For decades, music information retrieval (MIR) relied primarily on stereo audio files—finished products where drums, vocals, bass, and synths were inextricably baked together. While effective for basic analysis, stereo files presented an insurmountable "cocktail party problem" for tasks requiring instrument-specific data.
The largest multitrack collection holds the true source. When an AI is trained on these 1.2 million authentic stems, the result is a model that can split audio with 99.9% accuracy. Rumors suggest that both Google DeepMind and Sony have approached the ARSP to license the collection as "ground truth" data for next-generation audio AI. The Largest Multitrack Music Collection Ever- -...
The answer is lawyers.
For decades, the technology to properly separate the mono recordings of The Beatles into discrete multitracks was considered impossible. The band famously mixed down to four tracks, bouncing tracks together (sub-mixing) to free up space, effectively "baking" the ingredients together forever. When an AI is trained on these 1