Neural network to detect happy & angry emotion in voices (Keras, Tensorflow backend)

Voice dataset from:

1. English voice: http://smartlaboratory.org/ravdess/
2. German voice: http://emodb.bilderbar.info/start.html

data processing step

training with voice spectrograms

End-to-end folder contains jupyter notebooks which try several end-to-end models to take in raw audio signal as input (instead of spectrograms)

Findings: a) Spectrogram models seem to perform better than end-to-end models, possibly due to the case that information on several frequencies are not as comprehensively captured as compared to spectrogram models.

b) Model trained on classifying happy/angry voices are able to pick up other related emotion pairs like surprised/disgust, possibly due to overlap in voice information in similar emotions like (happy-surprised) & (angry-disgust).