But i’ve been refining a workflow that keeps 60-minute interviews at roughly 99% accuracy with a sub-3:1 turnaround. Yesterday’s benchmark was a 47-minute medical focus group delivered in 2:02 using a foot pedal, 1.25x playback, a quick AI pre-pass, and timestamps every 30 seconds — does that QC cadence line up with your results?
I’ve had better QC than “every 30 seconds” by anchoring timestamps to each moderator question and adding micro-stamps only during crosstalk; on a 47‑minute medical group it cut relistens by about 15% without dinging the 99%. Do you keep a term list handy for drug names or just lean on the pre-pass?
Static “every 30 seconds” stamps drive me nuts; with a foot pedal I map a 2s backstep and play‑on‑hold so I can stay at 1.25x for clean stretches and dip to 0.9x only on jargon — keeps me near your 2:02 on a 47‑min group. One concrete tip: feed the AI pre‑pass a custom vocab from the guide (drug/brand names) and QC only low‑confidence spans it flags instead of time blocks. @OP are you injecting a medical glossary or running the pre‑pass vanilla?
What’s worked for me is letting word‑level confidence drive QC: I highlight low‑confidence spans and diarization change points, then pedal through just those chunks to keep “99% accuracy” without babysitting clean stretches. If you’ve got Whisper or Trint handy, both expose per‑word confidence — are you using that, or something similar? Small caveat: on heavy jargon I still load a tiny term list so names don’t go off the rails.