what HIPAA actually requires of an AI transcription tool
9 min readHIPAA isn't just a BAA. the actual safeguards a transcription tool has to meet — physical, technical, administrative — and where on-device transcription removes whole categories of compliance work.
the diarization fragility test
8 min readspeaker diarization is the weakest part of every commercial AI transcription product in 2026. we tried to break it five different ways across every tool we could get our hands on. results.
the cleanup tax
6 min readevery transcription buyer pays a tax in time after the file arrives. it doesn't appear on the invoice. here's how big it is and where the money goes.
WER is a useless buyer metric
6 min readword error rate is the headline number every transcription vendor markets against. it tells you nothing about whether the transcript will save you time. what to ask vendors instead.
how to audit a browser-based transcription tool
7 min readany tool can claim 'on-device.' here's how to verify it: open the network tab, drop the audio in, watch what happens. the audit takes five minutes.
what's coming
posts in progress. when each one lands, the URL is announced to the list and added here.
- jefferson notation, automated. how we time pauses to a tenth of a second, detect overlap from word-level timestamps, and where the prosody-detection still falls short.
- the deposition format, derived. why 25-line pagination, why line numbers down the left, why the caption block matters, and what the model gets right and wrong about objections.
- NVivo, ATLAS.ti, MAXQDA: which schema wins. three CAQDAS packages, three import formats, one engine. tradeoffs, edge cases, and what we settled on.
- the diarization fragility test. we tried to break speaker diarization on every commercial transcription tool. crosstalk, identical-sounding voices, phone audio, swapped microphones. results.
- what HIPAA actually requires of an AI transcription tool. beyond the BAA — the safeguards, the retention, the access-log requirements, and where on-device transcription sidesteps the whole question.