MusicRL: Aligning Music Generation to Human Preferences

Geoffrey Cideron, Sertan Girgin, Mauro Verzetti, Damien Vincent, Matej Kastelic, Zalán Borsos, Brian Mcwilliams, Victor Ungureanu, Olivier Bachem, Olivier Pietquin, Matthieu Geist, Leonard Hussenot, Neil Zeghidour, Andrea Agostinelli
Proceedings of the 41st International Conference on Machine Learning, PMLR 235:8968-8984, 2024.

Abstract

We propose MusicRL, the first music generation system finetuned from human feedback. Appreciation of text-to-music models is particularly subjective since the concept of musicality as well as the specific intention behind a caption are user-dependent (e.g. a caption such as “upbeat workout music” can map to a retro guitar solo or a technopop beat). Not only this makes supervised training of such models challenging, but it also calls for integrating continuous human feedback in their post-deployment finetuning. MusicRL is a pretrained autoregressive MusicLM model of discrete audio tokens finetuned with reinforcement learning to maximize sequence-level rewards. We design reward functions related specifically to text-adherence and audio quality with the help from selected raters, and use those to finetune MusicLM into MusicRL-R. We deploy MusicLM to users and collect a substantial dataset comprising 300,000 pairwise preferences. Using Reinforcement Learning from Human Feedback (RLHF), we train MusicRL-U, the first text-to-music model that incorporates human feedback at scale. Human evaluations show that both MusicRL-R and MusicRL-U are preferred to the baseline. Ultimately, MusicRL-RU combines the two approaches and results in the best model according to human raters. Ablation studies shed light on the musical attributes influencing human preferences, indicating that text adherence and quality only account for a part of it. This underscores the prevalence of subjectivity in musical appreciation and calls for further involvement of human listeners in the finetuning of music generation models. Samples can be found at google-research.github.io/seanet/musiclm/rlhf/.

Cite this Paper

BibTeX
@InProceedings{pmlr-v235-cideron24a, title = {{M}usic{RL}: Aligning Music Generation to Human Preferences}, author = {Cideron, Geoffrey and Girgin, Sertan and Verzetti, Mauro and Vincent, Damien and Kastelic, Matej and Borsos, Zal\'{a}n and Mcwilliams, Brian and Ungureanu, Victor and Bachem, Olivier and Pietquin, Olivier and Geist, Matthieu and Hussenot, Leonard and Zeghidour, Neil and Agostinelli, Andrea}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {8968--8984}, year = {2024}, editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix}, volume = {235}, series = {Proceedings of Machine Learning Research}, month = {21--27 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v235/main/assets/cideron24a/cideron24a.pdf}, url = {https://proceedings.mlr.press/v235/cideron24a.html}, abstract = {We propose MusicRL, the first music generation system finetuned from human feedback. Appreciation of text-to-music models is particularly subjective since the concept of musicality as well as the specific intention behind a caption are user-dependent (e.g. a caption such as “upbeat workout music” can map to a retro guitar solo or a technopop beat). Not only this makes supervised training of such models challenging, but it also calls for integrating continuous human feedback in their post-deployment finetuning. MusicRL is a pretrained autoregressive MusicLM model of discrete audio tokens finetuned with reinforcement learning to maximize sequence-level rewards. We design reward functions related specifically to text-adherence and audio quality with the help from selected raters, and use those to finetune MusicLM into MusicRL-R. We deploy MusicLM to users and collect a substantial dataset comprising 300,000 pairwise preferences. Using Reinforcement Learning from Human Feedback (RLHF), we train MusicRL-U, the first text-to-music model that incorporates human feedback at scale. Human evaluations show that both MusicRL-R and MusicRL-U are preferred to the baseline. Ultimately, MusicRL-RU combines the two approaches and results in the best model according to human raters. Ablation studies shed light on the musical attributes influencing human preferences, indicating that text adherence and quality only account for a part of it. This underscores the prevalence of subjectivity in musical appreciation and calls for further involvement of human listeners in the finetuning of music generation models. Samples can be found at google-research.github.io/seanet/musiclm/rlhf/.} }
Endnote
%0 Conference Paper %T MusicRL: Aligning Music Generation to Human Preferences %A Geoffrey Cideron %A Sertan Girgin %A Mauro Verzetti %A Damien Vincent %A Matej Kastelic %A Zalán Borsos %A Brian Mcwilliams %A Victor Ungureanu %A Olivier Bachem %A Olivier Pietquin %A Matthieu Geist %A Leonard Hussenot %A Neil Zeghidour %A Andrea Agostinelli %B Proceedings of the 41st International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Ruslan Salakhutdinov %E Zico Kolter %E Katherine Heller %E Adrian Weller %E Nuria Oliver %E Jonathan Scarlett %E Felix Berkenkamp %F pmlr-v235-cideron24a %I PMLR %P 8968--8984 %U https://proceedings.mlr.press/v235/cideron24a.html %V 235 %X We propose MusicRL, the first music generation system finetuned from human feedback. Appreciation of text-to-music models is particularly subjective since the concept of musicality as well as the specific intention behind a caption are user-dependent (e.g. a caption such as “upbeat workout music” can map to a retro guitar solo or a technopop beat). Not only this makes supervised training of such models challenging, but it also calls for integrating continuous human feedback in their post-deployment finetuning. MusicRL is a pretrained autoregressive MusicLM model of discrete audio tokens finetuned with reinforcement learning to maximize sequence-level rewards. We design reward functions related specifically to text-adherence and audio quality with the help from selected raters, and use those to finetune MusicLM into MusicRL-R. We deploy MusicLM to users and collect a substantial dataset comprising 300,000 pairwise preferences. Using Reinforcement Learning from Human Feedback (RLHF), we train MusicRL-U, the first text-to-music model that incorporates human feedback at scale. Human evaluations show that both MusicRL-R and MusicRL-U are preferred to the baseline. Ultimately, MusicRL-RU combines the two approaches and results in the best model according to human raters. Ablation studies shed light on the musical attributes influencing human preferences, indicating that text adherence and quality only account for a part of it. This underscores the prevalence of subjectivity in musical appreciation and calls for further involvement of human listeners in the finetuning of music generation models. Samples can be found at google-research.github.io/seanet/musiclm/rlhf/.
APA
Cideron, G., Girgin, S., Verzetti, M., Vincent, D., Kastelic, M., Borsos, Z., Mcwilliams, B., Ungureanu, V., Bachem, O., Pietquin, O., Geist, M., Hussenot, L., Zeghidour, N. & Agostinelli, A.. (2024). MusicRL: Aligning Music Generation to Human Preferences. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:8968-8984 Available from https://proceedings.mlr.press/v235/cideron24a.html.

Related Material