Gemini 3.5 Transcribe: Meeting Transcripts, Access, and Alternatives (2026)
Google introduced Gemini 3.5 Transcribe on August 26, 2026. It is a speech-to-text model that turns raw audio into formatted text, with speaker attribution and word-level timestamps on pre-recorded files. This guide covers what shipped, where you can use it, and what still has to happen before a transcript becomes something you can send.
What Google announced on August 26
The announcement is short and specific. Google calls it "our most precise speech-to-text model yet, designed for intelligent voice interactions."
The framing is a contrast with older speech recognition. According to Google's announcement, conventional models "struggle with background noise, complex jargon, and disfluency cleanup." This one, Google says, "converts raw audio directly into accurate, polished, formatted text."
Two accuracy figures are published. As measured by Artificial Analysis, the model reaches an average Word Error Rate of 4.0% for streaming and 2.6% for non-streaming use cases.
Google also compares it to Chirp 3, the model it used before. Time to final output "improves by 70%," and on the FLEURS benchmark it records 5.50% WER streaming and 5.04% non-streaming.
Those numbers matter less than the format of the result. A cleaner raw text file means less editing before anyone can use it.
The two ways the model is offered
There are two separate APIs, and the split matters if meetings are your use case.
| Mode | Model ID | What it is built for |
|---|---|---|
| Real-time streaming | gemini-3.5-transcribe-live |
Continuous bidirectional streaming with sub-second latency, via the Live API |
| Pre-recorded audio | gemini-3.5-transcribe |
Recorded audio, meetings, and call logs, via the Interactions API |
The pre-recorded path is the one that carries meeting features. Google describes it as transcribing "recorded audio, meetings, call logs, and more with speaker attribution and word-level timestamps."
Speaker support has a stated ceiling. The model "accurately attributes speech in pre-recorded audio with timestamps for up to three speakers," and support beyond three is described as experimental.
What smart transcription actually changes
Four capabilities are listed, and they are the practical difference from a plain transcript.
Disfluency cleanup. The model handles self-corrections such as "let's meet Tuesday—no, Wednesday," removes filler words, and auto-formats the output.
Custom vocabulary. You can supply your own terms so specialised jargon and unusual spellings survive the process.
Alphanumeric accuracy. Google singles out "alphanumeric entities like postal codes and order IDs," which is where generic models tend to fail.
Language coverage. The model automatically detects over 85 languages, including regional accents and dialects.
There is also a function-calling capability worth noting. Google says the model "can delegate complex tasks (such as image generation and file analysis) to other Gemini models via function calls." That capability is currently listed for the Gemini macOS app only.
Where you can use it today
This is not an API-only release, which is unusual for a model launch.
| Surface | What it does there |
|---|---|
| Gemini API in Google AI Studio | Build mode, including coding apps by voice |
| Gemini Enterprise Agent Platform | Same model, enterprise deployment path |
| Gboard on Android | The Rambler feature turns speech into formatted text |
| Gemini app on macOS | Voice commands paired with screen context |
| Google Antigravity | Transcription that uses screen context and chat history |
| Chrome | Described as coming soon, for typing by voice in any field |
The macOS description is the most agent-like of the set. Google says the model makes it effortless "to summarize local files, repurpose text across apps, or generate images right at your cursor." All of that runs from voice alone.
Several developer platforms are named as building on the Live API, including LangChain, LiveKit, Pipecat, and Vercel.
One access note is easy to miss. The two API paths have separate model IDs and separate entry points. A live captioning build and a meeting-archive build are therefore two integrations, not one.
What the announcement does not cover
This is where a transcript stops being enough, and the gap is worth stating plainly rather than guessing at.
The announcement page describes text output. It does not describe producing a spreadsheet, a chart, a deck, or a written report from the audio. The words spreadsheet, Excel, CSV, slide, deck, report, and dashboard appear nowhere on it.
What it does describe is delegation: the model hands file analysis and image generation to other Gemini models. So the deliverable is assembled somewhere else, by something else.
That is a reasonable design. It is also the reason a good transcript does not close the loop on a meeting.
Turning a transcript into something you can send
A transcript records what was said. A meeting record usually needs three more things: the numbers that were on screen, the decisions, and who owns what next.
Powerdrill Bloom sits on that second half. You upload the audio and the file the meeting was actually about, then describe in natural language what you want out of it.
The speech to text page lists audio uploads in .mp3, .mp4, .m4a, .webm, and several other formats. It says the feature "gets transcripts, summaries, and main points from your audio in seconds."
Being fair about the boundary in the other direction: that page does not describe speaker attribution, word-level timestamps, or real-time streaming. Those are exactly what Google shipped. If you need diarised, timestamped output from a three-person call, the new model is the better tool for that step.
Where the two stop overlapping is the artifact. The AI report generator page covers turning source files into a written report, and Office document output is listed on the Pro plan.
Alternatives worth knowing
If your goal is meeting records rather than voice interfaces, three other routes exist.
Your meeting platform's own recap. Microsoft Teams collects the recording, shared files, notes, agenda, and follow-up tasks in one place after any recorded event. Intelligent recap features require Teams Premium or a Copilot licence.
A dedicated notetaker. These sit in the call, produce a summary, and keep the record inside their own product. Good when the meeting itself is the artifact.
Text output plus your source file. Slower to set up, but the only route where the numbers in the write-up come from the export rather than from someone reading them aloud.
Which one fits depends on a single question: is the valuable part of the meeting what people said, or what the data showed? The first three routes serve the former well. Only the last one serves the latter.
From transcript to deliverable
Gemini 3.5 Transcribe is a real step on a narrow problem. Cleaner text, three-speaker attribution, word-level timestamps, and 85-plus languages all reduce the editing that used to follow every recording.
What it does not do is decide what the meeting produced. That still comes from the data behind the discussion, which is why the transcript and the source file belong in the same place.
Our comparison with Gemini Deep Research covers the research side of the same question. To turn a recording and its source file into a finished summary, try Powerdrill Bloom.
Facts here are as of August 31, 2026, from Google's announcement page.
Frequently asked questions
What is Gemini 3.5 Transcribe?
It is Google's speech-to-text model announced on August 26, 2026. Google describes it as its most precise speech-to-text model yet, converting raw audio into polished, formatted text.
How accurate is Gemini 3.5 Transcribe?
Google publishes an average Word Error Rate of 4.0% for streaming and 2.6% for non-streaming use cases, as measured by Artificial Analysis. On the FLEURS benchmark the figures are 5.50% and 5.04%.
How many speakers can it identify?
Up to three speakers in pre-recorded audio, with timestamps. Google describes support for more than three speakers as experimental.
How do I access Gemini 3.5 Transcribe?
Through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. It also powers voice features in Gboard on Android, the Gemini app on macOS, and Google Antigravity, with Chrome described as coming soon.
Does it write the meeting summary for me?
The announcement describes transcription and delegation to other Gemini models rather than finished documents. Producing a written record with the underlying figures is a separate step, and it needs the source file as well as the audio.