Fast Jukebox: Accelerating Music Generation with Knowledge Distillation

The Jukebox model can generate high-diversity music within a single system, which is achieved by using hierarchical VQ-VAE architecture to compress audio in discrete space at different compression levels. Even though the results are impressive, inference stage tremendously slow. To address this issue, we propose Fast Jukebox, uses knowledge distillation strategies reduce number of parameters prior for compressed space. Since has shown highly diverse generation capabilities, used simple compilation songs experimental purposes. Evaluation obtained emotional valence show that proposed approach tendency towards actively pleasant, thus reducing time all levels without compromising quality.