ViaVoice

The genuine surprise of the video. Rather than matching waveforms against a dictionary, it treated speech as a probability problem using hidden Markov models — and after a single training session it was hitting roughly 90% accuracy. It died because you had to read fifty sentences before it would work at all, not because it didn't. Stripped-down descendants ended up in call centres, car infotainment and the first version of Siri.

• “I think this is really good. I think it might have just been ahead of its time. I just don't think there's enough people in the world who need to do this.”

  Watch at 5:41 ↗Testing Technology That Was Ahead of Its Time

speech recognitionhidden markovwrong applicationsiri ancestor