Speech Emotion Recognition

#1 on Kaggle

Can you teach a model to hear feelings? Consolidated 12,000+ audio samples across 5 datasets and extracted a rich set of time-series features including MFCCs, RMS energy, ZCR, spectral centroid, chroma, and mel spectrograms. The LSTM classifier, which was trained on 6 emotions, attained 90% accuracy landing #1 on Kaggle. Currently the notebook has 166 upvotes and 268 forks on Kaggle. Gender completely reshapes the feature profile, which I thought was interesting. Kaggle notebook →