Delayed auditory feedback

Why does hearing your own voice a fraction of a second late stop you speaking?

Put on headphones, speak into a microphone, and have your own voice played back to you about a fifth of a second late. Within a sentence or two most people slow to a crawl, repeat syllables, stretch vowels, and get louder. Bernard Lee described this in 1950 and called it an artificial stutter.

Here is what should bother us. Nothing was taken away. Your tongue works, your lungs work, you still know the sentence. All that changed is the timing of information you already had. So before we ask what a delay does, let us ask something smaller: why should speaking depend on hearing yourself at all?

Start with the obvious test. Cover your ears, or stand in a very noisy room, and speak. You can still do it. So hearing is not the thing that drives the muscles. Something else is issuing the commands.

Then ask how fast speech actually is. Syllables come at roughly four to seven a second, and the individual articulatory gestures inside them are shorter still. Now suppose the loop really did run through your ears: sound leaves your mouth, reaches your ear, gets analysed, and a correction comes back down. Would that be quick enough to steer a gesture that is already over? Almost certainly not — the round trip is on the order of the gesture itself.

So the system cannot be waiting on what it hears. The mainstream account is that it runs ahead: the motor system issues a command and, at the same time, predicts what that command should sound like. Hearing is then used to check the prediction, not to generate the movement. A mismatch means something drifted — a cold, a mouthful of food, a room that swallows your voice — and the system nudges rate, loudness, or articulation to bring things back.

Notice what that makes the ear: not a driver but a comparator. And a comparator is only meaningful if the two things it compares belong to the same moment.

Now put the delay back in. At 200 milliseconds — roughly the length of a syllable, and about where disruption peaks in the classic studies — what arrives at your ear is a syllable you have already finished, and it lands while the next one is being produced. The predictor says one thing; the ear reports something else entirely. There is a large error, and it is real: the signals genuinely disagree.

What does the loop then do? It corrects. But which syllable does the correction land on? Not the one that generated the error — that one is gone. It lands on whatever you are saying now. And that new syllable, corrected on the basis of a stale complaint, comes back late as well, producing another error. Slow down, get louder, repeat the syllable, and every one of those responses feeds the mismatch again.

That is the whole mechanism, and it is the classic recipe for an unstable control loop: a real error signal, a nonzero correction gain, and a loop delay comparable to the timescale of the thing being controlled. The loop does not fail because it is weak. It fails because it is fast and confident about information that is out of date.

Does that predict anything we can check? It predicts the disruption should be worst near a syllable's length, not simply worse the longer the delay. That is what is observed: very short delays are harmless, and delays beyond roughly half a second are heard as a separate echo, something you can dismiss as another voice rather than as your own mouth misbehaving. And it predicts that masking your own voice with noise should help, because a loop with no signal at least issues no false corrections. It does.

One honesty note: exactly which computation breaks is still argued. The prediction-and-comparator story above is the standard one, but some researchers put more weight on a mismatching stream of speech simply capturing attention. And there is a genuine puzzle sitting beside it — for many people who stutter, delayed or otherwise altered feedback makes speech more fluent, which is why such devices exist. Why the same manipulation should break fluent speakers and help stuttering ones is not settled.

The tempting summary is "your brain gets confused by the echo". That is close, but it misplaces the problem. The system is not confused about what it hears; it hears your voice perfectly well. What it cannot do is stamp it with the right time.

The second correction worth making is about direction. It is easy to assume feedback drives speech and that damaging it removes the steering. But silence — masking noise, or a soundproof room — barely hurts, while a mis-timed signal is devastating. A wrong loop is worse than no loop. That is a general property of control systems and not a quirk of talking.

And this is why the effect is a laboratory tool rather than a curiosity. If a fifth of a second can reduce a fluent adult to syllable repetition, then whatever is being disturbed is doing real work in ordinary speech, continuously, below awareness.

• The Lombard effect: why we automatically raise our voice in noise, and what that says about the same loop.
• Why altered feedback often improves fluency for people who stutter.
• Delayed visual feedback in typing and drawing — does the same instability appear?
• Loop delay and gain in engineered control systems, where the instability condition is stated formally.

Every term the collection defines is gathered in the glossary.