Emotional Speech: A Spectral Analysis

摘要

Feature extraction and dimensionality reduction may be found as the most imperative parts of emotional speech recognition problem. In this work, we propose a new set of speech features based on the distribution of energy in frequency domain. To investigate the applicability of the proposed model, we have set the first international audio/visual emotion challenge (AVEC 2011) as a benchmark. As for the modeling and dimensionality reduction, we have employed the lasso. It is shown how 15 explicit spectral energy features, as suggested in this work, can lead to a more accurate model than those of the participants in the audio sub-challenge. This is while this number of features is less than ten percent of the smallest set of features in the challenge.