Integrating Recurrence Dynamics for Speech Emotion Recognition

Abstract

We investigate the performance of features that can capture nonlinear\nrecurrence dynamics embedded in the speech signal for the task of Speech\nEmotion Recognition (SER). Reconstruction of the phase space of each speech\nframe and the computation of its respective Recurrence Plot (RP) reveals\ncomplex structures which can be measured by performing Recurrence\nQuantification Analysis (RQA). These measures are aggregated by using\nstatistical functionals over segment and utterance periods. We report SER\nresults for the proposed feature set on three databases using different\nclassification methods. When fusing the proposed features with traditional\nfeature sets, we show an improvement in unweighted accuracy of up to 5.7% and\n10.7% on Speaker-Dependent (SD) and Speaker-Independent (SI) SER tasks,\nrespectively, over the baseline. Following a segment-based approach we\ndemonstrate state-of-the-art performance on IEMOCAP using a Bidirectional\nRecurrent Neural Network.\n