Unsupervised Representations Improve Supervised Learning in Speech Emotion Recognition

Authors

Niloufar Faridani, Amirali Soltani Tehrani, Ramin Toosi

Abstract

Speech Emotion Recognition (SER) benefits from combining self-supervised acoustic representation learning with supervised classification. This paper uses a Wav2Vec-based feature extractor to obtain acoustic features from audio segments, then feeds the resulting feature maps to a custom CNN classifier. Evaluated on the ShEMO dataset, the approach outperforms SVM and transfer-learning CNN baselines and demonstrates the value of unsupervised deep feature learning for SER without hand-crafted audio features.