Automatic Speech Recognition at 60: Old and Immature

Automatic Speech Recognition (ASR) is a venerable discipline, with significant research papers going back to the early 1950's. Given this long history, ASR is often viewed as a mature field. However, like human beings, a research topic can be old without being mature. While major advances in speech engineering have occurred in the last 60 years, there has been limited progress in the basic principles; arguably, much of the improvement has been a fairly direct consequence of Moore's Law improvements in storage and computational clout, following some major advances in the 1970's and 1980's. In the meanwhile, ASR still fails quite badly in moderate amounts of noise, reverberation, and unexpected speaking styles or topics.

The presentation will review some of the history and present status of ASR systems, and end with some suggestions of promising directions, including recent developments in diagnostics that have some potential to provide a deeper understanding of the topic. It may be the case that the major limitations to ASR's use will not be possible before these diagnostics are developed and exploited.