I build transcription, and the files people actually send us look nothing like the benchmarks. A phone lying on a restaurant table. Three people talking over each other with one on speaker. A lecture recorded from the back row.
Curious what everyone else's worst case is, and whether anything actually handled it.