AI transcription accuracy, honestly

5-15% word error rate. Here’s what that actually means.

Modern AI speech-to-text achieves word error rates of 5-15% on general Australian English audio in good acoustic conditions, meaning 85-95% of words are transcribed correctly. After human review, accuracy typically exceeds 98%.

A 10% word error rate sounds significant. In a 30-minute interview of around 3,000 words, that is 300 errors. But errors are not randomly distributed. They cluster in proper nouns, technical terminology, heavily accented speech, and overlapping speech. The substantive content of candidate responses, the experiences described, the competencies demonstrated, is transcribed at much higher accuracy.

The right question is not whether AI transcription is perfect. No transcription method is. The right question is whether AI-assisted transcription with human review produces a more complete and consistent record than manual scribing alone.

Evaluate AI transcription against the manual alternative you actually use, not against a theoretical perfect record.
About Nguma

AI interview management and transcription built for Australian Government compliance.

Nguma automates APS interview scribing, Selection Reports, and meeting transcription. Everything processed onshore in Australia. No AI decision-making. Full audit trail for Merit Protection review. 51% Indigenous-owned.

Learn about Nguma →