Plenary Session Abstract

The modern era of automatic speech recognition (ASR) began in 1971 with the publication of the ARPA report advocating a research and development effort. Speech being the natural communication mode, its desirable recognition (transcription) was soon going to be possible due to advances in computer technology. Perhaps independently, a 1972 task force at IBM came to the same conclusion, moreover pointing out the business need for a sink for the uncontrollably increasing speed and storage capabilities of computers. Hence the birth of the first ARPA and IBM sponsored ASR teams.

This talk will discuss how and why the original acoustic-phonetic, rule- and knowledge- based approaches were transformed into today's hidden Markov model (HMM) paradigm that focuses on information extraction from speech data. The state-of-the-art self-organized statistical methods are only minimally informed by experimental results concerning speech perception and production, or by the nature of the speech signal. Their success enjoys a growing influence in other language engineering fields, such as machine translation, parsing, topic discovery and tracking, sense disambiguation, etc.

Prospects of the current efforts to break out of the straitjacket of the accepted problem formulation will be assessed.

Last Update:  February 4, 1999         Ingo Höntsch