I'm currently working on an audio classification project using CNNs. The problem is I'm having trouble training my CNN. I doubt if there are outliers in my dataset but I don't know how to detect outliers in an audio dataset. I've searched google and found nothing helpful.
A first think what comes to mind is to train an autoencoder, then identify abnormal data by these heuristics:
- Is the reconstruction error large, for example remove the top 5% of the data?
- Are the codes (outputs of the encoder) within a densely populated region, or are they outliers? You could calculate the distance to Nth nearest neighbor.