Tap2Drum with Transformer Neural Networks

In a paper recently published by the Google Magenta team, in the context of au-tomatic beat generation and beat humanization, a new task called Tap2Drum is described. In this task, the aim is to transform a tapped pattern into a full-fledged drum beat. This master thesis implements, trains and evaluates a Transformer Neural Network model for Tap2Drum beat generation, motivated by the creative possibilities of such a system for musicians and producers. The dataset used for training, as in our baseline (GrooVAE Seq2Seq), is the Groove MIDI Dataset. Sev-eral experiments have been carried out, of which the best models have been selected to be evaluated. The evaluation performed consists of a mixture of objective metrics and subjective observations based on our personal listening sessions. According to said evaluation, our model achieves competitive results, comparable to those of our baseline, although we cannot definitively claim it performs better.