Speech prosody: from language characterization to paralinguistic use

Prosody plays an important role in spoken language, encoding both linguistic and paralinguistic information. In this talk I will present two studies that touch upon each of these aspects. First, I will show how long-term information from the speech signal, captured by i-vectors, can be employed for automatic language characterization. 24 languages, spanning several families/sub-families, were considered for this study and their centroid i-vectors were computed using one hour of recordings from each language. The centroids were then correlated with typological features, revealing that they relate to syntactic information. Based on these findings, a simple syntactic feature prediction system was proposed, obtaining a 87% accuracy in a leave-one-out evaluation setting. Second, I will investigate the usefulness of prosodic information for the automatic detection of one of the most often found paralinguistic phenomenon in conversation, laughter. Employing features shown previously to discriminate between laughter and speech, a syllable-based laughter detection system was put forward. It obtained similar or higher performance to other detection systems, that use more generic features. It also revealed that features related to voice quality and rhythm characteristics are more discriminative than commonly-used features, such as speech intensity or fundamental frequency.