The Ledger · Entry 009 · Perception
Conversational speech transcription
Ceded to machinesSince 2016Criterion performanceLast reviewed 2026-08-17
Event
Microsoft reports a 5.9 percent word error rate on the Switchboard telephone corpus, matching its measured professional transcribers; 5.1 percent follows in 2017.
The parity claim rests on one benchmark of English telephone conversations recorded in the 1990s and on a specific measurement of human error; IBM argued the human floor was closer to 5.1 percent, which Microsoft reached the following year. On clean, well-resourced speech the machines are now beyond ordinary listeners. Not ceded: heavy accents, overlapping speakers, noisy rooms and most of the world's languages, where error rates remain far higher; the human transcriber's advantage is context, not hearing.