Phonemic restoration effect is a perceptual phenomenon where under certain conditions, sounds actually missing from a speech signal can be restored by the brain and may appear to be heard. The effect occurs when missing phonemes in an auditory signal are replaced with a noise that would have the physical properties to mask those phonemes, creating an ambiguity. In such ambiguity, the brain tends towards filling in absent phonemes. The effect can be so strong that some listeners may not even notice that there are phonemes missing. This effect is commonly observed in a conversation with heavy background noise, making it difficult to properly hear every phoneme being spoken. Different factors can change the strength of the effect, including how rich the context or linguistic cues are in speech, as well as the listener's state, such as their hearing status or age.

This effect is more important to humans than what was initially thought. Linguists have pointed out that at least the English language has many false starts and extraneous sounds. The phonemic restoration effect is the brain's way of resolving those imperfections in our speech. Without this effect interfering with our language processing, there would be a greater need for much more accurate speech signals and human speech could require much more precision. For experiments, white noise is necessary because it takes the place of these imperfections in speech. One of the most important factors in language is continuity and in turn intelligibility.

The phonemic restoration effect was first documented in a 1970 paper by Richard M. Warren entitled "Perceptual Restoration of Missing Speech Sounds". The purpose of the experiment was to give a reason to why in background of extraneous sounds, masked individual phonemes were still comprehensible.

In his initial experiments, Warren provided the sentence shown and first replaced the first 's' phoneme in legislatures with extraneous noise, in the form of a cough. In a small group of 20 subjects, 19 did not notice a missing phoneme and one person misidentified the missing phoneme. This indicated that in the absence of a phoneme, the brain filled in the missing phoneme, through top-down processing. This was a phenomenon that was somewhat known at the time, but no one was able to pinpoint why it was occurring or had labeled it. He again did the same experiment with the sentence:

He replaced the 'wh' sound in wheel and the same results were found. All people tested wrote down wheel. Warren later did much research for next several decades on the subject.

Since Warren, much research has been done to test the various aspects of the effect. These aspects include how many phonemes can be removed, what noise is played in replacement of the phoneme, and how different contexts alter the effect.

Neurally, the signs of interrupted or stopped speech can be suppressed in the thalamus and auditory cortex, possibly as a consequence of top-down processing by the auditory system. Key aspects of the speech signal itself are considered to be resolved somewhere in the interface between auditory and language-specific areas (an example is Wernicke's area), in order for the listener to determine what is being said. Normally, the latter is thought to be instantiated at the end stages of the language processing system, but for restorative processes, much remains unknown about whether the same stages are responsible for the ability to actually fill-in the missing phoneme.

Phonemic restoration is one of several phenomena demonstrating that prior, existing knowledge in the brain provides it with tools to attempt a guess at missing information, something in principle similar to an optical illusion. It is believed that humans and other vertebrates have evolved the ability to complete acoustic signals that are critical but communicated under naturally noisy conditions. For humans, while it is not fully known at what point in the processing hierarchy the phonemic restoration effect occurs, evidence points to dynamic restorative processes already occurring with basic modulations of sound set at natural articulation rates. Recent research using direct neurophysiological recordings from human epilepsy patients implanted with electrodes over auditory and language cortex has shown that the lateral superior temporal gyrus (STG; a core part of Wernicke's area) represents the missing sound that listeners perceive. This research also demonstrated that perception-related neural activity in the STG is modulated by left inferior frontal cortex, which contains signals that predict what sound listeners will report hearing up to about 300 milliseconds before the sound is even presented.