What is it about?
Many people use music to relax, but the effects of music are not the same for everyone—especially for people living with different mental health conditions. In this paper, we study how music listening is linked to short-term stress expression among people who disclosed having depression, anxiety, PTSD, or bipolar disorder. Using a large collection of public posts from Twitter (now X), we identify “listening moments” when a user shares a streaming link (for example from Spotify). We then measure stress-related language in the user’s posts 30 and 60 minutes after listening, and compare each group to carefully matched control users with similar activity patterns. We also test whether certain types of music are associated with different outcomes. Specifically, we analyze common genres (such as pop, rock, hip-hop, and jazz) and audio characteristics like tempo (slow vs fast), valence (sad-sounding vs happy-sounding), and how instrumental a song is. Our results show that the relationship between music and stress can differ across conditions and across music types, for example, some combinations are linked to higher stress rather than lower. Finally, we demonstrate a proof-of-concept recommendation approach that uses these findings to better rank songs that are more likely to be stress-reducing for a given listener.
Featured Image
Photo by Simon Noh on Unsplash
Why is it important?
Music is increasingly used in wellbeing apps and personalized recommendation systems, yet most tools assume that “calming music” works similarly for everyone. Our findings suggest this is not always true: music that helps one person may be neutral—or even linked to increased stress—for someone else, depending on their mental health context and the kind of music they listen to. By measuring stress changes shortly after real-world listening events, this work highlights why one-size-fits-all music recommendations can be risky when they are used for mood or stress management. This study is also important because it moves beyond small lab experiments and examines behavior at scale, in everyday settings, while still using careful matching and statistical modeling to make fair comparisons. The results can help researchers better understand how different groups respond to music in the wild, and help designers build more responsible, personalized, and harm-aware music experiences. Rather than promoting a single “best” genre or feature, we provide evidence that personalization should consider both the listener’s context and the musical characteristics that may be associated with better (or worse) short-term outcomes.
Perspectives
When I started working on this paper, I had a simple question in mind: people often say “music helps with stress,” but is that really true for everyone, in everyday life, and across different mental health conditions? I have personally seen how people turn to music during hard moments, sometimes to calm down and sometimes to sit with difficult emotions. That made me want to study music listening outside the lab, in a setting where people choose what to listen to on their own and then continue living their normal day. What stood out to me most is that the patterns were not one-size-fits-all. Some types of music that might be assumed to be calming were linked to higher stress for certain groups, while other combinations seemed more helpful, at least in the short term. I hope these findings encourage researchers and product teams to be more careful with “wellbeing” recommendations and to avoid oversimplified claims. My goal is not to label any genre as good or bad, but to show that personalization should be thoughtful, evidence-based, and sensitive to mental health context.
Parya Abadeh
University of Guelph
Read the Original
This page is a summary of: Music Listening, Mental Health, and Stress: A Computational Framework for Personalized Analysis and Recommendation, ACM Transactions on Information Systems, February 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3797892.
You can read the full text:
Contributors
The following have contributed to this page
