In recent years a number of deep learning models have been developed to convert tapped rhythmic ideas into fully-voiced, dynamic drum performances. This mas-ter thesis extends the research by introducing a number of controllable features, namely Density, Intensity and Genre, allowing users to meaningfully augment the output whilst retaining the core rhythmic pattern identity. Our proposed models are comparatively small, enabling real-time usage on modern laptops. After trial-ing a number of methodologies and hyperparameters, we introduce our final model: VAEDER (Variational Autoencoder for Disentangled Expressive Rhythms). In ad-dition to the model development, we introduce a number of open-source software packages that allow researchers to quickly deploy symbolic generation models into Digital Audio Workstations. We hope that this will enable a new level of partici-pation and collaboration between researchers and musicians in the field of artificial intelligence for music generation.