ESC-50 Audio data for binary classifier

Question

I'm building a binary sound classifier using the ESC-50 dataset. I have taken one class, "dog bark", to be positive and the rest of the 49 classes to be negative. As the dataset is imbalanced I'm running into lot of training issues. I tried building a model but couldn't get a f1-score greater than 0.3.

I'm using mfcc and fft as features. I have tried used LR and SVM to train without much success. Can't use Deep learning models as it's a real-time system and can't have much delay.

How can I approach this problem?

Answer 1

On the github page of the ESC-50 dataset, there is a list of tried classifiers as well as links to the relevant papers. The best ones are all using some kind of deep learning, mostly CNNs, and currently the best score is 85.5%. A baseline Random forest (notebook available here) achived 44.3% and SVM (available here) achieved 39.6%.

I would recommend you take a look at the approaches linked above and see if you can improve upon them. However, for higher accuracy you probably need some kind of deep neural network.