Gemini 3.5 Transcribe: The Essentials

Last updated: September 15, 2026

Covers Gemini 3.5 Transcribe

Gemini 3.5 Transcribe turns a spoken audio clip into an accurate written transcript. It auto-detects language across more than 85 languages including mid-clip code-switching, can label who said what with speaker diarization, can emit per-word timestamps for subtitle work, and can be steered toward names, acronyms, and jargon with a custom vocabulary list. The output is plain text, ready to paste into notes, a caption track, or a searchable archive.

How Gemini 3.5 Transcribe Works

Point the model at an audio file and it returns a clean transcript of everything spoken in it, no diarization or timestamps by default. Clips up to 60 minutes are supported at the base settings.

Source audio: a short first-person radio-style clip from an orbital station.
This is Commander Reyes aboard Orbital Station Four. We have a clear view of the storm system forming over the Pacific, and it is, without exaggeration, the most extraordinary thing I have ever seen from up here.

Plain transcript, no diarization or timestamps requested. · Open on Scenario

Speaker Diarization

Turning on diarization splits the transcript into one segment per speaker turn, each labeled speaker_0, speaker_1, and so on, with a timestamp range per segment. Useful for interviews, support calls, and multi-person meetings where who-said-what matters as much as what was said.

Source audio: a two-person customer support call.
1
00:00:00,100 --> 00:00:09,500
speaker_0: Hi, thanks for calling in today. I really appreciate it. So, before we get started, could you walk me through what happened when you first noticed the issue with the account? Take your time. There's no rush at all.

2
00:00:10,100 --> 00:00:23,500
speaker_1: Yeah, so it started, I think like two weeks ago. I logged in one morning and the dashboard was just completely blank. No data at all. And I figured it was just a loading glitch, but then it kept happening every single day after that.

Diarization on, timed segments labeled by speaker. · Open on Scenario

Turning on diarization halves the maximum supported audio length, from 60 minutes down to 30.

Word-Level Timestamps

wordTimestamps emits timed segments (start and end offsets) instead of one flat block of text, in a subtitle-ready numbered format. This is what to reach for when the transcript needs to drive a caption track or a searchable, seekable index into the audio.

Source audio: a team celebrating a product launch.
1
00:00:00,100 --> 00:00:20,100
I cannot believe we actually pulled this off. Six months of late nights, three failed prototypes, and a budget that nearly got cut twice, and we still shipped it. I am so incredibly proud of this team. Every single one of you fought for this product when it would have been so much easier to give up.

Word/segment timestamps on, subtitle-style numbered output. · Open on Scenario

Turning on word timestamps halves the maximum supported audio length, from 60 minutes down to 30, same as diarization. Turning both on at once still caps at 30 minutes, not further.

Custom Vocabulary

customVocabulary takes a list of domain terms, acronyms, and proper names (up to 1000) and biases recognition toward them, so a product name or technical term that a general-purpose model would mishear or approximate gets transcribed correctly. Keep the list to genuinely distinctive terms; loading it with everyday words dilutes the effect.

Custom vocabulary: ["Nexalume", "QRX-7"]

Source audio: a product briefing on a battery module.
Good morning, everyone. Today I want to walk you through the new Nexalume battery module and its diagnostic interface, the QRX-7 controller. We've run over 300 stress cycles, and the thermal readings stayed remarkably stable throughout each session. I'm genuinely excited about where this takes the product line next quarter.

Invented product names transcribed correctly via custom vocabulary biasing. · Open on Scenario

Custom vocabulary: ["Krebs cycle", "citric acid cycle", "acetyl coenzyme A", "oxaloacetate", "citrate", "nicotinamide adenine dinucleotide", "flavin adenine dinucleotide", "mitochondrial matrix"]

Source audio: a biochemistry lecture.
Today's lecture covers the Krebs cycle, also known as the citric acid cycle, and its role in cellular respiration within the mitochondrial matrix. We'll examine how acetyl coenzyme A combines with oxaloacetate to form citrate, and trace the subsequent oxidation steps that regenerate nicotinamide adenine dinucleotide and flavin adenine dinucleotide for the electron transport chain.

A dense list of scientific terms, all transcribed correctly. · Open on Scenario

Multilingual and Code-Switching

Leave languageCodes empty and Gemini 3.5 Transcribe auto-detects the spoken language across more than 85 options, including a clip that switches languages mid-sentence. Pass BCP-47 codes (like en-US or es-ES) only when a hint is needed, for example to disambiguate between close dialects.

