How do you know a song?

The question usually comes as a choice. Either the song is stored in your head like a recording, every note and word sitting there at once, or you do not store it at all and instead predict each next note from the ones before. The choice is less clean than it looks, and the way it fails is the interesting part.

Storage that can only be read forward

Consider what each answer would predict. A recording can be opened anywhere. If a song were stored that way, starting at the second verse would cost about as much as starting at the top. A pure predictor, something that produces the next note from what came before, can only be entered from the past. To reach the second verse it has to run through what leads there.

Now notice that the two collapse into each other. A store that can only be read forward, one part at a time, each part cued by the last, behaves exactly like a predictor: give it the past, and it hands you what comes next. And a predictor trained until it reproduces one song exactly has, in effect, stored that song. So "stored or predicted?" is the wrong question. The useful one is whether the song can be entered from anywhere, or only from what comes before.

The evidence says: from what comes before

In 1988 the psychologist Andrea Halpern ran four experiments on how people move through songs they know. In one, people saw a word from a familiar song's lyrics, then a second word, and said whether the two came from the same song. In another, harder task, they compared the pitches of the notes sung on two lyrics.¹ In both, the time to answer grew with the number of beats between the two lyrics in the actual song, and in some conditions it also grew with how far into the song the first lyric came.¹ Her conclusion was that song representations have "temporal-like characteristics."¹

People were not looking the lyrics up. They were, in some sense, playing their way from one to the other.

The chain does not stop at the end of a song. Anyone who has listened to an album many times knows the moment after a track ends, when the next one starts in your head before it starts in the room. Amber Leaver, Josef Rauschecker and colleagues used that moment. They played people the last 32 seconds of tracks from their favorite CDs, followed by silence, inside a brain scanner.² To be included, participants had to hum or sing the opening bars of the correct next track on at least 75 percent of trials.² During the silence, the scans showed strong activity in prefrontal and premotor areas, and the authors note how closely this resembles the learning of motor sequences, suggesting "common predictive mechanisms" in both.²

What links one track to the next is mostly the order in which someone heard them, over and over. The brain learned the order, and plays it forward unasked.

Why the middle is hard to reach

If knowing a song is knowing a chain, the everyday experiences fall into place. You restart from the beginning of a line because a line is where a link begins. You cannot sing a song backward with any fluency because the links run one way. A wrong word mid-song can derail everything after it, because the next link was cued by the word you were supposed to sing.

None of this requires a special faculty for music. It is what any one-way chain does.

What experts add

There is a group of people who cannot afford a chain alone: performers who play long pieces from memory in public, where a single broken link can stop the performance.

Roger Chaffin and Gabriela Imreh studied a concert pianist as she learned the final movement of Bach's Italian Concerto, recording her practice and her own account of the piece's structure and of the cues she meant to attend to while playing.³ They found that she used the formal structure of the piece as a retrieval scheme and her performance cues as retrieval cues, and that she practiced until retrieval from that conceptual memory was as fast and automatic as retrieval from motor and auditory memory.³

In other words, she built an index. The listener has the chain: hands and ears that know what comes next. The performer adds a map of the piece with entry points, so that if the chain breaks she can step back in at a marked place instead of starting over. Random access to music is not given. It is built, deliberately, by people who need it.

Where this breaks

The Halpern study is described here from its published summary, not its full text, and it measured lyrics and pitches, not whole performances.¹ The album study scanned nine people in its first experiment.² The pianist is one person.³ The pattern is consistent across them, but each is small.

A chain is also not the whole of what a listener knows. People recognize a song from a fragment heard in the middle, which suggests some access does not run from the start. Recognition and recall may work differently: hearing a part can find the song even when you could not have produced that part from memory.

And general expectation matters alongside the specific chain. Years of listening teach what usually comes next in any music, and that knowledge helps fill gaps in a particular song. The sources read here do not separate the two.

The answer

You know a song the way you know a path you walk every day: not as a map you can consult anywhere, but as a sequence in which each step shows you the next. That is storage and prediction at once. What you do not have, unless you build it the way performers do, is a way in from the middle.