Rank-based frame classification for usable speech detection in speaker identification systems

摘要

The performance of a speaker identification (SID) system degrades substantially when there is a mismatch between the training and testing conditions. Discriminating between temporal sections of speech signals which are speech-like (SID usable) and noise-like (SID unusable) while only retaining frames labeled SID usable can augment SID performance substantially. In this paper, a novel labeling system for SID usable and SID unusable frames is presented for a GMM based SID system. This is motivated by a control experiment demonstrating that very high SID accuracies are theoretically achievable by removing frames that contribute more to the scores of competing speakers rather than the true speaker. To blindly identify these SID usable and unusable frames, the Mahalanobis distance and an ensemble of decision tree classifiers (with boosting) were trained on a dataset which was different from the enrollment database for the SID system. The classifier based techniques yielded improvements over the base speaker identification system (all frames used) in all cases when the speech signal was corrupted with additive white or additive pink noise.