Why accuracy claims rarely survive contact with real audio
Every transcription vendor quotes an accuracy figure, and almost all of them were measured on clean, single-speaker American English. Feed the same engine a two-person interview recorded in a Birmingham café, with a Black Country accent on one side and trade jargon on both, and the word error rate can double or triple. That is not a reason to avoid AI transcription; it is a reason to test tools on your audio before committing to a subscription.
The good news is that the underlying speech models, including OpenAI's Whisper family, Google's speech models, and the engines behind tools like Otter, Rev and Descript, have improved substantially on UK regional accents in recent years. The differences now show up less in everyday vocabulary and more in three stubborn areas: proper nouns, domain jargon and overlapping speakers.
The three failure modes that matter for UK businesses
Regional accents
Strong Brummie, Black Country, Glaswegian, Scouse and Geordie accents remain harder than RP or Estuary English for most engines. Typical failures are vowel-driven substitutions: place names mangled, "canna" and "ain't" style contractions rewritten oddly, and dropped word endings guessed wrongly. Tools built on larger multilingual models generally cope better than lightweight real-time engines, which is why live-captioning accuracy often trails offline transcription of the same recording.
Jargon and proper nouns
Every industry has its landmines: SKU names, medication names, legal terms, product codes. Generic engines will confidently substitute a common word for your niche one, and because the sentence still reads plausibly, errors slip through review. The single most valuable feature to look for is a custom vocabulary or "boost" list, offered by most business-grade tools, where you preload client names, product names and acronyms.
Multiple speakers and crosstalk
Speaker diarisation (who said what) is the weakest link in the chain. Two polite speakers taking turns transcribe well. Four people interrupting each other on a speakerphone do not. If your use case is meetings, judge tools on diarisation quality, not raw word accuracy.
Need a hand with this?
Our team delivers AI & Machine Learning for UK businesses — with a free initial consultation, transparent fixed quotes and no lock-in contracts. Tell us what you're working on →
How to run your own 30-minute benchmark
Do not trust anyone else's league table, including this article's verdicts, without a quick test on your own material. The method:
- 1. Pick three real recordings of about five minutes each: your hardest accent, your most jargon-heavy content, and your messiest multi-speaker meeting.
- 2. Create a reference transcript of one two-minute chunk from each by hand. Tedious, but it is your ground truth.
- 3. Run all three files through each candidate tool on a free trial.
- 4. Count the errors per 100 words in the chunks you transcribed by hand: substitutions, deletions and insertions all count as one error.
- 5. Separately score diarisation: what percentage of speaker turns are attributed to the right person?
- 6. Note the time you spend correcting each transcript to publishable standard. Cleanup time, not word error rate, is the number that hits your payroll.
A tool that is 2% less accurate but has a dramatically better correction editor often wins on total cost.
Verdicts by use case
Interviews and podcasts
Editor-first tools such as Descript shine here because transcription is only step one; you will trim, correct and repurpose. Whisper-based workflows (either via apps like MacWhisper or a developer running the open-source model) offer excellent accuracy on UK accents at very low cost if you have someone slightly technical. Prioritise custom vocabulary and export flexibility.
Meetings and sales calls
Meeting assistants like Otter, Fireflies and the native transcription in Microsoft Teams and Google Meet trade a little accuracy for integration: calendar joins, summaries and action-item extraction. For sales teams, call-intelligence platforms that push transcripts into your CRM add the most value. Here diarisation and searchability matter more than a perfect verbatim record.
Video subtitles
Accuracy standards are highest here because errors are public. Use a tool with a proper subtitle editor and SRT/VTT export, and always human-review before publishing. Note that under UK accessibility expectations, and for anyone pursuing public-sector work, accurate captions are moving from nice-to-have to required, so build review time into your video budget.
Regulated or confidential content
For legal, medical or HR recordings, check where audio is processed and stored, whether the vendor trains models on your data (look for an explicit opt-out), and whether UK/EU data residency is available. A self-hosted Whisper deployment keeps everything in-house if confidentiality trumps convenience.
Pricing traps and data protection basics
Pricing models vary between per-minute credits, monthly seat licences and compute costs for self-hosting. Watch for these traps:
- Free tiers that cap monthly minutes low enough to force an upgrade in week two.
- Per-seat pricing when only one person actually transcribes; a shared workflow may need just one paid seat.
- Export paywalls: some tools let you read transcripts free but charge to export them.
- AI summary add-ons billed separately from transcription itself.
On data protection: recordings of identifiable people are personal data under UK GDPR. Tell participants calls are recorded and transcribed, have a retention period and delete old audio, and check your vendor's data processing agreement covers international transfers if processing happens outside the UK. None of this is onerous, but skipping it turns a productivity tool into a compliance headache.
Key Takeaway
Ignore vendor accuracy claims and run a 30-minute benchmark on your own audio: your hardest accent, heaviest jargon and messiest meeting. Score correction time, not just word errors, and check diarisation separately if you transcribe meetings. Choose editor-first tools (like Descript or Whisper workflows) for interviews, integrated assistants for meetings and calls, and always human-review subtitles. Load custom vocabulary lists, confirm the vendor does not train on your data, and treat recordings as personal data under UK GDPR.
Making the choice and wiring it into your workflow
Shortlist two tools per use case, run the 30-minute benchmark above, and pick on total cleanup time plus integration fit rather than headline accuracy. Then make the tool earn its keep: pipe meeting transcripts into your CRM or project tracker, turn interview transcripts into blog drafts and social clips, and publish subtitles on every customer-facing video. Transcription is cheap; the value is in what you do with the text. If you would like help choosing tools or integrating transcripts into your CRM and content pipeline, our team can help.