Scientists have trained a computer to analyze the brain activity of someone listening to music and, based only on those neuronal patterns, recreate the song.
The research, published on Tuesday, produced a recognizable, if muffled version of Pink Floyd’s 1979 song, “Another Brick in the Wall (Part 1).”
Before this, researchers had figured out how to use brain activity to reconstruct music with similar features to the song someone was listening to. Now, “you can actually listen to the brain and restore the music that person heard,” said Gerwin Schalk, a neuroscientist who directs a research lab in Shanghai and collected data for this study.
The researchers also found a spot in the brain’s temporal lobe that reacted when volunteers heard the 16th notes of the song’s guitar groove. They proposed that this particular area might be involved in our perception of rhythm.
The findings offer a first step toward creating more expressive devices to assist people who can’t speak. Over the past few years, scientists have made major breakthroughs in extracting words from the electrical signals produced by the brains of people with muscle paralysis when they attempt to speak.
But a significant amount of the information conveyed through speech comes from what linguists call “prosodic” elements, like tone — “the things that make us a lively speaker and not a robot,” Dr. Schalk said.
By better understanding how the brain metabolizes music, scientists hope to build new “speech prosthetics” for people with neurological diseases affecting their vocal production. The aim is for these devices to relay not only what someone is trying to say, but retain some of the musicality, rhythm and emotion of the organic speech.
To collect the data for the study, the researchers recorded from the brains of 29 epilepsy patients at Albany Medical Center in New York State from 2009 to 2015.
As part of their epilepsy treatment, the patients had a net of nail-like electrodes implanted in their brains. This created a rare opportunity for the neuroscientists to record from their brain activity while they listened to music.
The team chose the Pink Floyd song partly because older patients liked it. “If they said, ‘I can’t listen to this garbage,’” then the data would have been terrible, Dr. Schalk said. Plus, the song features 41 seconds of lyrics and two-and-a-half minutes of moody instrumentals, a combination that was useful for teasing out how the brain processes words versus melody.
Robert Knight, a neuroscientist at the University of California, Berkeley, and the leader of the team, asked one of his postdoctoral fellows, Ludovic Bellier, to try to use the data set to reconstruct the music “because he was in a band,” Dr. Knight said. The lab had already done similar work reconstructing words.
By analyzing data from every patient, Dr. Bellier identified what parts of the brain lit up during the song and what frequencies these areas were reacting to.
Much like how the resolution of an image depends on its number of pixels, the quality of an audio recording depends on the number of frequencies it can represent. To legibly reconstruct “Another Brick in the Wall,” the researchers used 128 frequency bands. That meant training 128 computer models, which collectively brought the song into focus.
The researchers then ran the output from four individual brains through the model. The resulting recreations were all recognizably the Pink Floyd song but had noticeable differences. Patient electrode placement probably explains most of the variance, the researchers said, but personal characteristics, like whether a person was a musician, also matter.
The data captured fine-grained patterns from individual clusters of brain cells. But the approach was also limited: Scientists could see brain activity only where doctors had placed electrodes to search for seizures. That’s part of why the recreated songs sound like they are being played underwater.
Other groups are doing similar experiments using noninvasive brain scanners, such as functional magnetic resonance imaging, or f.M.R.I., which gives a less detailed measure of activity but scans across the entire brain.
Yu Takagi, a neuroscientist at Osaka University, collaborated this year with scientists at Google to use f.M.R.I. data to identify the genre of music that a volunteer was listening to while in a brain scanner.
Dr. Takagi said the new study was significant because it showed that meaningful data could be collected from a relatively small number of neuronal clusters. “You don’t need that many electrodes to make something quality,” he said.
The new research also underscored what makes music different from speech. When the study volunteers heard a song, the right side of their brains tended to be more involved than the left, whereas the opposite happens when people hear plain speech. This finding, replicating previous research, helps explain why some stroke patients who can’t speak well can clearly sing sentences.
“It’s a technical tour de force,” said Robert Zatorre, a neuroscientist at McGill University whose lab established how the brain separates lyrics from music using brain scans. But to play back a song from someone’s head? “That’s a very interesting contribution,” he said.
