model route suits code and long recordings. In both, the output can be a summary rather than a copy, so ask for the speech word for word.

For the other two routes, read the guide to transcribing with Whisper, ChatGPT or Gemini. The question on Gemini's speech-to-text API goes deeper on the developer side.

Gemini can describe, summarize or transcribe audio in a prompt

Google's audio guide lists describing, summarizing and answering questions about audio content as uses of Gemini models, alongside transcription ¹. The same audio input can therefore produce a summary when you wanted a transcript. The prompt decides the output. Ask for a verbatim copy, and check the reply against the recording.

For a finished recording, use Gemini 3.5 Transcribe in the API

Google describes Gemini 3.5 Transcribe as its speech-to-text model, with automatic language detection, speaker diarization, word-level timestamps and custom vocabulary biasing ². This is the model to use when you want a transcript with timings and speaker labels.

The price is listed per million input audio tokens. On the paid tier it is $2.00 per million, which Google estimates as $0.003 per minute of audio ³. The per-minute figure is an estimate, so check the price list on the day you build.

A prompt-based route has a different limit. The Gemini API documentation gives a maximum of 9.5 hours of audio per prompt ⁴. That figure is for the developer API. The Gemini app sets its own limits, which vary by plan, so do not assume the API number applies to the app. Check Google's help page for the app before you upload a long recording.

The live model streams audio in sessions of up to 10 minutes

For captions or a voice application that transcribes while people speak, Gemini 3.5 Transcribe Live streams audio over a live connection. A live transcription session supports continuous streaming for up to 10 minutes ⁵. For a longer event, your code has to start a new session and keep the text in order.

The live model also cannot label speakers. Speaker diarization is not supported in live streaming sessions, and Google points to its non-streaming audio transcription endpoint for speaker labels ⁶. If you need to know who said what, transcribe the finished recording with the file route. The guide to speaker diarization explains how those labels work.

A chat reply can summarize, so ask for the speech word for word

A general Gemini model is designed to write useful text, not to copy speech exactly. Use a prompt that asks for a copy and forbids edits:

Transcribe this audio word for word, in the language spoken. Do not summarize, translate, shorten, reorder or correct grammar. Mark unclear words as [inaudible]. Output only the transcript.

For the difference between an exact copy and a tidier version, see verbatim vs clean verbatim. Name the filler words you want removed, if you choose the cleaner version.

What to do before you rely on a Gemini transcript

1. Choose the route. Use a prompt for one file in the app, and the Gemini API for many files or a timed transcript with speaker labels.
2. Check the limits on the day. Read the app's help page for the app and the price list for the API, because both can change.
3. Ask for a copy. Use the word-for-word prompt above, and split long recordings.
4. Check three passages. Compare one passage near the start, one in the middle and one near the end with the recording.
5. Check names, numbers and the end. A transcript that stops early or ends with a neat conclusion is probably a summary.

If the audio must never leave your computer, no cloud model meets that test. Whisper runs on your own machine, and how to use Whisper to transcribe audio shows the steps. For one recording and no API key, audio to text transcribes a file in your browser.

Related questions

• Can ChatGPT transcribe audio files?
• Can NotebookLM transcribe audio?
• Does Gemini have a speech-to-text API?
• How to transcribe audio to text, in the transcription overview.