Noise-robust speech emotion recognition

Most speech emotion models are trained on clean audio and fall apart in real environments. I trained CNN-based models on clean audio and on audio mixed with ten noise types at several signal-to-noise ratios. Combining magnitude and modified group delay spectrograms, adding synthetic noise in training, and adding attention each made the models noticeably more robust.

Related writing on this site:

• Adding noise to audio data
• Classifying emotions with neural networks
• Hiding emotions in plain sight