Google’s new Gemini 3.5 Transcribe does more than turn audio into words. It can remove filler words, process corrections, format text, identify speakers, and work with custom vocabulary. That is useful for voice notes and drafts, but users should not mistake polished output for a verbatim record.
Google’s Gemini 3.5 Transcribe is not just a speech-to-text tool. It is designed to turn messy spoken language into usable writing.
That is why it could be genuinely helpful.
Google announced Gemini 3.5 Transcribe on August 26 as a model for real-time and recorded-audio transcription. The company says it can remove filler words, handle self-corrections, add formatting, recognize custom vocabulary, identify speakers in recorded audio, and create word-level timestamps.
For everyday work, that can be more useful than old-fashioned dictation.
You could explain a customer follow-up after a service call, talk through a newsletter idea while walking, record meeting notes before they disappear from memory, or dictate a rough email. Instead of getting a page full of “ums,” repeated phrases, and unfinished thoughts, you can start with cleaner text to edit.
What is different
Traditional transcription tries to capture the words a person said. Gemini 3.5 Transcribe also tries to interpret what the speaker meant.
Google’s example is a correction such as, “Let’s meet Tuesday — no, Wednesday.” The model is designed to resolve that false start into clean text. Google says it can also remove filler words and use supplied vocabulary to better recognize specialized terms.
Its technical limits matter if you plan to build a workflow around it. Google’s documentation says recorded audio can be up to one hour per request. But processing is limited to 30 minutes when features such as speaker diarization or word-level timestamps are enabled. The live version is limited to 10 minutes per session.
That makes the tool more suitable for short voice notes, clips, interviews, and meetings than for dropping in an entire day of recordings without planning.
The important limitation
The cleanup feature is also the risk.
Ars Technica makes the key point: when an AI removes verbal stumbles and resolves corrections, it has changed the wording. That can be exactly what you want when the goal is a polished draft. It is not the same thing as an exact record.
Do not use a smart transcript as the only record of a legal interview, medical conversation, HR complaint, disciplinary meeting, financial instruction, or source interview. Keep the original audio where it is lawful and appropriate. Label the AI output as a draft. Check important quotes, names, numbers, dates, and commitments against the recording.
The same rule applies to meeting notes. Let AI produce a useful first pass, but make a person responsible for confirming decisions and deadlines.
A simple way to try it
Start with a low-stakes task:
- Record a two-minute explanation of an email you need to send.
- Run it through a transcription tool.
- Turn the output into a draft.
- Compare the result with what you intended to say.
- Edit before sending.
Gemini 3.5 Transcribe could make voice input much more useful for regular work. Just remember the difference between polished text and an exact transcript.
Bottom Line
Gemini 3.5 Transcribe's cleanup can make speech easier to use, but teams must distinguish readable reconstruction from a verbatim evidentiary record.