Speech emotion recognition via learning analogies

This work introduces the few-shot learning paradigm in speech emotion recognition domain. Emotional characterization of segments is carried out through analogies, i.e. by assessing similarities and dissimilarities between novel known recordings. More specifically, we designed a Siamese Neural Network modeling such relationships on combined log-Mel temporal modulation spectrogram space. We present thorough experimentations performance proposed solution holistically, where it demonstrated that reaches state art rates when following standard leave-one-speaker-out protocol, while at same time being able to operate non-stationary conditions, with limited knowledge speakers and/or emotional classes. Finally, investigated activation maps layer-wise manner order interpret predictions made model.