Fernandez, R. "A Computational Model for the Automatic Recognition of Affect in Speech"

Abstract

Spoken language, in addition to serving as a primary vehicle for externalizing linguistic structures and meaning, acts as a carrier of carious sources of information, including background, age, gender, membership in social structures, as well as physiological, pathological and emotional states. These sources of information are more than just ancillary to the main purpose of linguistic communication: Humans react to the various non-linguistic factors encoded in the speech signal, shaping and adjusting their interactions to satisfy interpersonal and social protocols.

Computer science, artificial intelligence and computational linguistics have devoted much active research to systems that aim to model the production and recovery of linguistic lexico-semantic structures from speech. However, less attention has been devoted to systems that model and understand the paralinguistic and extralinguistic information in the signal. As the breadth and nature of human-computer interaction escalates to levels previously reserved for human-to-human communication, there is a growing need to endow computational systems with human-like abilities which facilitate the interaction and make it more natural. OF paramount importance amongst these is the human ability to make inferences regarding the affective content of our exchanges.

This thesis proposes a framework for the recognition of affective qualifiers from prosodic-acoustic parameters extracted form spoken language. It is argued that modeling the affective prosodic variation of speech can be approached by integrating acoustic parameters from various prosodic time scales, summarizing information from more localized (e.g., syllable level) to more global prosodic phenomena (e.g., utterance level). In this framework speech is structurally modeled as a dynamically evolving hierarchical model in which levels of the hierarchy are determined by prosodic constituency and contain parameters that evolve according to dynamical systems. The acoustic parameters have been chosen to reflect four main components of speech thought to reflect paralinguistic and affective-specific information: intonation, loudness, rhythm and voice quality. The thesis addresses the contribution of each of these components separately, and evaluates the full model by testing it on datasets of acted and of spontaneous speech perceptually annotated with effective labels, and by comparing it against human performance benchmarks.