Meetings & calls
Business teams
Summaries, decisions and action items on top of the transcript — shared while the meeting is still fresh.
Gemini 3.5 Transcribe word-level timestamps, explained
Word-level timestamps give every word its own start and end time — and Gemini 3.5 Transcribe returns them through its API, per Google's launch post. Developers build subtitle generators, clip search and text-based editors on top of them. If you want the timecoded transcript rather than the app, Subanana turns the same class of engine output into editable cues — a 1-hour recording done in about 5 minutes.
What per-word timing is, what Google announced, and what still sits between raw timestamps and a file you can actually use.
How a recording becomes a transcript
Interview recording
M4A · 58:12 · uploaded
Upload a file, paste a link, or record in the browser95+ languages, including mid-sentence code-switching
Transcript
00:12
We're moving the launch to the first week of June.
00:47
Fine — but the pricing page has to be final by then.
Transcript
TXT · DOCX · XLSX · Markdown
Subtitles
SRT · VTT
Translation
95+ languages
Summary
Key points · action items
Answers
Ask the transcript anything
Why teams hand their recordings to Subanana
Not a feature list — the things that decide whether a transcript is usable without listening again.
A transcript you can read, not decode
- Speakers separatedEvery line carries who said it, identified automatically.
- Punctuation and paragraphsRestored automatically, so the text reads as prose — not as one unbroken wall.
- Tidied textFiller words are cleaned away while the meaning stays untouched.
- Ask the transcriptQuestion the recording in the editor and get answers grounded in what was said.
Accuracy that is engineered, not promised
- The best model per languageModels are benchmarked continuously; each file goes to the top performer for its language.
- Hallucination detectionSuspect output is caught and re-processed by another engine before it reaches you.
- Your terms, spelled your wayPin names, products and jargon in a glossary that applies across your projects.
- Propose-and-confirm fixesAI proofreading suggests corrections for misheard words; nothing changes until you approve.
Who turns speech into text here
The flow is the same — what differs is the deliverable: a transcript, minutes, subtitles or a summary.
Videos & podcasts
Creators
One transcript becomes subtitles, show notes and quotable lines, ready for every platform you publish on.
Interviews
Journalists & researchers
Quotes must be verbatim and attributed to the right speaker — and ready well before the deadline lands.
Lectures
Students & educators
Long recordings arrive summarized and searchable, so revision starts at the point that actually matters.
What word-level timestamps actually are
A transcription engine can report timing at several grains. Segment-level timing stamps each chunk of speech — usually a sentence or phrase — with one start and end time. Word-level timing goes finer: every word carries its own start and end, typically to the millisecond. The difference sounds academic until you build with it. With segment timing you know roughly when a sentence was said; with word timing you can place a caption the instant a word lands, jump a video to the exact word a viewer searched for, or delete a filler word from the audio by deleting it from the text. Per Google's launch announcement, Gemini 3.5 Transcribe returns word-level timestamps with speaker attribution on pre-recorded audio through the Interactions API, and serves real-time use via a separate streaming mode, the Live API, with sub-second latency.
What per-word timing enables
Four things people build with word-level timestamps — each one an application of the same data.
- Subtitles and captionsCue boundaries can fall exactly where words do, so captions sync tightly to speech. The timestamps are the raw material; the cue segmentation — how many words per cue, line lengths, reading speed — is a layer you build on top.
- Search and clippingWhen every word knows its position in the audio, a search hit is a seek target. Podcast apps, lecture archives and video editors use this to jump straight to the moment a phrase was said, or to cut a clip by selecting text.
- Highlight-as-you-listenKaraoke-style read-along views — the word lights up as it is spoken — need per-word timing by definition. The same mechanism drives accessibility readers and language-learning players.
- Text-based editingEditors like the ones podcasters use let you delete a sentence from the recording by deleting it from the transcript. Word timing is what maps the text operation back onto the waveform.
Gemini 3.5 Transcribe timing specs, per Google's announcement
What the launch materials state — plus the access model any API workflow implies.
- Timing grain
- Word-level timestamps on pre-recorded audio, via the Interactions API
- Speaker attribution
- Speech attributed with timestamps for up to 3 speakers; 3+ speakers marked experimental
- Languages
- 85+ languages automatically detected, including regional accents and dialects
- Streaming mode
- Continuous bidirectional streaming with sub-second latency via the Live API
- Reported accuracy
- Average word error rate (WER) of 2.6% for pre-recorded audio, 4.0% for streaming, per Google
- How you access it
- The Gemini API, with a key from Google AI Studio — the consuming application is yours to write
The gap between timestamps and a finished deliverable
Word-level timestamps are an ingredient, not a product. A subtitle file needs cues with sensible boundaries, line lengths a viewer can read, and a reading speed that doesn't outrun the eye. A usable meeting transcript needs speaker labels attached to readable paragraphs, names spelled consistently, and a way to fix the lines the engine got wrong. A clip workflow needs an editor that understands both the text and the video. If you are building an application, that layer is your product, and an API with per-word timing is exactly the right foundation. If you just need the subtitle file or the transcript, the honest arithmetic is different: you would be building a one-off pipeline that managed tools already run at production quality.
