Writing
Notes on building a transcription tool, including where ours falls over. We publish the failure modes because finding one in print is worse than reading about it here.
How to calculate word error rate, with a worked example
The WER formula is one line. Getting a number you can trust is the rest — aligning the two transcripts, deciding what counts as the same word, and knowing why the result can go past 100%.
·5 min
Why your transcript repeats the same phrase over and over
Sometimes a speech recogniser stops transcribing and writes one short phrase again and again. Why it happens, how to spot the quieter versions of it, and what to do with the part of the recording it swallowed.
·6 min
AI transcription hallucinations: what the research found, and how to catch them
Speech recognisers sometimes write sentences nobody said. What the best-known study measured, what the widely shared numbers do and don't mean, and a practical way to catch invented text before it reaches anyone.
·6 min
How to quote an interview accurately from a transcript
A transcript is a draft of what was said, not the record. Where machine transcripts go wrong in ways that change a quote's meaning, and a routine for checking every quote against the audio before it goes out.
·6 min
SRT vs VTT: what actually differs, and what you lose converting
SubRip and WebVTT look almost identical, and most tools treat them as interchangeable. They aren't quite — here's where they differ, which one to use where, and what a conversion quietly throws away.
·6 min
How to transcribe an interview with speaker names
Automatic transcription gives you "Speaker 1" and "Speaker 2", not names — and sometimes gets who said what wrong. A practical workflow, from recording to export, for an interview transcript you can quote from.
·5 min
Preparing interview transcripts for thematic analysis
What a transcript needs before you start coding — the level of detail, who said what, a way back to the audio, participant codes, and the ethics questions to settle first — and where automatic transcription helps and where it doesn't.
·6 min
When "auto-detect language" quietly translates instead of transcribing
In bilingual recordings, automatic language detection can hand you fluent English for something that was said in Chinese — and because the output reads perfectly, nobody catches it.
·9 min
Why a three-hour interview comes back with eleven speakers
Speaker labels drift on long recordings, and the reason isn't that the model is bad at voices. It's that long audio gets cut into pieces, and each piece counts from one.
·6 min
What "99% accurate" actually means
Every transcription company advertises the same number. Here's where it comes from, why your recording won't reproduce it, and the question worth asking instead.
·7 min