Deep Convolutional Neural Networks for Predominant Instrument Recognition in Polyphonic Music Using Discrete Wavelet Transform

In this article, a new multi-input deep convolutional neural networks (deep-CNNs) model architecture is addressed for the recognition of predominant instruments in polyphonic music using discrete wavelet transform (DWT). The proposed deep-CNNs model employs a fusion of Mel-spectrogram and Mel-frequency cepstral coefficient (MFCC) features as its first input and a concatenation of statistical features extracted from decomposed signals obtained through DWT as its second input. Particle swarm optimization (PSO), a feature selection algorithm, is employed to minimize the feature dimensionality by excluding the irrelevant features. The proposed model is experimentally tested on the IRMAS dataset using fixed-length single-labeled train data for model training and variable-length multi-labeled test data for model evaluation. The proposed model is evaluated using several DWT feature dimensions, and a feature dimension of 250 yields the best outcomes. The model performance is assessed by averaging the precision, recall, and F1 measures on a micro- and macro-level. For a set of optimal model hyperparameter values, our proposed model can reach micro and macro F1 measures of 0.695 and 0.631, which are 12.28% and 23.0% greater as compared to the benchmark Han et al. (IEEE/ACM Trans Audio Speech Lang Process 25(1):208-221, 2016. https://doi.org/10.1109/taslp.2016.2632307) CNN model, respectively.