Music genre identification from Spectrogram

In this project we developed a model to detect 6 different music genres using ResNet CNN model. The CNN takes spectrogram for every 4 sec and predicts the class label and the model also uses global pooling strategy to pool features for variable length audios.
The model is trained on ~15Hrs of music data collected from youtube and tested on 3hrs. The model achieves accuracy of 93% on the heldout test data.