Shouted / normal speech classification using speech-specific features

This work explores the classification of shouted speech from normal speech which are typical scenarios encountered in news debates. The shouted and normal speech are studied in terms of the characteristics of the excitation source, vocal tract system and long term information. The different behavior of these speech characteristics in the shouted and normal class motivates to define features which provide good classification for the two classes. The Peak to side lobe ratio (PSR) of Hilbert envelope (HE) of Linear Prediction (LP) residual represents the source feature and the sum of ten largest peaks of the spectrum represents the vocal tract shape feature. The modulation spectrum energy represents the long term feature which indicates the slowly varying temporal envelope components in speech. These features are non-linearly mapped and combined for threshold based classification. It is observed that the sum of ten largest spectral peaks performed the classification task with highest F-score for almost all thresholds.