tabahi / SER-LSTM-test

Speech emotion recognition in the wild

Real-time SER-LSTM-test

Speech emotion recognition in real-time using LSTM

Conventional LSTM based demo: https://tabahi.github.io/SER-LSTM-test/

Uses tensorflow JS library to predict emotions from speech MFCC features.

Trained Model

The model is trained with following layers and parameters using the IEMOCAP database to predict 4 basic emotions (Anger, Happy, Sad, Neutral, and Silence).

• Dense, 33
• LSTM, 16
• LSTM, 8
• Drop-out, 0.8
• Time Distributed Dense, 5
• Softmax

Note: The latency increases as the MFCC buffer size increases.

The new proposed method that performs better is demonstrated at: https://realtime-speech-emotion.netlify.app/