Music Generation with Long Short-Term Memory Networks from MIDI Data of Classical Music

The generation of music has long been a challenging problem in the field of artificial intelligence. While rule-based approaches have had some success, recent advances in deep learning and sequence modelling present an opportunity for data-driven music composition. In this work, we leverage long short-term memory (LSTM) networks to generate new melodies based on a corpus of MIDI files. Our approach begins by extracting all note events from a datasets of MIDI files to form a corpus of musical sequences. We pre-processes this data by converting notes to a normalized representation and creating an encoded vocabulary of unique note events. An LSTM language model is then trained on these sequences for 200 epochs to learn the statistical patterns and dependencies present in the music corpus. At generation time, the model is seeded with a priming sequence and samples the next note event conditioned on the previous events. This process iterates autoregressively to generate a new melody reflecting the learned musical styles and patterns. We evaluate our generated melodies both quantitatively, in terms of held-out likelihood, and qualitatively through human listening studies. Our results demonstrate that LSTMs can effectively capture long-range dependencies in musical sequences to produce coherent and pleasing melodies resembling the original data distribution. We discuss strengths, limitations, and future directions for this promising approach to data-driven music generation.