8.5 Phonemic restoration revisited

We can now solve one of the puzzles from movement 1.

Consider the sentence “It was found that the eel was on the axle”, where the bracketed phoneme has been replaced by a cough or noise burst. Listeners hear “wheel”. In a different context (“It was found that the eel was on the orange”) the same masked phoneme is heard as “peel”. The brain has restored a phoneme it did not actually receive, based on context.

In Bayesian terms: the sensory input (a noisy phoneme) has a vague likelihood — many phonemes could have produced this acoustic input. The context provides a strong prior — given the rest of the sentence, certain phonemes are much more probable than others. The posterior follows the prior. The listener perceives “wheel” or “peel” with full clarity, even though the acoustic data alone could not distinguish them.

Phonemic restoration is one of the cleanest demonstrations of top-down perceptual inference in action. It is everywhere in everyday speech understanding — you fill in missing or noisy phonemes all the time, without noticing, because the prior overwhelms the noisy likelihood. Speech understanding in noisy environments is, in significant part, a Bayesian-inference performance.