Insights

AI vs human transcription: when the machine is not enough

AI transcription is fast, cheap and often very good. It is also wrong in a particular way, at a particular moment. Here is where it works, where it struggles, and how to decide which your recording needs.

The short answer Use AI transcription for clean recordings of one or two people speaking clearly into a good microphone: it is fast, cheap and often very good. Pay for a human transcript when the recording is hard, with several speakers, a noisy room, a phone line, strong accents, specialist terms or English and Afrikaans in the same sentence, or when a wrong word would cost more than the transcript. The real difference is not the average accuracy. It is what happens to the words a machine cannot hear.

Where AI transcription is genuinely good

Automatic transcription has become remarkably capable, and on the right recording you should use it. A single speaker close to a decent microphone, a quiet room, everyday vocabulary and an accent the system has heard often: dictated notes, a recorded lecture you need the gist of, a podcast made with proper microphones. On recordings like these the best AI transcription is quick, inexpensive and close enough to right that a careful read against the audio will catch what it missed.

Two habits make it safer. Check every name, number and date against the recording, because those are the words a machine is most likely to replace with something plausible. And keep the audio: a transcript you cannot check is a transcript you have to trust.

Where it struggles

The trouble starts when the recording stops being clean, which is most of the time in real work.

  • People talking over each other. In a meeting, an interview that warms up, or an argument, two voices at once are hard for a machine to separate. One voice usually wins, and the other is lost or merged into it.
  • Distance and rooms. A phone on a boardroom table hears the room as much as the people. Researchers building one of the best-known meeting transcription systems found “a substantial performance degradation” when the same system worked from microphones on the table rather than headsets (Hain and colleagues, 2007).
  • Phone and video calls. A call compresses the voice, drops syllables when the line breaks up, and adds echo when someone listens on speaker.
  • Accents and South African speech. Systems do best on the accents they were trained on most. Regional accents, fast speech and mumbled asides stretch them.
  • Specialist terms and names. Legal, medical and research vocabulary, family names, place names and organisations are exactly the words a general-purpose system is least sure of, and exactly the words that matter.
  • English and Afrikaans in one sentence. A speaker who moves between languages mid-sentence, as so many South Africans do, asks a system to switch languages without warning. That is where words are dropped, or rendered as English that sounds similar.

Why an average accuracy figure hides where the errors fall

Transcription accuracy is usually given as a single percentage, and a high one sounds reassuring. But a percentage is an average across every word in the recording. At 98 percent, a thousand words still carry twenty errors, and a one hour interview runs to roughly 9,000 words, so even that figure means around 180 wrong words.

More important than how many is where they fall. The easy stretches, one person speaking clearly, come back almost perfect. The errors gather in the hard stretches: the overlap, the accent, the poor line, the one specialist term the matter turns on. So the two percent that is wrong is rarely a random two percent. It is often the part you most needed. And a headline figure is usually measured on clear test recordings, not on a boardroom, a focus group or a call from a moving car.

The real risk: a confident guess

When a person cannot hear a word, they can say so. When a machine cannot hear a word, it usually writes the most likely one and moves on, in the same confident type as everything around it.

Take a consultation recording in which an attorney says that the amounts in dispute are set out at paragraph twelve of the affidavit. On a poor line a machine can render “affidavit” as “after David”: fluent, plausible, and a person who does not exist. Nothing on the page tells you to check it.

Sometimes it goes further. A 2024 study of a widely used speech recognition model found that roughly 1 percent of transcriptions contained entire phrases or sentences that were never spoken, and that 38 percent of those invented passages included explicit harms, such as references to violence. The inventions were more common for speakers with longer pauses (Koenecke and colleagues, 2024). The rate will differ between tools and recordings, but the pattern is the point: invented text reads as fluently as real text.

A human transcriber works the other way round. Where a word will not come clear, it is marked with its position in the recording, as [inaudible 00:14:32], or [unclear: affidavit? 00:14:32] where the word is probable but not certain. You lose the illusion of a perfect transcript and gain something more useful: you know exactly where to listen.

AI vs human transcription: a simple decision guide

A free or AI tool is usually enough when

  • one or two people speak clearly, close to the microphone;
  • the room is quiet and nobody talks over anyone;
  • you need the gist, or searchable notes, rather than a record;
  • you will check the transcript against the audio yourself;
  • nothing rides on any single word.

A human transcript is worth paying for when

  • there are three or more speakers, or people talk over each other;
  • the recording is a meeting room, a phone line or a video call;
  • the speakers have strong accents, or move between English and Afrikaans;
  • the recording is full of names, numbers or specialist terms;
  • the transcript will be quoted, cited, coded or relied on: a thesis interview, a statement, a hearing, a board record;
  • you need to know where the recording goes. Before uploading confidential material to any service, ask where it is processed and whether it can be used to train a model.

At Transcribe Refinery there is no machine draft at all. A person transcribes every recording against the audio, works the hard stretches until the wording is certain, and marks what will not come clear. That is how we work.

Which service fits your recording

Our human transcription services are set out by the kind of recording:

Not sure how closely the transcript should follow the recording? Our guide to clean verbatim vs full verbatim shows the same interview both ways. Every rate is on the pricing page, from R14.50 per recorded minute, all inclusive.

Tell us about the recording the machine got wrong.

Describe it and what the transcript is for. We reply with a written quote, usually within one working day. Nothing is sent, and nothing is charged, until you accept.

Get a written quoteWork out a price

Sources

  1. A. Koenecke, A. S. G. Choi, K. X. Mei, H. Schellmann and M. Sloane, Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’24). doi.org/10.1145/3630106.3658996
  2. T. Hain et al., “The AMI System for the Transcription of Speech in Meetings”, Proceedings of IEEE ICASSP 2007. cstr.ed.ac.uk

Published 28 October 2026 by Shireen Bruinders, founder of Transcribe Refinery, who brings 22 years of professional experience, including transcriptions for attorney disciplinary hearings at the Law Society of the Northern Provinces. More about Transcribe Refinery.