1.6 Phonemic restoration

What is happening

Phonemic restoration was first reported by Richard Warren in 1970, in an experiment as elegant as it is unsettling. Warren took spoken sentences, replaced a single phoneme — typically a consonant — with a brief burst of noise (a “cough”), and asked listeners what they heard.

They heard the missing phoneme. Intact. They could not reliably say which phoneme had been replaced. The brain had filled it in.

The fill-in is context-dependent. The same noise-replaced syllable in the sentence “The wheel was on the axle” is heard as /w/ in “wheel”, and the same physical noise-replaced syllable in “The peel was on the orange” is heard as /p/ in “peel”. The acoustic input is identical; the perceived phoneme differs because the prior — the listener’s expectation about what is likely in this sentence — differs.

The mathematical content is again Bayesian. The likelihood of the noisy syllable under a phoneme hypothesis is vague (the noise could have masked many phonemes); the prior, conditioned on the rest of the sentence, is sharp (English makes some phonemes much more likely than others). The posterior follows the prior, and the listener perceives the high-prior phoneme with full clarity.

We will treat this in 9.5, where we apply the full predictive-coding machinery from earlier sections of movement 9.

The puzzle for now is that you did not, in the demonstration above, perceive a “blank” or a “cough” inside a sentence with one missing phoneme. You perceived an intact sentence. The brain interpolated, the brain committed, and the brain reported its commitment as if it were the data. That is exactly the thesis of the essay.