Mining Intonation Corpora Using Knowledge Driven Sequential Clustering

This work presents a mining methodology designed to cope with the usual data scarcity problems of intonation corpora which arises from the high variability of prosodic information. The methodology is an adaptation of a basic agglomerative clustering technique, guided by a set of domain constraints. The peculiarities of the text-to-speech intonation modelling problem are considered in order to fix the initial configuration of the cluster and the criteria to merge classes and stopping their split- ting. The scarcity problem poses the need to apply a sequential selection mechanism of prosodic features, in order to obtain the initial set of classes in the cluster. A searching strategy to select the best class among a set of alternatives is proposed, which provides useful prediction models for accurate synthetic intonation. Visualization of final classes by means of a modified decision tree brings graphical cues about contrastable prosodic information of the intonation corpus.