Reading between the lines : modelling and recognition of paralinguistic phenomena

Abstract

The ability to understand the paralinguistic content of speech is central to the design of more natural human computer interfaces [106]. In this thesis we explore the detection of two types of speaker traits, emotion and speaker appeal, from the voice. Within the literature a wide range of tools and techniques have been proposed for the classification of paralinguistic traits. We explore some of the more popular features, such as prosodic measures, Mel frequency cepstral coefficients, and Teager energy operator features, for use in affect classification. We also apply a novel set of voice-quality features to the problem, and demonstrate that these are more robust to a vocal phenomenon called creak, which is often used to signal emotion.