Does Gemini have a speech-to-text API?

Yes. Google's Gemini Developer API has a dedicated speech-to-text model, Gemini 3.5 Transcribe. Google describes it as a speech-to-text model with automatic language detection, speaker diarization, word-level timestamps and custom vocabulary biasing ¹. A second model, Gemini 3.5 Transcribe Live, streams audio and returns text while people are still speaking. You can also ask a general Gemini model to transcribe a file inside an ordinary prompt, which Google's audio guide describes as speech-to-text. The rest of this page covers the prices, limits and gaps you need to check before you build on it. For the wider market, see our guide to speech-to-text APIs.

Gemini 3.5 Transcribe is the file route, and it lists at $0.003 per minute

For a recording you upload, the model to use is Gemini 3.5 Transcribe. On the paid tier, Google lists its input at $2.00 per million audio tokens, and gives an estimate of $0.003 per minute of audio ². The free tier shows input as free of charge on the same price list. The per-minute figure is an estimate, not a meter reading. Google calculates it at 25 audio tokens per second of input ³, so the bill for a real file depends on its length in tokens.

Set against the cloud speech product from Google, the model is cheaper on the list rate:

The two rows are not a like-for-like contest. The Google Cloud figure is for a different product with its own features, quotas and data terms, and the Speech-to-Text pricing comparison lists the other providers. Use the table to see the order of magnitude, then check your own volume and features. For the Cloud product, see whether Google Speech-to-Text is free and how to use the Google API.

The live model streams audio for up to 10 minutes per session

For live captions or a voice application, Gemini 3.5 Transcribe Live opens a bidirectional connection over WebSockets, takes raw PCM audio and returns incremental text as speech occurs. Google's live transcription guide describes this model and its limits. The most important limit is session length: a live transcription session supports continuous streaming for up to 10 minutes ⁵.

For a meeting or a lecture longer than that, plan to split the stream into sessions in your own code, and to keep the text from each session in order. The limit is one of the first things to check, because a design that works in a five-minute test can fail in a full-length event.

The live model also has a gap that matters for meetings. Speaker diarization is not supported in live streaming sessions, and Google says that for speaker labels you should use its non-streaming endpoint ⁶. A live transcript therefore cannot say who spoke, so if you need labels, transcribe the finished recording with the file route. The guide to speaker diarization explains what the labels are and how they are scored.

A general Gemini model can transcribe a file, but it counts audio differently

Any Gemini model that accepts audio can produce a transcript when you ask for one in the prompt. Google's audio guide lists transcription and translation, speaker diarization, emotion detection and timestamps as uses, and it says you can upload larger files through the Files API. The guide's technical section gives the audio rate for these models as 32 tokens per second of audio, or 1,920 tokens per minute ⁷.

That rate is not the one Google uses for the transcription model, which bills at 25 tokens per second ³. If you estimate the cost of a transcription job from the 32 rate, your estimate will not match the price list for Gemini 3.5 Transcribe. Use the price list for the model you call.

A prompt-based route also gives you a free-form answer rather than a fixed transcript format. You can ask for a summary, a translation or a list of action items in the same request. If all you need is a plain transcript with word timings, the dedicated model is the more direct choice.

Google's own pages disagree about real-time speech-to-text

Google's documentation is not consistent on this point. The general audio guide says that for dedicated speech-to-text models with real-time transcription, you should use the Google Cloud Speech-to-Text API ⁸. The live transcription guide, read on the same day, describes a real-time transcription model inside the Gemini API.

We think the live transcription guide describes the newer product, because it names the model and its limits. We cannot prove which guidance Google intends to keep, so do not rely on either page alone. Before you build a real-time system, check the model name on the page for the platform you use, and test the streaming endpoint itself. The Gemini API and Vertex AI are separate products with separate documentation. In our search results, the live model appeared on the Gemini API page, not on the Vertex AI Live API overview, so confirm which platform offers it before you build.

What to do before you build on Gemini speech-to-text

1. Choose the route by audio type. Use the file route for recordings, and the live route for audio that is still being spoken.
2. Read the price list on the day you build. Check the per-minute estimate, the token rate it assumes and any rate change dates, because the page lists some prices that start on a later date.
3. Test on your own audio. Transcribe a set of real recordings, then measure the word error rate with the same normalization for every system you compare. Count invented sentences separately, since a clean error rate can hide them.
4. Design around the live limits. Split long live sessions, and do speaker labels on the finished file rather than on the stream.
5. Read the data-use row for your tier. The price list has a row for each model that says whether your data is used to improve Google's products. The row can differ between free and paid tiers, so check the one you will use before you send customer audio.

If the audio should never leave your computer, no cloud API meets that test. Yap, which we make, is an on-device dictation app for Mac and Windows, and offline dictation on Mac explains how that works. For a single recording you can also try free online dictation or file transcription in the browser, which need no API key.

For the broader question of how speech recognition works and what its accuracy means, see our overview of speech recognition. Related answers: whether there is a free speech-to-text API, what OpenAI's speech-to-text API costs and whether Gemini can transcribe audio.