Speech Emotion Recognition Using Multihead Attention in Both Time and Feature Dimensions

To enhance the emotion feature and improve performance of speech recognition, an attention mechanism is employed to recognize important information in both time dimensions. In dimension, multi-heads modified with last state long short-term memory (LSTM)'s output match accumulation characteristic LSTM. scaled dot-product replaced additive that refers method update LSTM construct attention. This means a nonlinear change replaces linear mapping classical Experiments on IEMOCAP datasets demonstrate could emotional recognition.