Method and device for dynamically extracting speech emotion features

The invention discloses a method and a device for dynamically extracting speech emotion characteristics, which are characterized in that forward input data and reverse input data of speech data are respectively imported into N frame-level characteristic encoders, and the imported speech data are firstly used for dynamically extracting the frame-level fusion characteristics of the speech through a frame-level dynamic fusion unit; then cross-scale information between frames is obtained through a one-dimensional time sequence convolution unit, after normalization and activation processing, attention weight distribution is conducted on the obtained information through an attention unit, and the information acts on imported voice data. And respectively outputting the N forward voice emotion features and the N reverse voice emotion features, and importing the N forward voice emotion features and the N reverse voice emotion features into a global feature encoder for emotion fusion to obtain final advanced voice emot