Why speech can sound whole when a sound is missing

In a laboratory experiment reported in 1970, Richard M. Warren played students a recorded sentence about state governors and legislatures. A cough interrupted the recording. Nineteen of twenty listeners said every speech sound was present. None placed the cough at its correct position.

The recording disagreed with them. Warren had removed a short section containing the first “s” sound in “legislatures”, including nearby transitions, and inserted the cough. The missing sound was not quietly playing underneath it. When he used silence instead, the missing speech became apparent. Read Warren's original experiment.

That difference poses a useful question. Why should adding an interruption make a damaged sentence sound more complete? The answer concerns how hearing uses evidence, not a hidden repair to the recording.

A noise can conceal the absence

In ordinary conversation, a clatter can cover part of a voice while the speaker keeps talking. The ear receives a mixture. In a restoration experiment, researchers can instead delete the speech from that interval and leave only noise. These are physically different signals, even if they can sound surprisingly alike.

A speech sound that distinguishes words, such as the “s” in this experiment, is called a phoneme. Experiencing a missing one as present is phonemic restoration. The listener does not need to consciously announce a guess. The word itself can seem intact.

The interruption must work with the remaining speech, rather than simply occupy time. A noise capable of hiding the absent segment can support continuity; the surviving fragments must also provide enough information. In 1992, Bashford, Riener and Warren found that filling repeated speech gaps with noise could improve word intelligibility under suitable acoustic and contextual conditions. Noise supplied no deleted words, yet listeners understood more. Read the study's abstract.

Hearing a word is not the same as reporting one

Understanding a damaged word alone would not prove that its missing sound was heard. You can identify a familiar tune without hearing every note. A listener might also prefer to answer that a sentence was complete because its meaning was obvious.

Arthur Samuel addressed that problem in 1981 by comparing two kinds of recording: noise replacing a speech sound, and noise added over a sound that remained present. Listeners judged whether each word was intact. Their correct and incorrect answers let him separate difficulty distinguishing the signals from a tendency to choose one answer.

Restoration was stronger for real words than for pronounceable nonwords, and it depended on the relationship between the replacement noise and the missing sound. But sentence context affected reporting bias in that study. “Context helps” therefore does not automatically mean “context changes the sound you hear”. The task and measure matter. Read Samuel's methods and findings.

A later test asked whether restored sounds changed subsequent sound judgments. Samuel found that repeated words with a restored “b” or “d” shifted listeners' labeling of later ambiguous syllables. Control conditions without restoration did not produce the same adaptation. This supported an effect on sound processing, beyond a report that the word made sense. It still did not recover the exact deleted waveform. Read the 1997 experiment.

Looking for completion in the brain

Brain measurements offer another kind of evidence. In a study published in 2009, Antoine Shahin, Christopher Bishop and Lee Miller used functional MRI, which measures changes associated with blood oxygenation, while people judged interrupted words. They found different activity patterns associated with perceived continuity and with their proposed repair process. These results supported a network account, but the division into two mechanisms was an interpretation of the comparisons, not a direct photograph of repair. Read the original study.

In 2016, Matthew Leonard and colleagues recorded electrical activity directly from the brain surface in five participants. They used word pairs such as “faster” and “factor”, replacing the distinguishing sound with noise. On some repeated presentations, the same altered recording was heard as one word; on others, as the other.

Activity in auditory cortex, brain tissue involved in processing sound, resembled the speech sound the participant reported hearing. Earlier activity in left frontal cortex also predicted the later report. That is evidence that the missing sound's features were represented during listening, not merely named afterward. It does not by itself prove that frontal activity caused the restoration, or identify every necessary step. Read the direct-recording study.

More than the ears supplies evidence

Seeing the speaker can change the effect. Shahin and Miller compared interrupted words accompanied by matching mouth movements, reversed movements or a still mouth. Matching movements allowed perceived continuity across longer noise-filled interruptions. The visual information did not insert speech into the headphones; it supported the listener's interpretation of the incomplete sound. Read the audiovisual experiment.

Language experience matters too, but not as a simple ranking of listeners. Mako Ishida and Takayuki Arai compared native English speakers with Japanese speakers using English as a second language. In their similarity-rating task, native speakers showed a stronger real-word advantage over nonwords for replaced sounds; the non-native group did not show that same advantage. Both groups showed other restoration effects. This is a finding about particular sounds, groups and measurements, not proof that second-language listeners cannot fill gaps. Read their 2016 study.

Why more noise is not a listening strategy

The benefit depends on the signal reaching the ear. In a realistically simulated reverberant room, Nirmal Kumar Srinivasan and Pavel Zahorik found that periodically interrupted sentences were more intelligible with source-level silent gaps than with noise-filled gaps. Reverberation, the lingering sound produced by reflections, filled those nominally silent intervals with decaying speech energy.

Their comparisons suggested that this changed the usual advantage of noise. The “silent” source interval was no longer silent at the ear. A room can change the evidence before perception begins. The finding does not mean reverberation always helps, or that noise generally improves speech. Read the room-acoustics experiment.

The cough in Warren's recording therefore reveals both a strength and a limit. Hearing can combine surviving sound with learned patterns to produce a useful, continuous experience. But a complete-sounding word is not proof that every sound reached the ear. Perceptual completion is an inference constrained by evidence, not a physical reconstruction of what was removed.