Assessing an Application of Spontaneous Stressed Speech - Emotions Portal

Abstract

Detecting and identifying emotions expressed in speech signals is a very complex task that generally requires processing a large sample size to extract intricate details and match the diversity of human expression in speech. There is not an emotional dataset commonly accepted as a standard test bench to evaluate the performance of the supervised machine learning algorithms when presented with extracted speech characteristics. This work proposes a generic platform to capture and validate emotional speech. The aim of the platform is collaborative-crowdsourcing and it can be used for any language (currently, it is available in four languages such as Spanish, English, German and French). As an example, a module for elicitation of stress in speech through a set of online interviews and other module for labeling recorded speech have been developed. This study is envisaged as the beginning of an effort to establish a large, cost-free standard speech corpus to assess emotions across multiple languages.