Cocktail party effect

Why can your own name cut through a conversation you were paying no attention to at all?

You are deep in one conversation. Across the room, in a group you could not have summarised a word of, someone says your name — and you turn.

There is a contradiction sitting in that sentence. If you were not processing that conversation, nothing in it could have been recognised as your name. If you were processing it, in what sense were you not listening? Either you were doing more than you thought, or "not listening" means something narrower than switching off.

Before we can ask what gets through, we have to notice a harder problem that comes first. Your eardrum does not receive several voices. It receives one summed pressure wave — every talker, the music, the glasses, added into a single wiggling line. There are no conversations in that signal until the auditory system builds them.

So how are they built? From regularities that hold for a real source and not for a coincidence: sounds that start at the same instant probably came from one throat; harmonics sharing a common fundamental frequency belong together; the tiny differences in arrival time and level between your two ears say where a source stands. From these the system assembles streams. Only then is there anything to attend to. Selection does not pick sounds out of a mixture; it picks a stream out of a set the system has already carved.

Now the selection itself. The tidiest early proposal was a filter placed early, before meaning: choose a channel by its physical properties — that ear, that voice pitch, that direction — and discard the rest unanalysed. It explains a great deal. Play one message to each ear, ask someone to shadow the left, and they genuinely cannot report what the right ear said. They often cannot say what language it was in.

But that proposal makes a hard prediction: nothing meaningful can ever cross, because nothing on the other side was ever analysed for meaning. Which is exactly what your name at the party refutes. And it was tested deliberately — own name planted in the unattended channel is reported noticeably more often than an unremarkable control word planted in the same place. That asymmetry is the whole argument. If your name broke through no more often than a stranger's, the hard filter would survive and we would be left explaining occasional lapses of attention instead.

So the filter is not a switch. The better account is that the unattended stream is attenuated — weakened, but still passed on for analysis. Every word then has to clear a threshold before it can influence what you are aware of, and thresholds are not equal. Your name has been paired with everything that matters to you since infancy; its threshold is about as low as a word's can get. A weakened signal is still enough to clear it. "Fire" behaves similarly, and so does anything you happen to be primed for that day.

That is the mechanism, and it is worth saying plainly what remains unsettled. Only a minority of people show the effect in a controlled dichotic test, and the fraction moves around between studies enough that I will not quote one. More awkwardly, the finding does not by itself prove parallel semantic analysis, because an observer cannot see where attention was actually pointed in the instant the name arrived: a rival reading says attention had briefly drifted, and the name was caught by a channel that was momentarily the attended one. That account is weaker but not dead, and honest textbooks say so.

Two refinements hold the account together. First, the separation of streams and the selection among them are different jobs, and only the second is attention. This matters because it explains the situations where the party effect fails altogether: when the room is loud enough that the streams cannot be built, no amount of attending recovers a voice, which is a segregation failure rather than an attention failure.

Second, "attenuated" should not be read as "consciously heard but ignored". You genuinely do not experience the unattended stream as speech. What survives is enough activation to nudge word representations, not enough to construct a heard sentence — which is why breakthrough is a word, almost never a clause.

It is worth marking the boundary against two neighbouring puzzles. In inattentional blindness the unattended object produces no report at all, and the interesting fact is the total absence; here the interesting fact is that something specific escapes. And in binocular rivalry the loser is genuinely suppressed and the alternation is involuntary; here you choose the winner, and the loser is turned down rather than switched off.

• How the auditory system decides which harmonics belong to one voice, and why that fails first in noisy rooms.
• Why hearing aids and conferencing software still handle the segregation problem far worse than a listener does.

Every term the collection defines is gathered in the glossary.