Recognition of Continuous Speech: Segmentation and Classification using Signature Table Adaptation. AIM-213

This report explores the possibility of using a set of features for segmentation
and recognition of continuous speech. The features are not necessarily
"distinctive" or minimal, in the sense that they do not divide the phonemes
into mutually exclusive subsets, and can have high redundancy. This concept of
feature can thus avoid apriori binding between the phoneme categories to be
recognized and the set of features defined in a particular system.

An adaptive technique is used to find the probability of the presence of a
feature. Each feature is treated independently of other features. An unknown
utterance is thus represented by a feature graph with associated probabilities.
It is hoped that such a representation would be valuable for a hypothesize-test
paradigm as opposed to a one which operates on a linear symbolic input.