Most English in the world is spoken by people who did not grow up in London or Chicago. A typical recording uploaded here is a project call with colleagues in Bangalore, Warsaw and Manila, or a conference talk given by a German engineer. The speech model was trained on that mix, not only on newsreader English, so General American, Southern British, Indian, Australian, Irish, South African and Nigerian English are all written as ordinary text.
Accuracy still varies. Clear speech in a mainstream accent on a decent microphone gives the cleanest result. Broad Glaswegian, Geordie, strong Jamaican patois, rural Appalachian speech and very heavy non-native accents at speed produce more errors, mostly in short function words and in names. Two people talking over each other is harder than any accent. Timestamps let you jump to an uncertain line and check it by ear.
English has two main spelling traditions and a transcript has to pick one. The model tends to follow the speaker: a British voice usually yields "organise" and "centre", an American one "organize" and "center". This is a tendency and not a setting, and a recording with mixed speakers may mix conventions. If the text is going into a report or a publication with a house style, run a spell check in the variant you need.
Homophones are the other classic problem. "Their" and "there", "affect" and "effect", "principal" and "principle" sound identical, and the model chooses by context, which is right most of the time and wrong occasionally. Filler words such as "um" and "you know" are mostly left out, and false starts are lightly cleaned, so the transcript reads as text and is not a strict verbatim record. For legal or linguistic work that needs every hesitation, plan a manual pass.
Business English is dense with abbreviations: KPI, SLA, EBITDA, Q3, B2B. Common ones are written correctly. Internal project names, product codes and a company's own acronyms are unknown to any model and get spelled as they sound, so "Kubernetes" is fine while your internal tool "Kestrel-9" may not be. Medical and legal terms are handled well when pronounced clearly, less well when abbreviated in passing.
Personal names deserve a check. "Sean", "Shaun" and "Shawn", "Catherine" and "Kathryn", "Stephen" and "Steven" cannot be told apart by sound. Names from other languages spoken in the middle of an English sentence are the least reliable part of any transcript. Compare them with the meeting invitation or the slide deck. Figures, dates and amounts are written in digits and are worth verifying whenever a decision depends on them.
English recordings here are mostly work and study: team meetings and client calls, university lectures, research interviews, webinars, podcasts and YouTube videos that someone wants to quote or summarise. Music is a different task: sung lyrics over instruments transcribe poorly, so use this for speech. Guests get 10 transcriptions of up to 3 hours each and read a free preview; the free allowance is 5 hours in total, and a free account keeps the full transcripts. AI chat with the recording, automatic summaries and PDF or Word export come with the paid plans, and subtitles can be translated into other languages.
The speech models run on servers we operate. Your audio is not forwarded to a third-party speech API, which is usually the first question a security or legal team asks before a client call may be uploaded anywhere.