Which transcript did you get?

Overview

The transcript lands an hour after the call, and you read it as a record of what was said. It is a record of one of two things, and it almost never tells you which.

Transcription has been sold as two separate products for years. Full verbatim keeps every word — the fillers, the repetitions, the false starts. Clean verbatim strips those out and hands you something readable. Legal transcription services make you pick before they start, and for testimony the full-verbatim version is the one that counts: a cleaned transcript of a deposition is not the record. Nobody asks you.

Why now. In August 2026 Google shipped a transcription model with both of those settings named in the API — verbatim and smart — and the documentation states a trade-off the launch post itself never mentions.

So here is the habit: before you rely on a transcript, find out whether it was written word-for-word or cleaned up, and when the exact words matter, check the tidy version against what was actually said.

The content

What “cleaning” actually does. The tidying isn’t cosmetic. Google’s own launch post describes smart mode handling self-corrections — their example is a speaker who says Tuesday and then corrects themselves to Wednesday — along with removing filler words and auto-formatting the result. On a scheduling slip that is a favour. Now run the same operation over a commitment. “I think we could probably ship in October” and “We could ship in October” are not the same sentence, and only one is a promise. Hesitation, hedging and self-repair are exactly the material that cleaning is designed to remove, and exactly where a speaker’s real level of commitment lives.

The trade is stated in the manual. This is the detail worth carrying away. In Google’s API documentation, smart transcription is “incompatible with timestamp_granularities and diarization_mode” — the timestamps and the speaker labels. You can have the readable transcript, or you can have the two features that let you find a passage in the recording and confirm who said it. Not both. The tidy version isn’t merely edited; it is edited and stripped of the handles you would use to check the edit.

The default is not the scary one. It would be neater to say everything you read has been quietly smoothed. Not here: verbatim is the documented default in that API, and a developer has to opt in to the cleaning. The honest version is duller and more useful. The setting exists, it varies by tool, and it is almost never surfaced to the person reading the output.

The market already prices this. Rev charges an extra $0.50 per audio minute for a verbatim transcript, and states plainly that without that box ticked you get the standard one, lightly edited for readability. GoTranscript sells clean verbatim for readability and full verbatim for research and legal proceedings, in those words. The distinction has carried a price tag since long before any current model.

The counterargument deserves a hearing. Transcription houses say clean verbatim removes only speech habits that do not change meaning, and much of the time they are right. The trouble is not that the claim is false. It is that you cannot test it from the tidy text — the removed material is, by construction, no longer there to inspect. That is why the check runs against the recording, not the transcript.

And it compounds. Whatever the transcript loses, the summary built on top of it inherits. Work on clinical dialogue summarisation puts it flatly: the errors transcription introduces can “significantly degrade the performance of downstream tasks like summarization.” Once the recap exists, nobody goes back to the transcript. Once the action items exist, nobody goes back to the recap. The wording gets one chance.

Try it

Do this once, on a tool you already use, and then only when it matters.

1. Find the setting. In whatever transcribes your meetings — Teams, Zoom, Otter, Meet, your notes app — open the transcription or recording settings and look for anything naming verbatim, raw, filler words, or “clean up” / “smart” formatting. Note which way it is set. Check what your workspace has enabled: on a managed account this may be an admin’s choice rather than yours, and if the option isn’t exposed anywhere, that is your answer — assume the readable version, and keep the recording for anything that could be contested.
2. Test it against something you were in. Take a recent transcript of a meeting you attended and find a passage where someone made a commitment, gave a number, or agreed to a scope. Read it back against the recording. You are not looking for errors. You are looking for hedges that are no longer there.
3. Ask for the words, not the gist. Paste the transcript into your assistant and ask: “Quote every sentence in this transcript where a speaker qualifies, hedges, or corrects themselves — quote exactly, do not paraphrase, and list them under the speaker’s name. If there are none, say so.” An hour of unhedged speech from four people is itself a finding, and usually means you are reading the cleaned version.

Where this breaks. The check is only as good as the recording, and plenty of meetings are transcribed without one being kept. With no audio to return to, the transcript is not evidence of wording whichever mode produced it — so if a commitment matters, get it in writing from the person who made it.

Additional reading

• Google, Audio transcription (Gemini API documentation, accessed 2 September 2026) — both modes in Google’s own words, the incompatibility with timestamps and diarization, and verbatim marked as the default. A living page; check it before quoting.
• Google, Intelligent transcription with Gemini 3.5 Transcribe (26 August 2026) — the launch post, including the self-correction example.
• Rev’s transcription add-ons and GoTranscript — the commercial distinction: an extra $0.50 per audio minute for verbatim, standard “lightly edited for readability”, and full verbatim named as the format for research and legal proceedings.
• Binici et al., MEDSAGE (August 2024) — transcription errors degrading the summaries built on them, in a domain where the cost is obvious.
• Dinkar, Clavel & Vasilescu, Fillers in Spoken Language Understanding (Traitement Automatique des Langues 63(3), 2022) — a survey across both traditions: psycholinguistics finds fillers informative, while much speech-recognition work has treated them as noise to detect and remove. The authors argue against that narrowness.

Editor’s note

I am cautious when it comes to these features. On one hand, they can be great for the obvious reasons. On the other hand, I have had a smart transcription cut out a particularly damning affirmative “Yeah.” from the record that was clearly audible in the audio recording and consequential in terms of the interpretation of the exchange (noise by the tool’s definition, evidence by mine). The last two points in the body (the counterargument, and the way it compounds) are the ones worth internalising.