Harpy was a speech understanding system built at Carnegie Mellon University by Bruce Lowerre and Raj Reddy, completed in 1976. It was the culminating result of DARPA’s Speech Understanding Research (SUR) program, a multi-year effort that ran from 1971 to 1976 and funded a handful of groups to push past the small, isolated-word recognizers of the 1950s and 1960s.
Harpy could handle connected speech - words spoken in continuous sentences rather than one at a time with pauses - over a vocabulary of 1,011 words, often described as roughly the vocabulary of a three-year-old. To search the enormous space of possible word sequences efficiently, Harpy compiled its knowledge of grammar, pronunciation, and acoustics into a single large network and used beam search, keeping only a narrow band of the most promising candidate paths at each step and pruning the rest. Beam search proved so effective that it remained a standard technique in speech recognition for decades.
Harpy met the SUR program’s headline goals, demonstrating that a structured search over a unified knowledge representation could turn continuous speech into text on a usable vocabulary. It was still far from open-ended dictation, and it ran on expensive research hardware, but it established methods and ambitions that later commercial systems would build on.
For business readers, Harpy is a reminder of how much modern consumer AI rests on government-funded research. The continuous-speech techniques refined under DARPA in the 1970s flowed, decades later, into dictation software and the voice assistants now built into phones and speakers.