SC-ERM: Speaker-Centric Learning for Speech Emotion Recognition

We propose a novel deep learning-based model for speech emotion recognition, SC-ERM, which focuses on speaker-centric learning. This model effectively estimates emotions and demonstrates the ability to generalize to unseen speakers. Our proposed model utilizes speaker-specific emotion characteristics in two steps: first, it extracts emotion representations using an emotion encoder, and second, it employs speaker-centric learning by incorporating speaker style embeddings as a condition through a speaker mask generator. We evaluate our model's performance using an emotional dataset and find that it demonstrates out-standing performance in recognizing emotional states. Notably, it achieves a 9.2% relative improvement in accuracy compared to the baseline when classifying emotions for speakers not seen during training. Overall, our model demonstrates promising performance in accurately identifying emotions across a range of emotional expressions, irrespective of the speakers involved.