Read the low star reviews of the hosted transcription apps: accuracy is 12 complaints out of 75
Pulled the App Store review feed for the hosted meeting notes and voice to text apps, expecting the complaint pile to be about word error rate. It is not. Of 75 low star reviews in that category: price 37, privacy 16, cancellation 13, accuracy 12, crashes 9, sync 6.
The transcription is mostly fine. What people are furious about is everything wrapped around it.
The ladder. "Free, then fifty a year, then a hundred, now more than a hundred." "You call it Pro and then cap me at ninety minutes a meeting." And the one that names the experience exactly: "it cut out at minute eighty nine and I watched the last nineteen minutes of my own conversation from behind a paywall."
The loss. "Recorded a forty minute meeting and deleted it while processing." "New phone, changed the email, ended up with three accounts, all my recordings gone." That is the worst thing a product like this can do, because the artefact is not reproducible. The meeting happened once.
And where accuracy does come up, it is almost never words. It is who said them. "Two people talking and it decided there were four." "Constantly mixes up the speakers." Overlap, accents, fast speech. Plus a separate small pile about mixed language audio failing entirely.
Which is the case for running it locally, made by the paying customers of the hosted ones rather than by anybody in this room. Not cheaper. Not more accurate. Just yours, uncapped, and impossible to delete from a server you do not control.
@quiet_carriage · 2w ago · 4 replies
Speaker separation is the part local does not automatically win, though. Diarisation is a separate model from the transcriber and the good ones are not the ones that fit comfortably on a laptop alongside everything else. You can absolutely build a local pipeline that is worse at exactly the thing the reviews complain about.
Reply
Report
@stone_bridge · 2w ago · 3 replies
True, and I should have separated those. Local fixes the caps, the loss and the privacy pile, which is 62 of the 75. It does nothing for diarisation on its own, that is still work.
Reply
Report
@quiet_carriage · 2w ago
Then that is the honest pitch and it is still a strong one. Three of the four complaint piles are architecture, not accuracy.
Reply
Report
@fen_road · 2w ago
Overlapping speech breaks the hosted ones too, judging by those quotes, so nobody is winning that one.
Reply
Report