MUSICGEN: simple and controllable music generation

Summary

Paper link: https://arxiv.org/pdf/2306.05284v1.pdf
We solve the task of conditional music generation. We introduce MUSICGEN, a single language model (LM) that operates on several streams of compressed discrete music representations (i.e., tokens). Unlike previous work, MUSICGEN consists of a single-stage transformer LM and an effective token interleaving pattern, which eliminates the need to cascade multiple