Text-Independent Phoneme Segmentation via Learning Critical Acoustic Change Points

The conventional methods of automatic text-independent phoneme segmentation detect phoneme boundaries via calculating the acoustic changes along speech signals followed by a peak picking procedure according to user-defined rules. Instead, this paper presents a learning-based method in which the phoneme boundaries are viewed as critical points in the acoustic change context of speech signals. First, we adopt a metric learning procedure in the calculation of acoustic changes, in order to make the acoustic changes at phoneme boundaries more discriminative. Then, latent-dynamic conditional random field is used to model the acoustic change context of speech signals for the detection of phoneme boundaries. The experiments demonstrate that our method outperforms the rule-based methods reported in previous work.