Assessing Speaking Modes in Radio News Using Topic Classification and Acoustic Parameters

Authors: Sven Grawunder, Ute Gradmann

Abstract:

This study investigates an approach to the automatic classification of speaking modes in German radio news. Based on a large, manually annotated corpus of broadcast news, we combine acoustic analysis, text-based methods, and positional metadata to characterize and distinguish news modes. Acoustic parameters, lexical distributions, and normalized segment positions are analyzed separately and jointly. A supervised machine learning model integrating these features achieves nearly 90% classification accuracy across most modes. An ablation study highlights the relative contribution of textual, acoustic, and positional features. The results demonstrate the viability of transparent, multimodal approaches for speaking-mode classification in broadcast news.