Speaker Segmentation Based on Subsegmental Features and Neural Network Models

摘要

In this paper, we propose an alternate approach for detecting speaker changes in a multispeaker speech signal. Current approaches for speaker segmentation employ features based on characteristics of the vocal tract system and they rely on the dissimilarity between the distributions of two sets of feature vectors. This statistical approach to a point phenomenon (speaker change) fails when the given conversation involves short speaker turns (< 5 s duration). The excitation source signal plays an important role in characterizing a speakers voice. We use autoassociative neural network (AANN) models to capture the characteristics of the excitation source that are present in the linear prediction (LP) residual of speech signal. The AANN models are then used to detect the speaker changes. Results show that excitation source features provide better evidence for speaker segmentation as compared to vocal tract features.