A DTW-based directed acyclic graph (DAG) optimization method is proposed to exploit the interaction information of speech and speaker in feature component. We introduce the DAG representation of intra-class samples based on dynamic time warping (DTW) measure and propose two criteria based on in-degree of DAG. Combined with (l — r) optimization algorithm, the DTW-based DAG model is applied to discuss the feature subset information of representing speech and speaker in text-dependent speaker identification and speaker-dependent speech recognition. The experimental results demonstrate the powerful ability of our model to reveal the low dimensional performance and the influence of speech and speaker information in different tasks,and the corresponding DTW recognition rates are also calculated for comparison.
展开▼