Bruce Lowerre ’ s HARPY speech recognition system

1. Bruce Lowerre’s HARPY speech recognition system

2. Introduction • Developed by Lowerre & Reddy in 1979 • A turning-point for recognition systems development • Significant break with architectures of Hearsay and HWIM • Hearsay and HWIM artificial intelligence approach to speech understanding

3. How did it work? • 1st Large vocabulary continuous speech recognition (LVCSR) system • Reasonable performance • Replaced multiple knowledge sources with integrated network of spectral templates • Used rule firing by graph search

4. Significance • Showed good recognition could be achieved through good engineering • Showed good quality linguistic knowledge was not required • Did threaten to split field into two • People who accept any computational framework for recognition • People who sought an explanation of human processing using familiar symbolic manipulation

5. What was learnt • Dennis Klatt studied Harpy to deduce a cognitive model called LAFS (Lexical Access from Spectra) • Alan Newell attempted to link Harpy’s integrated search into AI production systems • However both failed to be taken seriously

6. Criticism by Donald Norman • Harpy’s performance was considerably worse than a human • Harpy’s architecture was only one of many potential architectures for speech recognition • Harpy did not show how higher level linguistic constraints could be incorporated in the search

7. Conclusion • High performance came from postponing decisions • Not to identify segments until such time as top-down information about potential word sequences were available • Harpy started the divergence between word accuracy and contemporary linguistic or psycho-linguistic wisdom