The Brain That Anticipates Sound

Imagine you're listening to a familiar melody, and one note suddenly disappears. Your brain still hears it. Not by magic, but because it had already anticipated it would be there. This phenomenon, far from being anecdotal, sits at the heart of one of the most influential theories in neuroscience today: predictive coding.

A Brain That Bets on the Future

For a long time, the ear and brain were thought of as a purely receptive system: sound comes in, the brain processes it. The reality is far more active. According to the predictive coding framework, the brain does not simply record what it hears. It constantly generates predictions about what it is about to hear, drawing on past experience and immediate context.

In practice, when a sound arrives, the brain compares it to its prediction. If everything matches, little of note happens. But if the sound deviates from what was expected, an error signal fires — what researchers call a "prediction error." It is precisely this error signal, not the sound itself in its entirety, that gets passed up to higher levels of auditory processing.

Neuroscience studies have captured this phenomenon in striking ways. When a listener is played a regular sequence of sounds and one is omitted, neural responses appear synchronized with the moment the sound should have occurred. The brain reacts to an absence as though it were a real event. It had predicted, and reality failed to deliver.

A Hierarchy of Predictions

What makes this theory particularly compelling is its layered organization. Researchers have shown that prediction errors are organized hierarchically along the auditory pathway: they are detectable as early as subcortical structures and grow stronger as signals move toward the auditory cortex. In other words, the brain does not predict at just one level. It predicts at every stage, from the lowest to the most abstract.

This architecture has profound implications for how we perceive music, speech, and any complex sound environment. It is not only the physical sound that determines what we hear, but also what the brain expects to hear. Context, culture, habit — all of these shape perception, upstream of the signal itself.

When Predictions Go Wrong: Tinnitus

Predictive coding also sheds new light on auditory conditions such as tinnitus. According to research published in the journal Brain in 2023, a brain deprived of acoustic input through hearing loss does not simply go quiet. It compensates by amplifying the internal noise of its own neural network in an attempt to restore the missing signal. That noise, reinterpreted as genuine sensory input, can be perceived as a persistent phantom sound.

Within the predictive coding framework, this mechanism can be understood as follows: when the bottom-up signal coming from the ear weakens, the brain over-weights its own internal predictions, and the resulting perception drifts away from acoustic reality. Tinnitus would then be a kind of uncorrected prediction — a sound the brain expects and can no longer disprove.

The Same Principle Inside Your Audio Files

What is fascinating is that this principle does not belong to neuroscience alone. It was applied independently, and much earlier, in the field of signal processing. As far back as the 1950s, engineers developed what is known as Differential Pulse-Code Modulation (DPCM): rather than recording each audio sample in full, the system predicts the value of the next sample from the previous ones, and stores only the difference between the prediction and reality.

The logic is exactly the same as in the brain: if I already have a good idea of what comes next, I do not need to transmit everything. I focus on what surprises.

This approach is the foundation of Linear Predictive Coding (LPC), which has underpinned decades of audio compression, from early telephony applications to modern codecs. FLAC, the widely used lossless audio format, is built directly on this principle: it predicts audio samples and encodes only the residuals — the gaps between prediction and reality.

Artificial Intelligence Takes It Further

Today, neural codecs carry this story forward. Systems such as LPCNet (Mozilla) and EnCodec (Meta) combine classical predictive coding with neural networks trained on vast audio datasets. The result is remarkable audio quality at very low bitrates — sometimes below 2 kbps for intelligible, natural-sounding speech.

These systems learn, in essence, to predict sound the way a trained brain would. They internalize statistical regularities so that only the essential gets transmitted: the unexpected, the prediction error.

The circle is complete. What engineers had intuited to compress data, and what neuroscientists later formalized to explain perception, obeys the same fundamental principle: to listen is, first of all, to anticipate.

Have you ever experienced that moment when you "hear" a missing note, or when an unexpected sound jolts you out of concentration? That is your predictive coding system firing an error signal. Share it in the comments — these small auditory surprises often reveal just how actively our brains construct the sound of the world around us.