FairSpeech

Written in June 2025, [Code]

Speech Emotion Recognition (SER) models are often biased due to gender-based expression differences and imbalanced datasets. To address this, we propose FairSpeech, a Transformer-based model that uses adversarial bias mitigation to significantly reduce gender bias while maintaining competitive performance. FairSpeech works by explicitly minimizing the emotion loss disparity between male and female speakers. (This is a project for the course CS371 - Introduction to Deep Learning)