Comparison of Heterogeneous Feature Sets for Intonation Verification

The assessment of intonation, to which intonation verification belongs, has many applications, such as health-impaired people training – from individuals with Parkinson’s disease to children with Autism Spectrum Disorders – and second language learning. Most of the approaches that are found in the literature are based on intensive preprocessing of the audio signal and hand-crafted feature extraction methods, and most of those works do not tackle the particularities of the Portuguese language. In this paper, we present our work on intonation assessment, developed from a database of binarily-labelled Portuguese intonation imitation. This has been done using the set of Low Level Descriptors (LLDs) and eGeMAPS, both extracted with the openSMILE toolkit, and Problem-Agnostic Speech Encoder (PASE) features. We have taken the most informative feature subsets for prosody out of these. Distances between stimulus and imitation – the so-called similarity intonation scores – have been computed applying Dynamic Time Warping (DTW) for different feature subsets, and have afterwards been used as input features of a binary classifier. Performance achieves up to 66.9% of accuracy in the test data set when considering only one feature set, and it increases up to 77.5% for a set of seven features.