When computers learn to generate or classify music, they often represent notes as a long sequence of tiny tokens, like describing each note's pitch and duration separately. This creates very long sequences that are inefficient for machine learning models to process.
This paper applies Byte Pair Encoding (BPE), a compression technique from natural language processing, to music tokens. BPE merges the most common pairs of tokens into new, larger tokens, building a vocabulary of several thousand meaningful musical patterns from the data.
Using BPE cut the sequence length by about half while improving the quality of music generation and classification accuracy. It also made the models run faster. This means more efficient and better-sounding AI music systems for users.