I dictate most of what I write. Every dictation is stored twice — what Whisper heard, and what a language model turned it into. These charts read straight from that data.
Amsterdam time
I dictate with OpenWhispr: Whisper Turbo on Groq for transcription, GPT-OSS for the cleanup pass.
Tue, Mar 3 to Sat, Aug 29 · Amsterdam time
1,408
dictations
19.7h
spent talking
106,387
words kept
90
usable words / min
84×
faster than real time
1.8s
key press to text
Volume first: what dictating instead of typing adds up to.
same numbers, different units
median 32s · 52 words
The longest single run was 6m 57s of uninterrupted talking.
112 of 180 days active
Whisper transcribes everything I say. A language model then decides what was worth saying.
14.9% of every word cut
The biggest casualty is “that”: 51% of the times I say it, it never reaches the page.
cut share 8.9% → 16.9%
The cut share has climbed from 8.9% to 16.9% — the more I trust the cleanup, the sloppier I let myself talk.
share of words changed
56% of dictations change by less than a tenth; 53 came out longer than what I said.
peak 11:00
37% of all dictations happen after 18:00.
Fri is quietest
What happens in the gap between releasing the hotkey and the words appearing.
1.8s median, end to end
70% of the processing time is the cleanup pass, not Whisper.
415 timed runs
The longest clips (over 5 min) still transcribe in about 1.4s — roughly 230× faster than real time.