Scientists have developed a new way to identify hit songs that people will love by measuring listeners’ brain activity responses using machine learning AI.
This approach was much more accurate than asking people if they liked each song.
Key Facts:
- Researchers measured people’s brain activity while they listened to 24 new songs.
- They tracked emotional engagement and attention levels using wearable sensors.
- Asking people to rate how much they liked each song did NOT predict popularity.
- But brain responses did – they identified hits with 69% accuracy.
- Using machine learning on brain data boosted accuracy to 97% for hits.
- Even just 1 minute of brain data predicted hits with 82% accuracy.
Source: Front. Artif. Intell. 20 June 2023 (Sec. Machine Learning and Artificial Intelligence)
The Struggle to Pick Hit Songs
Every day, thousands of new songs are released worldwide.
But only a tiny fraction will become hits.
This creates a huge challenge for music streamers and radio stations trying to identify songs people will love.
Traditionally, experts analyze song elements like lyrics, tempo, genre etc. to predict hits.
But accuracy remains low. For example, even veterans can’t predict hit movies any better than random guessing.
People want to hear new music. But studies show we prefer songs similar to ones we already enjoy.
So when asked to rate new, unfamiliar songs, people tend to give low ratings.
This makes it hard to predict what will be a hit.
Could Brain Data Do Better?
To try a new approach, researchers at Claremont Graduate University teamed up with a music streaming company.
The company provided 24 new song releases – 13 hits and 11 flops based on streaming data.
They brought in 33 participants aged 18-57 to listen to clips of these songs in the lab.
The goal was to compare ratings of “liking” each song to actual brain responses.
Could neural activity predict hits better than self-reports?
Measuring Emotional Reactions
Participants wore wireless sensors that measured heart rate as they listened to the songs.
Heart rate data was fed into algorithms that gauge emotional engagement and attention levels.
Why track emotions? Studies show music activates brain areas linked to emotion and memory – not just the auditory cortex.
Emotions likely drive whether songs become popular.
The sensors provided a composite immersion score combining attention and emotional resonance.
Researchers also extracted two additional measures from the data – peak immersion and retreat.
Peak immersion captured heightened engagement during a song.
Retreat showed decreased attention. They hypothesized hits would show higher immersion and lower retreat.
Self-Reports Failed to Predict Hits
After each song, participants rated how much they liked it from 1-10.
For songs they were unfamiliar with, self-reported liking did NOT differ between hits and flops.
Liking also didn’t correlate with actual popularity measured by streaming numbers.
Asking people to predict hits or recommend songs was similarly useless.
Brain Responses Showed Promise
In contrast, hit songs showed significantly higher immersion than flops based on brain activity. Retreat trended lower too.
Using just immersion and retreat, a simple linear regression model correctly classified hits 69% of the time – far better than chance.
This was a huge improvement over past studies that tried predicting hits from brain data.
Many focused only on attention, but combining attention and emotion here seemed key.
Machine Learning Boosts Accuracy
Next, researchers tested if machine learning could improve accuracy further by capturing non-linear relationships in the brain data.
They generated synthetic datasets based on the lab brain recordings to provide enough samples to train models.
An ensemble technique combined multiple algorithms to maximize predictive power.
This boosted hit classification to a remarkable 97% accuracy for hits and flops overall.
The brain features were most heavily weighted by a k-nearest neighbors algorithm.
This excels at handling the inherent nonlinearity in emotional neural signals.
Brain Data Predicts Hits in 1 Minute
Remarkably, brain activity from just the first minute of each song was enough to predict hits with 82% accuracy.
This demonstrates the potential value of rapid neural testing.
The approach could help music services quickly gauge initial audience engagement to new songs.
Then they can identify the most promising to promote.
Overcoming Biases in Self-Reports
This study demonstrates a better way to identify hit worthy songs than relying on self-reports.
We’re notoriously bad at predicting what content will become popular based on reflection.
Asking people if they “like” a new song they’ve never heard is especially biased.
Familiarity colors perceived liking.
But brain activity provides an unfiltered emotional reaction.
Machine learning can then detect subtle patterns linking neural signals to future popularity.
Emotionally Engaging = Share-worthy
The researchers speculate that highly engaging content activates our brains in a way that drives sharing and virality.
Catchy songs may literally get stuck in our heads.
So while conscious ratings fail, neural indicators of immersion and retreat may implicitly signal a song’s potential to spark contagious enthusiasm.
This could extend beyond music to predicting the success of movies, videos, social media posts, and more forms of viral content.
A New Tool for Content Creators
These findings offer an exciting new tool for content creators aiming to produce hits that resonate.
Neural testing could help identify the most engaging ideas to invest in.
For established artists, it could help select the best songs for albums and leading singles.
Marketers can also more efficiently target content likely to go viral.
Benefits for Streaming Services
Music streamers may benefit most.
Testing user brain activity while browsing new songs could improve recommendations and playlists.
By detecting high engagement in the first minute, promising new additions can be identified for personalized playlists and discovery features.
This neural approach could finally help streaming services overcome the major challenge of identifying the next hit among thousands of options.
Limitations and Next Steps
The study had a modest sample size of just 24 test songs.
The next step is validating results on a larger music database.
Researchers also plan to try this technique for predicting hits in other entertainment areas like video content.
The use of synthetic datasets for training machine learning models could overweight certain relationships in the data.
Directly testing larger real-world samples would be ideal.
Finally, it will be important to assess whether findings generalize across geographic regions, genres, and demographics.
But the study provides a highly promising proof of concept.
Harnessing Neural Technology
Thanks to advancing wearable sensor technology, it’s now practical to gather useful neural data outside the lab.
This could make neural hit prediction scalable for streaming services and content platforms.
As the technology improves further, the technique may even work using sensors in headphones or VR headsets.
Smart AI systems could then curate personalized, emotionally engaging content based on an individual’s neural feedback.
The ability to reliably predict hits would benefit consumers, artists, marketers, and distributors alike.
It could finally solve the enduring challenge of hit prediction that has puzzled experts for decades.
This study demonstrates the power of combining neural technology and machine learning to unlock valuable emotional insights straight from the human brain.
References
- Study: Accurately predicting hit songs using neurophysiology and machine learning
- Authors: Sean H Merritt et al. (2023)