Source audio: a talk on sustainable urban design, in Spanish.
Buenos días a todos. Hoy quiero hablarles sobre la importancia de la sostenibilidad en el diseño urbano. Las ciudades que priorizan espacios verdes y transporte público no solo reducen la contaminación, sino que también mejoran significativamente la calidad de vida de sus habitantes.

Spanish, auto-detected with no language hint. · Open on Scenario

Source audio: a webinar introduction, in German.
Guten Tag und herzlich willkommen zu unserem Webinar. Heute sprechen wir über die neuesten Entwicklungen im Bereich der erneuerbaren Energien und wie Unternehmen davon profitieren können.

German, auto-detected with no language hint. · Open on Scenario

Source audio: a personal story that switches between English and Spanish mid-sentence, hinted with ["en-US", "es-ES"].
So my abuela always said, "Mija, if you want to learn a language, you just have to jump in and start hablando, no matter how many mistakes you make." Honestly, that advice changed everything for me. I stopped being afraid of sounding tonta and just started talking to people.

Code-switching between English and Spanish within one clip, transcribed correctly in both languages. · Open on Scenario

Parameters

audio

Required. The audio to transcribe. Up to 60 minutes at base settings, or 30 minutes if diarization or word timestamps are turned on.

languageCodes

Optional array of BCP-47 language hints, such as en-US or fr-FR. Leave empty to auto-detect across 85+ languages, including code-switching within one clip. See "Multilingual and Code-Switching" above.

customVocabulary

Optional array of domain terms, acronyms, and proper names (up to 1000) to bias recognition toward. Keep the list to distinctive terms; everyday words dilute the biasing effect. See "Custom Vocabulary" above.

wordTimestamps

Boolean, default false. Emits timed, numbered segments with start and end offsets instead of one flat transcript. Halves the maximum audio length to 30 minutes. See "Word-Level Timestamps" above.

diarization

Boolean, default false. Attributes each segment to a distinct speaker (speaker_0, speaker_1, and so on), splitting the transcript at each speaker turn. Halves the maximum audio length to 30 minutes. See "Speaker Diarization" above.

Use Cases

  • Video captioning and subtitles: run wordTimestamps on a video's audio track to get a subtitle-ready, timed transcript.

  • Meeting and interview notes: turn on diarization so notes clearly show who said what, without manual re-listening.

  • Podcast show notes and search: transcribe an episode to make it searchable and to draft show notes or pull quotes from.

  • Technical and medical dictation: load customVocabulary with product names, drug names, or jargon so specialized terms come through correctly.

  • Multilingual content archives: transcribe clips in different languages, or clips that switch languages mid-conversation, without manually tagging the language per file.

  • Customer support QA: diarized, timestamped transcripts of support calls make it easy to review specific moments in a conversation.

Tips for Better Results

  1. Only turn on diarization or word timestamps when the workflow actually needs them. Both halve the maximum clip length from 60 to 30 minutes, and a plain transcript is faster to scan for straightforward note-taking.

  2. Load custom vocabulary with the specific terms that matter, not a general glossary. A short, targeted list of names and jargon biases recognition more effectively than a long list of common words.

  3. Leave languageCodes empty for mixed-language or unknown-language audio. Auto-detection handles code-switching within a single clip; a language hint is only useful to disambiguate close dialects.

  4. Split audio longer than the cap before submitting. A 60-minute base clip or a 30-minute diarized/timestamped clip is the hard ceiling; longer sources need to be chunked first.

  5. Use diarization for any recording with more than one speaker where attribution matters. Interviews, support calls, and panel discussions all benefit from per-speaker segments over a flat block of text.

Known Limitations

  • Diarization and word timestamps each cap audio at 30 minutes. The 60-minute ceiling only applies when both are off.

  • Output is plain text, not a file format like SRT or VTT. The word-timestamp output is a numbered, timed block that reads like SRT but is not exported as a ready-to-use subtitle file; reformat it before feeding it to a video editor's subtitle track.

  • Speaker labels are generic. Diarization returns speaker_0, speaker_1, and so on, not real names; map labels to names manually if needed.

  • No built-in translation. The model transcribes in the spoken language; it does not translate the output into a different language.