MuseNet

🎵 Dive into the endless creativity of music with MuseNet! 🎶 This deep neural network crafts 4-minute compositions, blending styles from country to Beatles. Not programmed with music knowledge, MuseNet learned from MIDI files. 🎹✨ Experiment, blend, and share your masterpieces! #AI #MusicGenius

MuseNet is a deep neural network that generates 4-minute musical compositions with 10 instruments, blending styles from country to Mozart to the Beatles.
Rather than being explicitly programmed with music knowledge, MuseNet learned patterns of harmony, rhythm, and style from hundreds of thousands of MIDI files.
MuseNet uses unsupervised technology similar to GPT-2, predicting tokens in sequences of audio or text.
Users can blend music styles in novel ways with MuseNet and experiment with different composer and instrumentation tokens for control.
MuseNet's long-term structure understanding is facilitated by a 72-layer network with 24 attention heads and a context of 4096 tokens.
Training data for MuseNet was collected from various sources including ClassicalArchives, BitMidi, and the MAESTRO dataset.
During training, MuseNet transposes notes, augments volumes and timing, and uses mixup on token embeddings for better generation.
The model also includes different kinds of embeddings for structural context, such as for tracking time and note relationships.
MuseNet encourages user creativity and offers the option to share generated pieces on platforms like Instaudio.