Quasi-continuous local codebook features for multilingual acoustic phonetic modelling

In this article we present a method for defining the question set used for the induction of acoustic phonetic decision trees. The method is data driven resulting in an ordered feature space in contrast to the usual categorical one consisting of phonetic attribute values. Visualization of the feature space verifies that the derived characteristics are meaningful. We apply the features to a multilingual speech recognition task, showing that comparable results to the standard method, using question sets devised by human experts, can be derived.