Dragon Systems

Contribution to AI

Statistical modelling won the argument about speech long before anyone bought a copy of the software. James Baker's mid-1970s work at Carnegie Mellon recast recognition as probabilistic decoding — hidden Markov models weighing acoustic evidence, rather than hand-written rules about how phonemes behave — and the venture that grew out of it existed to test whether that approach could be pushed onto hardware ordinary people owned. It could. A licensing deal with Apricot in 1984 produced the first personal computer that would open a file on a spoken command. DragonDictate-30K, shipped in 1990, was the first large-vocabulary general dictation system for PCs: it adapted to an individual speaker and accepted words one at a time, each separated by a pause. The pauses kept the market small but not the consequences. People who could not use a keyboard could write, work and study at a machine, and assistive use became the first durable commercial market for speech recognition. NaturallySpeaking, in 1997, took the pauses out and still ran on a 133 MHz Pentium, arriving years earlier than most of the field thought continuous dictation was feasible. The ending is a lesson in how loosely inventors hold their inventions: the engine outlived the sale to Lernout & Hauspie and that firm's collapse in accounting fraud, passing to ScanSoft, then Nuance, then Microsoft, while the Bakers were left with worthless stock.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).