Self-attention transfer networks for speech emotion recognition

A crucial element of human–machine interaction, the automatic detection emotional states from human speech has long been regarded as a challenging task for machine learning models. One vital challenge in emotion recognition (SER) is how to learn robust and discriminative representations speech. Meanwhile, although methods have widely applied SER research, inadequate amount available annotated data become bottleneck that impedes extended application techniques (e.g., deep neural networks). To address this issue, we present method combines knowledge transfer self-attention tasks. Here, apply log-Mel spectrogram with deltas delta-deltas input. Moreover, given emotions are time-dependent, Temporal Convolutional Neural Networks (TCNs) model variations emotions. We further introduce an attention mechanism, which based on algorithm order long-term dependencies. The Self-Attention Transfer Network (SATN) our proposed approach, takes advantage autoencoders source task, then recognition, followed by transferring into SER. Evaluation built Interactive Emotional Dyadic Motion Capture (IEMOCAP) demonstrates effectiveness novel model.