Resources · Educational

AI Transcription Accuracy: What Organisations Should Realistically Expect

AI transcription vendors often cite impressive accuracy statistics. Before relying on these numbers to make procurement decisions, organisations should understand what they measure, what they do not measure, and what factors affect accuracy in real-world use.

How AI transcription accuracy is typically measured

AI transcription accuracy is typically measured using Word Error Rate (WER) — the percentage of words in the transcript that differ from the ground truth. A WER of 5% means roughly one word in twenty is wrong. A WER of 1% means roughly one word in a hundred is wrong. These statistics are generated under controlled conditions: clear audio, standard accents, neutral vocabulary. Real-world conditions in organisational settings — variable microphone quality, overlapping speakers, unfamiliar names and terminology, regional accents — typically produce higher error rates than benchmark statistics suggest.

What accuracy statistics do not tell you

Accuracy statistics do not tell you where in the transcript the errors will occur. AI transcription errors are not randomly distributed — they cluster in specific categories: unfamiliar proper nouns (people's names, place names, organisation names), technical terminology, low-energy speech (quiet voices, trailing sentences), overlapping speakers, and audio with background noise or recording artefacts. For organisations transcribing formal proceedings, this matters. The most important words in a council meeting — a councillor's name, a bylaw number, a street address — are exactly the kinds of words AI transcription is most likely to get wrong. A 98% accurate transcript that misidentifies the name of a bylaw being amended, or attributes a motion to the wrong councillor, is not a reliable record of that meeting.

How to evaluate accuracy for your specific use case

Rather than relying on benchmark statistics, organisations evaluating AI transcription should test accuracy with their own audio. Take a representative sample of recordings — typical meeting audio, in the environment where meetings are usually held, with the typical mix of speakers — and measure how many corrections a reviewer needs to make. The relevant question is not 'what is the WER?' — it is 'how long does it take a reviewer to bring the transcript to an acceptable standard?' If a reviewer can correct a one-hour meeting transcript in 20 minutes starting from an AI draft, that is a significant efficiency gain over manual transcription, even if the AI draft has noticeable errors. If the AI draft requires extensive correction, the efficiency gain is smaller.

Frequently Asked Questions

What can organisations do to improve AI transcription accuracy?
The most important factor is audio quality. Clear microphones positioned close to speakers, minimal background noise, and good meeting room acoustics all improve accuracy significantly. Custom vocabulary configuration for organisation-specific names and terminology also helps.
Does AI transcription accuracy improve over time?
AI models are updated regularly and accuracy does improve over time for common conditions. Custom vocabulary and model fine-tuning can also improve accuracy for specific organisational contexts.
Why does human review matter even for high-accuracy transcripts?
Even a 99% accurate transcript of a two-hour meeting contains approximately 40-60 errors (assuming average speaking pace of about 130 words per minute). Whether those errors affect the reliability of the record depends on where they occur. Human review catches the errors that matter most.