Researchers Trace The Outline Of Imagined Music Straight From The Brain

Brain Signals Can Reveal the Contour of an Imagined Melody, Study Finds

A song gets stuck in the head. No sound comes out, no lips move, yet somewhere in the brain’s electrical activity, that melody is playing. Researchers have now shown they can partially trace its outline. The result was not a recording of the music; the system predicted relative note categories and arranged them into sequences that often followed the original melody’s rise and fall.

Scientists have reconstructed that contour from brain activity recorded directly from the surface of the human brain, working not by detecting specific musical notes, but by tracking the relationships between them, the same mental shortcut humans often rely on when recognizing a tune.

Most people cannot identify a single piano note by its frequency alone, but almost anyone can recognize “Twinkle, Twinkle, Little Star” whether it is played high or low. That ability, recognizing how notes move relative to each other, is one of the main ways people recognize and remember melodies. A team of researchers from institutions in South Korea built their decoding system around that same principle, and published their findings in the journal eNeuro.

Patients Imagined Melodies on Cue While Wired for Brain Signals

Researchers worked with ten patients who already had electrodes placed on the surface of their brains for epilepsy treatment; no additional surgery was performed for research purposes. Of 17 patients initially recruited, three were excluded for difficulty understanding the task and four for not being able to accurately sing the trained melodies, a predefined criterion for task compliance. The final group included three males and seven females, with a mean age of 27.3 years.

Each participant completed eight sessions in a single day. During each session, they listened to two bars of a familiar children’s song, silently imagined the next two bars, then hummed what they had imagined. Songs included tunes like “Yankee Doodle” and “Twinkle, Twinkle, Little Star,” presented in three different musical keys so the brain was not simply tracking specific note frequencies. Visual cues kept mental time at 120 beats per minute, with each imagined note lasting half a second.

Participants were instructed to avoid any overt movement during the imagery phase, though researchers note subtle movement-related activity could not be entirely ruled out, since no electromyography recordings were made to confirm the phase was motion-free.

Brain recordings captured electrical activity across many frequencies at once. Researchers focused on high-frequency activity associated with active brain processing to identify which electrode sites were genuinely engaged during the imagery task. Only sites showing meaningfully elevated activity during imagination, compared to a quiet baseline, made it into the analysis.

Brain Signals Revealed Relative Pitch, Not Absolute Pitch

Rather than trying to identify the exact musical key of each note, researchers asked where each note falls relative to the others, like “Do, Re, Mi.” That approach proved far more effective. Across the group, computer models trained on this relative-pitch framework classified imagined notes more accurately than expected by chance. During an independent test session, average accuracy reached 26.7%, with the best individual subject hitting 31.2%. Chance level for a six-option problem was 16.7%.

Tested on absolute pitch instead, the same brain data performed no better than chance. That result fits the idea that the brain encodes and replays melodies in terms of note relationships rather than fixed frequencies, though this study offers evidence for that framework rather than a definitive or universal conclusion.

At the sequence level, results were genuinely telling. Even though individual note predictions were imperfect, stringing them together produced reconstructed sequences that matched the rise and fall of the original songs, tracking closely with the real melody’s shape across most of the notes. Put plainly, even when a single predicted note was wrong, the overall shape of the melody was often right.

Sound and Motor Regions Both Drove Accurate Decoding

Not all brain areas contributed equally. The most useful signals came from the superior temporal gyrus, a region on both sides of the brain associated with processing sound and language. A motor-related region on the left side also contributed meaningfully. Researchers suggest this pattern may reflect a kind of inner rehearsal, similar to how people sometimes mouth words silently while reading, though they present this as an interpretation rather than a confirmed mechanism.

Certain combinations of brain region and signal frequency were reliably useful across multiple subjects, even though electrode placement varied person to person based on individual medical needs. When researchers tested each participant using a model trained on everyone else, decoding performance began to surpass chance once motor-region signals were added. Researchers treat this cross-subject result as exploratory and preliminary, suggestive rather than evidence of a broadly shared decoding scheme.

Familiar Melodies, Not Novel Ones, Made Decoding Possible

This study describes itself as a proof of concept, and that framing matters. The songs used were well-practiced, not improvised or recalled from scratch. Memory and mental rehearsal likely shaped the brain signals being decoded, and researchers acknowledge as much. Extending this work to new melodies, broader pitch ranges, or real-time decoding would take substantially more development.

Earlier studies have picked up on pieces of imagined music, things like rhythm and rough sound texture, but reconstructing the shape of a melody note by note is new ground. The brain has been humming to itself all along. Science is now beginning to trace the outline.

---

---