Human or machine transcription: how to tell which one your file needs

Machine transcription is free and often good enough. Here is where it stops being good enough, what human transcription costs per minute in the UK, and the five file characteristics that decide it.

Automatic transcription got good. A clean recording of one person speaking unaccented English, in a quiet room, into a decent microphone, will come back from a machine at an accuracy you would struggle to beat by hand — in seconds, for roughly nothing.

So the useful question is not whether machine transcription works. It is where it stops working, and whether your file is on the wrong side of that line.

Here is the short version.

Machine transcription is the right call for clean single-speaker audio in a major language, for internal notes, for searching your own archive, and for any draft you are going to read rather than publish.

Human transcription is the right call when the audio is messy, when more than two people talk over each other, when the vocabulary is specialist, when the language is not one of the big handful, or when something downstream depends on the transcript being right.

The five things that decide it

Machine accuracy does not fall off gradually. It falls off at specific, predictable features of a recording.

1. Crosstalk. Two people speaking at once is the single most reliable way to break automatic transcription. Focus groups, panel discussions, arguments, anything recorded on one room mic. Machines either drop a speaker or blend both into one sentence that neither person said.

2. Accents, and distance from the training data. Automatic systems are strongest on the accents they saw most. A broad regional accent, a second-language speaker, or a speaker switching between languages mid-sentence all cost accuracy, and they cost it silently — the output still looks confident.

3. Specialist vocabulary. Drug names, case citations, part numbers, product names, people's names. A machine will render an unfamiliar term as the nearest familiar word and move on. That is worse than leaving a gap, because nothing in the transcript flags it.

4. Language. Machine transcription in English is a different product from machine transcription in Malayalam, Amharic or Jamaican Patois. We transcribe in over 100 languages with native speakers, in the same language as the original, and the gap between human and machine widens sharply as you move away from the most-resourced languages. Do not assume parity across a language list.

5. What happens next. A transcript that becomes subtitles, a translation, evidence, a published quote or a compliance record carries every error forward into something more expensive to fix.

Verbatim or clean-read — decide before you order

This choice matters more than most people expect, and it is not an upsell.

Verbatim captures every word exactly as spoken, including interjections, false starts and cues like sighs or laughter. Use it when you need the raw content for analysis, editing or evidence — research interviews, anything where how something was said is part of the record.

Clean-read removes fillers and false starts and gives you a transcript that reads as finished prose. Use it when the transcript is going to be read by people — case studies, publication, internal documentation, subtitles.

Ordering the wrong one means paying twice. If you are unsure, send thirty seconds of the audio and ask.

What human transcription costs in the UK

Most providers quote on request. Ours are published, so you can build a comparison without a sales call:

• From £8 per minute of audio or video for English (€10 / $12).
• From £10 per minute for all other languages (€13 / $15).
• Minimum project fee from £150 (€175 / $200; £150 / €200 / $250 for non-English).
• Delivered as a Word document with timecodes. SRT and other subtitle files are priced on our subtitling rates.

All exclude VAT, and the final figure moves with audio quality, number of speakers, language and turnaround.

So a 30-minute English interview is a £240 job rather than a free one. The honest comparison is not £240 against zero. It is £240 against the hours someone on your team spends correcting a machine draft — and for a clean recording, the machine draft genuinely may need almost no correcting, in which case use it and keep the £240.

Where the money is actually wasted

Not on choosing human over machine. On choosing machine and then fixing it by hand.

Correcting a bad automatic transcript is slower than typing from scratch, because you are reading, comparing and second-guessing instead of just listening. If a junior person on your team is spending an afternoon per recording on cleanup, the machine was never free.

The second waste is ordering human transcription for files that did not need it. A clean webinar recording of one presenter does not need a linguist. We will tell you that.

What "human" should actually mean

"Human transcription" is used loosely. Worth checking what you are buying:

• Is the transcript produced by a person, or produced by a machine and skimmed by one?
• Is the person a native speaker of the language being transcribed?
• Is there a documented quality process behind it? Ours sits under ISO 9001 (quality management) and ISO 14001 (environmental management) certification.
• Does one team carry the project from source file to final transcript, or does it change hands?

Those four questions separate most providers faster than any price comparison.

A 60-second decision

Answer these. Two or more yesses and the file wants a person.

1. Do two or more people ever talk at the same time?
2. Are there strong accents, second-language speakers, or language switching?
3. Does it contain names, drugs, case law, part numbers or product names?
4. Is it in a language other than English, Spanish, French, German or Mandarin?
5. Will anything be built on this transcript — subtitles, a translation, a publication, a record?

All five no? Run it through a machine. You do not need us for that file.

Not sure which side your file is on? Send it — or thirty seconds of it — and we will tell you straight, including when the answer is "use a machine". We transcribe in over 100 languages, verbatim or clean-read, with timecodes, from our London studios, and an itemised quote normally comes back within four working hours.