The short version
- Using electrodes on the brain surface in 10 epilepsy patients, researchers decoded imagined melodies at about 26% per-note accuracy, versus 16% by chance.
- The decoder recovered the rise-and-fall shape of familiar songs but could not identify absolute pitch above chance.
- The method needs surgical implants and familiar tunes, so it tells us how imagining music works, not how to read minds.
A study published October 7 in eNeuro by Jii Kwon, Chun Kee Chung and colleagues at Seoul National University used electrocorticography, electrode arrays laid directly on the brain surface, in 10 patients with intractable epilepsy who already had the implants for surgical planning. Participants heard children’s songs in three keys, silently imagined the next section, then hummed it aloud. A machine-learning decoder trained on high-gamma activity (60–150 Hz) from the superior temporal gyrus and left precentral gyrus predicted relative pitch (do, re, mi and so on) in half-second segments.
The numbers are modest: about 26% of single notes correct against a 16% chance level, and absolute pitch could not be decoded above chance. What did come through was shape. PsyPost reports the reconstructed sequences “generally matched the ascending and descending patterns of the original melodies.” The authors list the limits themselves: low per-note accuracy, only highly familiar songs, invasive electrodes that rule out consumer devices, and no test of novel or improvised melodies.
The finding matters for a different reason than the headlines suggest. It adds direct, high-resolution evidence that imagining music recruits the same auditory and motor regions as hearing it, the same networks that music therapy for stroke and dementia tries to reach. The brain’s “inner ear” is measurable, just not readable.
Ten patients, familiar nursery songs and a one-in-four hit rate is where the field stands.
