In 1970, Richard Warren replaced a single phoneme in a sentence with a cough. Nearly all subjects heard a complete word. The ones who noticed something odd couldn't say where in the sentence the gap was — not even approximately — even when told a phoneme had been replaced and asked to point to it.
The restoration works through two interacting mechanisms: top-down context prediction (what the sentence meaning expects at that position) and acoustic plausibility (whether the masking sound is loud enough that it could have hidden a phoneme). Both must cooperate. The output, when they do, is perceptually indistinguishable from a sentence with no gap.
What the simulation cannot show: whether the restored phoneme is phenomenally identical to a genuine percept, or whether there is a subtle experiential difference that subjects cannot access or report. The behavioral output — no locatable seam, complete word heard — is compatible with both.
The isolated-word preset shows what happens when sentence context is absent (the word is presented alone, or in a semantically unrelated sentence). Context drops to near-uniform; any phoneme is roughly equally plausible. Restoration still occurs if masking is sufficient, but with lower confidence, and the wrong phoneme may win.
The plausibility constraint is genuinely acoustic. A 2012 study found that in a reverberant room — where prior speech decays into silence, filling gaps with residual energy — silent gaps become more intelligible than noise-filled ones. The masking logic recalculates based on the full acoustic scene, not just the gap itself.
Warren's localization finding: subjects who detected any disruption still could not identify which phoneme was affected. The knowledge that a gap exists does not grant access to where it is. The generation erases its own location.