Exploring Music Representation Learning for Detection of Finer-grained Details

The art of music is profound and includes the fine-tuning of multiple fine-grained aspects. Taking the example of one such attribute, the music tonic, this paper investigates if machine learning can be effectively applied to detect fine-grained features from various representations of music. It describes the methodology and experiments to determine the effectiveness of the audio representations and non-linear machine learning classifiers such as Support Vector Machines, K-Nearest Neighbors, and tree-based methods in capturing and detecting tonic informa-tion. The effectiveness of Mel-frequency cepstral coefficients is compared with Bark-frequency cepstral coefficients and Mel spectrogram data, which are other audio representations. The feature extraction of the Mel-frequency cepstral coef-ficients is performed using two different tools for comparison. A spectral analysis of the music renditions in a chosen dataset using Principal Component Analysis, and other dimensionality reduction techniques, t-distributed stochastic neighbor embedding, auto-encoder, and Uniform Manifold Approximation and Projection provided valuable insights into the representation learning of music in general. Classification using fine-tuned machine learning models demonstrated the fea-sibility of automating the detection and possible prediction of the music tonic. The work also resulted in a dataset with vocal renditions of the second and third authors of this paper that was used for some of the experiments.