Dynamic selection of acoustic features in an Automatic Speech Recognition system

A general approach for integrating different acoustic feature sets and acoustic models is presented. A strategy for using a feature set as a reference and for scheduling the execution of other feature sets is introduced. The strategy is based on the introduction of feature variability states. Each phoneme of a word hypothesis is assigned one of such states. The probability that a word hypothesis is incorrect given the sequence of its variability states is computed and used for deciding the introduction of new features. Significant WER reductions have been observed on the test sets of the AURORA3 corpus. Using the CH1 portions of the test sets of the Italian and Spanish corpora, word error rate reductions respectively of 16.42% for the Italian and 29.4% for Spanish were observed.