Beat Tracking with Autoencoders

This work is a study into the e˙ectiveness of autoencoders applied to beat tracking. We propose an autoencoder architecture based on a state-of-the-art beat tracking network, the purpose of which is to create a new set of features for the beat tracking network to be trained on. The autoencoder is trained on Mel spectrograms extracted from pieces of music. After training, we pass the Mel spectrograms through the encoder half of the autoencoder only, and use the output as a new set of features. The new set of features is then used to train a beat tracking system, the results of which are compared with a reference beat tracking system trained on the Mel spectrograms. The aim of using autoencoders is to perform feature fusion on the input, resulting in an eÿcient encoding of the input to reduce its dimensionality, which should allow for an increase in performance when using this encoding to train the beat tracking system. Our testing showed on par performance with the reference system when using the test loss of the beat activation function as our evaluation metric, and a slight increase in performance when using more typical beat tracking evaluation metrics, such as the F -measure and continuity based metrics