Raw timestamps vs. a finished transcript
The API hands you the ingredient; Subanana hands you the dish. Same class of engine underneath.
| Gemini API output — you build on it | Subanana output — ready to use | |
|---|---|---|
| What you hold afterwards | Words with start and end times, via the Interactions API | Timecoded cues with speaker labels, open in an editor |
| Fixing what the engine misheard | Review tooling you build | AI proofread suggestions you approve, plus automatic re-runs of suspect segments on another engine |
| Getting a subtitle file | Cue segmentation and SRT rendering in your code | SRT or VTT in one click, reading-speed warnings included |
| Getting a document | Your own formatting layer | TXT, DOCX, XLSX or Markdown export |
| Asking questions of the recording | Retrieval you build | Chat with the transcript — answers grounded in the recording |
This compares two workflows, not two models. Subanana continuously benchmarks frontier engines of the same class and routes each file to the best one for its language — 98% average accuracy, 200K+ users.
Questions we getFrequently asked questions
Word-level timestamps are per-word timing data: the transcription engine returns a start time and end time for every individual word, rather than one timecode per sentence. They enable tightly synced captions, search that jumps to the exact moment a word was said, and editing video by editing text.
Yes — Google's launch announcement lists word-level timestamps with speaker attribution as a capability for pre-recorded audio via the Interactions API. Real-time streaming is served separately through the Live API with sub-second latency.
Segment-level timing gives one start and end time per chunk of speech, usually a sentence or phrase. Word-level timing gives a start and end for every word, typically to the millisecond — fine enough to highlight words as they are spoken or cut audio by deleting text.
You group words into cues — deciding where each cue starts and ends, how many characters fit a line, and how fast a viewer can read — then render the result in SRT timecode format. That segmentation logic is code you write and tune, or a subtitle tool's job: Subanana returns transcripts already segmented into editable cues and exports SRT or VTT in one click.
Per Google's announcement, the model attributes speech with timestamps for up to three speakers in pre-recorded audio, and support for more than three speakers is experimental.
Google reports an average word error rate (WER) of 2.6% for pre-recorded audio and 4.0% for streaming. WER is the share of words transcribed incorrectly, so lower is better; results on any engine vary with audio quality, noise and jargon.
No — subtitles are built from cues, and segment-level timing can produce good ones. Word-level timing gives finer control: cue boundaries exactly on word edges, per-word highlighting, and precise clip cutting. What matters most for subtitle quality is the segmentation and reading-speed logic applied on top of whichever timing grain you have.
Yes. Upload audio or video to Subanana — or paste a public YouTube, Instagram or Facebook link — and the transcript comes back segmented into timecoded cues with speaker labels, ready to edit and export as SRT, VTT, TXT, DOCX, XLSX or Markdown. A free account previews the first 15 minutes of each file, 3 files a month; exporting the files is a paid feature.
Questions we getFrequently asked questions
Word-level timestamps are per-word timing data: the transcription engine returns a start time and end time for every individual word, rather than one timecode per sentence. They enable tightly synced captions, search that jumps to the exact moment a word was said, and editing video by editing text.
Yes — Google's launch announcement lists word-level timestamps with speaker attribution as a capability for pre-recorded audio via the Interactions API. Real-time streaming is served separately through the Live API with sub-second latency.
Segment-level timing gives one start and end time per chunk of speech, usually a sentence or phrase. Word-level timing gives a start and end for every word, typically to the millisecond — fine enough to highlight words as they are spoken or cut audio by deleting text.
You group words into cues — deciding where each cue starts and ends, how many characters fit a line, and how fast a viewer can read — then render the result in SRT timecode format. That segmentation logic is code you write and tune, or a subtitle tool's job: Subanana returns transcripts already segmented into editable cues and exports SRT or VTT in one click.
Per Google's announcement, the model attributes speech with timestamps for up to three speakers in pre-recorded audio, and support for more than three speakers is experimental.
Google reports an average word error rate (WER) of 2.6% for pre-recorded audio and 4.0% for streaming. WER is the share of words transcribed incorrectly, so lower is better; results on any engine vary with audio quality, noise and jargon.
No — subtitles are built from cues, and segment-level timing can produce good ones. Word-level timing gives finer control: cue boundaries exactly on word edges, per-word highlighting, and precise clip cutting. What matters most for subtitle quality is the segmentation and reading-speed logic applied on top of whichever timing grain you have.
Yes. Upload audio or video to Subanana — or paste a public YouTube, Instagram or Facebook link — and the transcript comes back segmented into timecoded cues with speaker labels, ready to edit and export as SRT, VTT, TXT, DOCX, XLSX or Markdown. A free account previews the first 15 minutes of each file, 3 files a month; exporting the files is a paid feature.
Updated 2026-08-26