Conceptual
Browse Instruct Curate

Discriminative Feature Learning with Center Loss for Speech Emotion Recognition

An end-to-end speech-emotion-recognition model that learns discriminative features from variable-length spectrograms by jointly minimizing softmax cross-entropy loss (which separates emotion classes) and center loss (which pulls same-class features toward a learned class center, shrinking intra-class variance). A CNN plus bidirectional GRU encodes the spectrogram to a fixed feature vector, avoiding the pair/triplet sampling and separate SVM of earlier metric-learning approaches.