I benchmarked whether ASR captures a dictated phone number exactly, in German and French. Whisper: 34% exact. Same text written down: 91% [P]

A transcript can read perfectly and still carry the wrong phone number. The caller says +33 6 05 36 23 71, Whisper writes +33 6 75 36 23 71 on a simulated phone line, the sentence looks fine, and every form downstream no