Most of what we publish falls into two buckets: how to get usable text out of a recording, and what a given speech model can or cannot do. Both age quickly — an export format changes, a hosted model moves behind a waitlist, an accuracy figure turns out to describe a different test set. This page sets out the standard those pages are held to.
Every workflow we describe is run end to end first: a real file goes in, and the steps, settings, and exported output are recorded from that run rather than from a feature list. For coverage of a third-party model we read the provider's own documentation, note the date we read it, and keep a link to it. Where a claim comes from a launch video or a press page rather than documentation, the page says so in plain language.
Word error rates, language counts, and speed multipliers travel a long way from their original test conditions. We attribute each figure to whoever produced it and do not restate a vendor figure as an independent benchmark. Should we ever run our own evaluation, the page carries the audio set, the settings, the date, who ran it, and what the result does not cover — without those, a number is decoration.
A page is revised when a change would send a reader down the wrong path: a renamed export option, a model that stopped accepting new sign-ups, a pricing tier that no longer exists. Cosmetic drift is left alone. If something here is wrong or stale, send the URL and whatever source shows it to support@gpt-transcribe.org; substantiated reports are corrected on the page itself rather than in a footnote.
Work published under the GPT Transcribe Editorial Team byline comes from the people who build and support the product. It is an organizational byline, not a stand-in for a named researcher, and we do not present it as an independent testing lab.