Multi-resolution modulation-filtered cochleagram feature for LSTM-based dimensional emotion recognition from speech

Abstract

Continuous dimensional emotion recognition from speech helps robots or virtual agents capture the temporal dynamics of a speaker’s emotional state in natural human–robot interactions. Temporal modulation cues obtained directly from the time-domain model of auditory perception can better reflect temporal dynamics than the acoustic features usually processed in the frequency domain. Feature extraction, which can reflect temporal dynamics of emotion from temporal modulation cues, is challenging because of the complexity and diversity of the auditory perception model. A recent neuroscientific study suggests that human brains derive multi-resolution representations through temporal modulation analysis. This study investigates multi-resolution representations of an auditory perception model and proposes a novel feature called multi-resolution modulation-filtered cochleagram (MMCG) for predicting valence and arousal values of emotional primitives. The MMCG is constructed by combining four modulation-filtered cochleagrams at different resolutions to capture various temporal and contextual modulation information. In addition, to model the multi-temporal dependencies of the MMCG, we designed a parallel long short-term memory (LSTM) architecture. The results of extensive experiments on the RECOLA and SEWA datasets demonstrate that MMCG provides the best recognition performance in both datasets among all evaluated features. The results also show that the parallel LSTM can build multi-temporal dependencies from the MMCG features, and the performance on valence and arousal prediction is better than that of a plain LSTM method.