Can a closure contain voicing?

Yes, depending on the realization.

Is closure duration the same as voice onset time?

No. They use different starting landmarks.

Does every quiet interval indicate a stop?

No. Several acoustic and recording conditions can produce low energy.

AI Glossary / Natural language and speech / Advanced

Closure, stop

Definition

A stop closure is the interval during production of a stop consonant when the relevant articulators form a complete obstruction in the vocal tract. In an oral stop such as [p], the lips or tongue block the oral passage before the obstruction is released. Closure is one phase of the consonant, distinct from its release and any following aspiration.

What it means

In saying a sound such as [p], the speaker brings the lips together before opening them. That closed interval contributes to the sound's timing and acoustic pattern even though it may have much less energy than a neighboring vowel. The closure is therefore part of speech production, not simply an accidental empty space in a recording.
For speech technology, this distinction matters when labeling sounds, measuring timing, or deciding which quiet intervals to remove. A waveform region with little energy can belong to a consonant. Conversely, a quiet region alone does not prove that a stop closure occurred: pauses, recording conditions, and other events can also reduce the signal.
The term here concerns phonetics. It does not refer to closing a software process, mathematical closure, or a conversational instruction to stop.

How it works

Articulators move into contact to create an obstruction. The relevant place depends on the sound: lips for bilabial stops, for example, or the tongue against part of the mouth for other oral stops. When the obstruction is released, a brief acoustic burst may occur.
A closure is often visible as reduced energy compared with surrounding vowels, but it need not be complete acoustic silence. Vocal-fold vibration may continue during part or all of a voiced closure. Its acoustic evidence should therefore be inspected rather than replaced with a universal zero-amplitude rule.
An annotator considers the waveform, spectrogram, surrounding sounds, and a stated labeling convention. Closure onset, release, and voicing onset are separate landmarks. The duration from closure onset to release measures closure duration under the convention used here. Voice onset time instead compares the onset of voicing with release. These measurements answer different questions and should be stored separately.
Not every stop realization has an easily visible release burst. Where a boundary is uncertain or absent, an annotation system should represent that limitation rather than inventing a precise timestamp.

Three practical examples

A phonetics project measures how long a speaker maintains a stop closure, using consistent boundary definitions across recordings.
A speech-data team checks whether an automatic alignment places a consonant boundary at the closure onset or only at the later release burst.
An audio-processing team reviews quiet-region removal to ensure that it does not delete consonant closures and alter the timing of spoken words.

Think of it this way

Think of a stop closure as the held-closed phase of a small gate. Opening the gate is a separate event. Measuring how long it stays closed is different from measuring how long after opening another activity begins. The analogy describes timing; speech also involves voicing and changing acoustic resonances.

Worked example

Consider an original hypothetical annotation of an oral stop between surrounding speech sounds. The closure begins at 1.200 seconds, the release occurs at 1.280 seconds, and voicing for the following vowel begins at 1.315 seconds. These timestamps are invented for instruction, not measured from an attached recording.
Closure duration is 1.280−1.200=0.080 seconds, or 80 milliseconds. Voice onset time is 1.315−1.280=0.035 seconds, or positive 35 milliseconds. The interval from closure onset to vowel voicing onset is 115 milliseconds. Calling that entire interval the closure would overstate closure duration by 35 milliseconds under the stated convention.
At a sampling rate of 16,000 samples per second, the 80-millisecond closure corresponds to 1,280 sample intervals. The 35-millisecond delay corresponds to 560 sample intervals. Together they span 1,840 sample intervals. These conversions describe the recording's time grid; they do not imply that an annotator can identify every boundary to the precision of one sample.
Now consider a second hypothetical stop with closure onset at 2.000 seconds and release at 2.080 seconds. Suppose voicing starts during the closure at 2.050 seconds and continues through release. Closure duration is again 80 milliseconds, but voice onset time is 2.050−2.080=−0.030 seconds, or negative 30 milliseconds.
The two closures have equal durations, yet their voicing timing differs. This is why closure duration and voice onset time cannot be substituted for each other. The numbers alone also do not establish a language's phoneme label: interpretation depends on the language, context, and additional acoustic evidence.

Illustrative case study

Fictional example: ClearRead prepares speech recordings for an internal reading assistant. A preprocessing rule removes every low-energy interval longer than a chosen threshold. Reviewers hear unnatural joins in some words and discover that the rule has removed regions associated with stop closures.
The team compares processed clips with the originals and separates trimming outside an utterance from editing intervals inside speech. It also records uncertain boundaries for review instead of assuming that all low-energy regions have the same cause. This scenario illustrates a failure mode; it reports no measured improvement or recommended universal threshold.

Common misconceptions and mistakes

A closure is not the release burst itself. It is also not necessarily the whole duration assigned to a consonant in every annotation scheme. Different projects may include different phases in their segment labels, so the measurement definition must be explicit.
Another mistake is equating oral closure with absence of voicing. The oral articulators and vocal folds participate in different aspects of production. Finally, a tiny waveform or a dark spectrogram region is evidence to interpret in context, not a complete phonetic diagnosis by itself.

Business use

Closure-aware review can help teams assess speech alignment, pronunciation analysis, and audio preprocessing. Document which landmarks measurements use and preserve the original recording for comparison. When evaluating an automatic system, inspect boundary errors as well as average duration error. Two wrongly shifted boundaries can produce a plausible duration while locating the closure incorrectly.

Check your understanding

How long is the first closure?

Answer: Eighty milliseconds.

Why is the second voice onset time negative?

Answer: Voicing begins before release.

How many sample intervals represent 80 milliseconds at 16 kHz?

Answer: 1,280.

Related terms

Stop (consonant)
Spectrogram
Place of articulation
release burst
aspiration
voice onset time
voicing
phonetic alignment

Learn next

Mark closure onset, release, and voicing onset separately in a suitable annotated recording. Compare the resulting durations and note boundaries that cannot be located confidently.

Sources

Jurafsky and Martin, Speech and Language Processing, chapter 15, stop articulation and waveform interpretation: https://web.stanford.edu/~jurafsky/slp3/15.pdf Phon manual, Voice Onset Time report, independent release and voicing landmarks: https://www.phon.ca/phon-manual/report/acoustic_data_reports/vot_report.html

Previous: Closeness Next: Cloud analytics