Musenet

MuseNet, developed by OpenAI, is a deep neural network that has the capability to produce 4-minute musical compositions featuring up to 10 different instruments. This innovative technology combines elements from various genres like country, Mozart, and the Beatles. Similar to GPT-2, MuseNet is built on a versatile unsupervised technology that can predict the next token in a sequence, whether it be audio or text. The model is trained on sequential data, where it learns to anticipate the upcoming note based on a given set of notes. To achieve this, chordwise encoding is utilized, treating each combination of notes played simultaneously as an individual "chord" and assigning a token to each one. Moreover, the inclusion of composer and instrumentation tokens allows for greater control over the type of music MuseNet generates. Notably, the model excels in creating compositions that seamlessly blend different styles and instruments, while also retaining a coherent long-term structure. To train the model, a diverse dataset from sources like Classical Archives, BitMidi, and the MAESTRO dataset is utilized.