Basic filters for convolutional neural networks applied to music: Training or design?

When convolutional neural networks are used to tackle learning problems based on music or other time series, raw one-dimensional data are commonly preprocessed to obtain spectrogram or mel-spectrogram coefficients, which are then used as input to the actual neural network. In this contribution, we investigate, both theoretically and experimentally, the influence of this pre-processing step on the network's performance and pose the question whether replacing it by applying adaptive or learned filters directly to the raw data can improve learning success. The theoretical results show that approximately reproducing mel-spectrogram coefficients by applying adaptive filters and subsequent time-averaging on the squared amplitudes is in principle possible. We also conducted extensive experimental work on the task of singing voice detection in music. The results of these experiments show that for classification based on convolutional neural networks the features obtained from adaptive filter banks followed by time-averaging the squared modulus of the filters' output perform better than the canonical Fourier transform-based mel-spectrogram coefficients. Alternative adaptive approaches with center frequencies or time-averaging lengths learned from training data perform equally well.