Speech Emotion Recognition

Abstract

The in-depth learning based on the
focus on speech recognition and natural
language processing has become very popular.
With the focus system, the relevant encoding
context vectors make up the major chunk of the
decoding content construction, simultaneously
minimizing the effect of the irrelevant ones.
Driven by this idea, a methodology is proposed
in this work for the active selection of sub
pronounced representations to construct
discriminative pronunciation representations.
In comparison with the basic standard of a
model based on unified focus, a focused model
improves weighted accuracy by 1.46% in the
emotion classification work. Furthermore, the
selection distribution has evolved to a better
understanding of the sub-pronunciation piece of
an emotional value by users.