The Switchboard corpus
Two and a half thousand telephone conversations between strangers were collected: spontaneous speech with interruptions, pauses and laughter.
Why it matters
The measure stopped being reading aloud and became how people actually talk, and error rates immediately doubled.
The conversations were recorded automatically over telephone lines at Texas Instruments: participants were given a topic and spoke as they liked. It exposed how inflated laboratory results were, as systems that read news well broke down on spontaneous speech. Switchboard remained the main reference point for a quarter of a century; it is the set on which parity with humans was announced in 2017.