Convolutional Neural Networks with Batch Normalization for Classifying Hi-hat, Snare, and Bass Percussion Sound Samples
Nicolai Gajhede, Oliver Beck, H.‐G. Purwins · 2016
After having revolutionized image and speech processing, convolutional neural networks (CNN) are now starting to become more and more successful in music information retrieval as well. We compare four CNN types for classifying a dataset of more than 3000 acoustic and synthesized samples of the most prominent drum set instruments (bass, snare, hi-hat). We use the Mel scale log magnitudes (MLS) as a representation for the input of the CNN. We compare the classification results of 1) a CNN (3 conv/max-pool layers and 2 fully connected layers) without drop-out and batch normalization vs. three variants, 2) with drop-out, 3) with batch normalization (BN), and 4) with both drop-out and BN. The CNNs with BN yield the best classification results (97% accuracy).