Excitation Source Feature for Discriminating Shouted and Normal Speech

Abstract

Dynamics of shouted speech production significantly vary from that of normal speech. These variations can be analyzed from excitation source information by using differenced electroglottogram (DEGG) signal. This work has two contributions. First, the proposal of a novel Glottal Open Phase Tilt (GOPT) feature derived from DEGG signal for discrimination of shouted and normal speech. Second, the construction of a database of speech and corresponding EGG signals for performance analysis of the proposed feature. In case of shouting, vocal folds vibrate faster and close abruptly. This leads to relative proximity of glottal opening and the following closing instances. This motivated the proposal of tilt feature for discriminating shouted from normal speech. The proposed feature is also extracted from ILPR signals that are known to approximate DEGG signals. Experiments on the collected dataset have provided shouted speech detection rate of 90.9% for DEGG and 76.37% for ILPR signals,