Continuous speech recognition

Abstract

The authors focus on a tutorial description of the hybrid HMM/ANN method. The approach has been applied to large vocabulary continuous speech recognition, and variants are in use by many researchers, The method provides a mechanism for incorporating a range of sources of evidence without strong assumptions about their joint statistics, and may have applicability to much more complex systems that can incorporate deep acoustic and linguistic context. The method is inherently discriminant and conservative of parameters. Despite these potential advantages, the hybrid method has focused on implementing fairly simple systems, which do surprisingly well on large continuous speech recognition tasks, Researchers are only beginning to explore the use of more complex structures with this paradigm. In particular, they are just beginning to look at the connectionist inference of language models (including phonology) from data, which may be required in order to take advantage of locally discriminant probabilities rather than simply translating to likelihoods. Finally, the authors' current intuition is that more advanced versions of the hybrid method can greatly benefit from a perceptual perspective.>