Emotion Recognition Challenge 2019

This is a very first challenge that I participated in Deep Learning, because I've just been studying in this field for 3 months...

Final Result:

• 11th place in the 1st Round.
• 4th place in Final Round - Honourable Mention ( Exceeding our expectations O.O )

Dataset:

Train: https://drive.google.com/drive/folders/1TeguARxkKENBuEbBsZup1ZPt7Z-rBoeQ

Test: https://drive.google.com/drive/folders/1UU1H3dwPKS6CjviROqoW9RXpGo6hms6Z

Our Solution:

• First and foremost, we generated 3 types of sound signal to image for training this problem - STFT, Mel-spectrogram-RGB, Mel-spectrogram-GreyScale.
• Besides that, essembling 3 models, including RNN and CNN models with their precision point based on class (Turning Point) maked our result better .
• In addition, we tried to use mel spectrogram images in tranposed form due to operating principle of CNN - sliding frame in horizontal-order

References:

[1] https://haythamfayek.com/2016/04/21/speech-processing-for-machine-learning.html
[2] https://www.kaggle.com/ejlok1/audio-emotion-part-6-2d-cnn-66-accuracy