Investigations into phonological attribute classifier representations for CRF phone recognition

Classifier combination has long been a staple for improving robustness of ASR systems; we present an experiment where introducing phonological feature scores from another lab’s system [1] into our system gives a statistically significant improvement in Conditional Random Field-based TIMIT phone recognition, despite a standalone system based on their features performing significantly worse. The second part of the paper explores the reasons for this improvement by examining different representations of phonological attribute classifiers, in terms of what they are classifying (binary versus n-ary features) and representation of scoring functions. The analysis leads to the conclusions that while binary phonological feature estimates usually are worse than n-ary features, the combination of the two can be quite good if there are also differences in the feature definitions or training paradigm.