MathRaaj/ser-fast-cnn-bilstm

Speech Emotion Recognition (CNN-BiLSTM-Attention)

This model was trained from scratch on the RAVDESS and TESS datasets.

Model Architecture

• Front-end: 4-block CNN for feature extraction from Mel Spectrograms.
• Mid-section: Bidirectional LSTM for temporal dependencies.
• Pooling: Multi-head Attention pooling.
• Back-end: Fully connected classifier.

Classes

0: neutral, 1: calm, 2: happy, 3: sad, 4: angry, 5: fearful, 6: disgust