Towards a Computational Model of Human Speech Recognition – Partha Niyogi (Bell Labs)
How should we do speech recognition research if we are to engage our twin goals of understanding how humans recognize speech and getting computers to do the same?
I will explore this issue and discuss in particular the possibility of approaches that are based on the notion of distinctive features. In order to mediate the mapping from linguistic invariance to acoustic variability, one will need the incorporation of statistical learning models of various sorts. The construction of such models will have to be guided by what we know about human speech production, perception and linguistics. I will articulate such a point of view, discuss phonetic-feature learning models within the structural risk minimization framework of Vapnik, and provide results obtained on various speech problems with such an approach.
Partha Niyogi obtained the B.Tech. degree from IIT, N. Delhi and the S.M. and Ph.D. degrees in EECS from MIT, Cambridge, USA. After a brief stint as a researcher at MIT, he joined Bell Laboratories in Murray Hill, NJ. Over the years he has worked on topics at the interface of learning theory and problems in speech and language. This has included work on language acquisition and change, formal learning theory, speech perception and recognition and so on. He is the author of one book: The Informational Complexity of Learning and several articles. He is also affiliated with the University of Chicago.