Corrective focus detection in Italian speech using neural networks

Abstract

The corrective focus is a particular kind of prosodic prominence where the speaker is intended to correct or to emphasize a concept. This work develops an Artificial Cognitive System (ACS) based on Recurrent Neural Networks that analyzes suitable features of the audio channel in order to automatically identify the Corrective Focus on speech signals. Two different approaches to build the ACS have been developed. The first one addresses the detection of focused syllables within a given Intonational Unit whereas the second one identifies a whole IU as focused or not. The experimental evaluation over an Italian Corpus has shown the ability of the Artificial Cognitive System to identify the focus in the speaker IUs. This ability can lead to further important improvements in human-machine communication. The addressed problem is a good example of synergies between Humans and Artificial Cognitive Systems.