An unsupervised deep-learning project (CSE425) that learns musical structure from the MAESTRO piano dataset and writes new music. MIDI files become 88-key piano rolls cut into fixed-length windows; an LSTM autoencoder learns to rebuild them and a sequence VAE samples new passages from its latent space, with a transformer added later for longer pieces. The output is judged on pitch-histogram distance, rhythm diversity and repetition, and the write-up is candid about where the samples stay conservative.
