Knowledge Distillation with Latent Representations of Emotions
Ensemble of students taught by probabilistic teachers reduces confusion among students and at the same time improves their performance as well.
-
Speech emotion recognition system that gives scalable, reliable and consistent predictions, tailoring it to real-world applications
-
Knowledge transfer between teacher and student using the learned latent emotional representations by the teachers
-
Incorporating model diversity and preserving prediction uncertainty using Monte Carlo dropout and emsembling
-
Developed inference models that are faster and consistent in their predictions and achieved up to 5% increase in the prediction scores for emotional attributes
Video
[code]