gemini-3.5-transcribe — Speech Recognition AI Model
Gemini 3.5 Transcribe turns up to an hour of speech into text, detecting the language on its own and following speakers who switch languages mid-sentence. It can label who is speaking, tidy fillers away, or favour your own terms. Runs on Google's side against your API key.
What it does
Gemini 3.5 Transcribe turns a recording of up to an hour into text. It detects the language itself across 85+ locales and follows a speaker who switches language mid-sentence; a language hint helps when you already know it.
Verbatim keeps every filler and false start, Smart cleans them up and formats lists, numbers and dates, and Verbatim + Speakers labels each turn by speaker. Custom vocabulary steers it toward names and jargon.
When to use it
- Subtitles and transcripts for interviews, meetings and podcasts
- Speaker-labelled transcripts of a conversation with up to eight voices
- Clean, formatted notes from a voice memo using Smart mode
- Getting product names and acronyms right with custom vocabulary
Details
- CategorySpeech Recognition
- Year2026
- LicenseGoogle Cloud Terms
Compliance & Provenance
- ProviderGoogle · GPAI
- EU AI Act RiskMinimal Risk
- Art. 50 TransparencyNot applicable
Inputs & Outputs
- SoundInputsound
Speech recording to transcribe (up to one hour)
- TextOutputstring
Transcript
Parameters
- modelstringdefault Gemini 3.5 Transcribeoptions: Gemini 3.5 Transcribe
Model variant
'Gemini 3.5 Transcribe' is the only tier.
- transcribe_modestringdefault Verbatimoptions: Verbatim, Smart, Verbatim + Speakers
Transcription mode
'Verbatim' (default) — exact words, fillers and false starts kept. 'Smart' — fillers removed, self-corrections resolved, lists/numbers/dates formatted. 'Verbatim + Speakers' — verbatim with one line per speaker turn ('Speaker 1: ...'); cannot be combined with custom vocabulary.
- languagestringdefault Autooptions: Auto, Korean, English, Japanese, Chinese (Mandarin), Spanish, +6 more
Spoken language hint (Auto detects it)
Optional language hint as a BCP-47 code. 'Auto' (default) detects the language, including switches mid-sentence. Set one when the language is known to improve accuracy.
- custom_vocabularystringdefault advanced
Custom vocabulary (comma-separated)
Optional comma- or newline-separated terms to bias recognition toward (names, product terms, acronyms), up to 1,000; best under 100. Leave blank when using 'Verbatim + Speakers' — the API refuses the combination.
Limitations
- Up to one hour of audio per run, or 30 minutes with speaker labels.
- Custom vocabulary cannot be combined with speaker labels; the run is refused when both are set.
- Runs on your own Google AI API key. Without one configured in the workspace the node cannot run.
Tags
Alternatives in Audio Understanding
- Speech Recognition
Recognize speech in a video or audio file and convert it to text.
- Speech Recognition (Clips)
Recognize speech in each clip of a video list and convert it to timed transcripts.