AI Glossary / Natural language and speech / Intermediate

Filled pause

Definition

A filled pause is a vocalized hesitation, such as “uh” or “um” in English, occurring within the flow of speech. Unlike a silent pause, it contains an audible utterance. Filled pauses are common in spontaneous conversation and can contribute to how speakers manage planning, timing, and interaction.

What it means

People normally formulate much of their speech while speaking. They may pause to choose a word, plan the next phrase, or manage a response. An “um” can occur during that process without being part of the main factual content of the sentence.
Calling it a filled pause does not mean it is meaningless noise or evidence of poor ability. Research has examined its communicative functions as well as its relationship to speech planning.
Interpretation depends on context. A single hesitation does not reliably reveal a speaker’s confidence, honesty, competence, or emotional state. For language technology, the practical question is how to recognize and represent the audible event appropriately for the intended task.

How it works

In a recording, a filled pause occupies time and has acoustic properties like other vocal material. A transcription convention may represent it as an ordinary-looking token, a special label, or a timed annotation.
Its treatment depends on the goal. A verbatim transcript may retain “uh” and “um,” while a carefully edited reading copy may omit selected hesitations. An analysis of conversational timing should preserve their location rather than treating them as silence.
Clark and Fox Tree proposed that English “uh” and “um” can signal anticipated delays, with differences in the kinds of delay signaled. They prefer the term “fillers” because they analyze these items as words. That is a research account, not a rule for diagnosing the meaning of every individual occurrence.
Filled pauses should also be distinguished from repetitions, abandoned words, and later repairs, even though these phenomena can occur together.

Three practical examples

• A speech-recognition evaluation checks whether its verbatim reference and system output apply the same policy to “um.”
• A conversation researcher annotates filled pauses separately from silent intervals when studying turn timing.
• A transcript editor produces a readable version while preserving access to the original audio and unedited transcript.

Think of it this way

Think of an audible marker placed while a speaker prepares what comes next. It can help a listener interpret the flow, much as a visible pause in an action can signal that the action is continuing. The analogy is limited because speakers do not always consciously choose fillers, and their functions vary across contexts and languages.

Worked example

Consider an invented recording: “Please send, um, the revised drawing.” Suppose “um” lasts 0.3 seconds and is followed by 0.2 seconds without speech before “the.”
A timing annotation can represent one filled pause lasting 0.3 seconds and one silent pause lasting 0.2 seconds. Calling the whole 0.5 seconds silence would misrepresent the audible filler.
A verbatim word sequence might be “please send um the revised drawing,” containing six tokens under a simple whitespace convention. An edited sequence might be “please send the revised drawing,” containing five.
If a recognizer produces the edited sequence but is scored against the verbatim sequence using ordinary word edit distance, the missing “um” is one deletion. With no other errors, that gives 1 / 6, approximately 16.7% word error rate for this tiny example.
This does not prove the edited transcript is unsuitable for every use. It shows that transcription purpose and evaluation policy must agree. The duration measurements do not establish why the speaker hesitated.

Illustrative case study

In a fictional project, Rowan Interviews prepares transcripts for a public archive. One version is intended for close study of spoken interaction; another is a lightly edited reading copy.
The team keeps filled pauses in the research transcript and documents its editing policy for the reading copy. Reviewers check that deleting a hesitation does not accidentally remove nearby words or merge a self-correction into a different statement.
The imagined workflow recognizes two legitimate uses of the same recording. It does not claim that removing all fillers improves communication or that retaining every filler is necessary for every audience.

Common misconceptions and mistakes

A filled pause is not silence. It contains vocal material, even if it interrupts the main wording of an utterance.
It is not interchangeable with every disfluency. Repetition, word fragments, and repairs describe other phenomena.
Nor should a filler automatically be treated as a mistake that must be removed. The desired transcript or analysis determines what should be preserved.
Finally, counting “ums” does not provide a dependable standalone measure of truthfulness, intelligence, or speaking quality. Such conclusions go far beyond identifying the speech events themselves.

Business use

Understanding filled pauses helps teams set realistic expectations for speech recognition, transcription, dialogue systems, and audio review. A product should make clear whether it aims to reproduce spoken wording or provide an edited text.
Quality checks should compare outputs under matching conventions and inspect errors around hesitations and repairs. These regions can contain changes of wording that matter to the speaker’s intended message.
For conversational systems, timing analysis should distinguish ongoing vocalization from silence. That distinction can inform turn handling, but no single filler or pause duration should be treated as a universal signal that the user has finished.

People also ask

Are “uh” and “um” the same as a silent pause?

No. They are audible vocalizations, whereas silence contains no spoken filler.

Should an automatic transcript remove them?

That depends on whether the target is verbatim speech, an edited reading copy, or another form of annotation.

Do all languages use the same fillers?

No. Forms and usage vary, so an English-specific transcription or detection rule should not be assumed to work unchanged elsewhere.

Check your understanding

A speaker says “I need, um, the blue folder.” Does the presence of “um” prove that “blue” is a correction of an earlier word?

Answer: No. A filled pause can occur without a replacement or repair. The surrounding speech must show an actual correction before it is labeled that way.

Related terms

• Disfluency
• Endpointing
• Automatic Speech Recognition
• Word error rate
• Prosody
• silent pause
• speech repair
• spontaneous speech
• turn-taking
• verbatim transcription

Learn next

Learn the main types of speech disfluency, then compare a short recording with both verbatim and edited transcripts. Mark filled and silent pauses separately. Finally, examine how a chosen speech-recognition evaluation handles fillers so its score reflects the intended output convention.

• Disfluency

More in Natural language and speech

• Extraposition - Extraposition is a grammatical construction in which a clause or phrase appears later in a sentence than in a corresponding more basic arrangement.
• F-measure - F-measure is a family of evaluation scores that combine precision and recall into one number. The best-known member, F1, is their harmonic mean.
• Factoid questions - Factoid questions seek a specific factual answer that can usually be expressed briefly, such as a person’s name, a place, a date, or a quantity.
• Filtered-popping recursive transition network - A filtered-popping recursive transition network, abbreviated FPRTN, is a recursive transition network that permits a return from a subroutine only when the...
• Final fall - A final fall is a downward movement of pitch toward the end of an intonational phrase.
• Flap (phonetic) - A flap is a consonant produced with a very brief articulatory contact rather than a sustained closure or repeated vibration.

Sources

Clark and Fox Tree, “Using uh and um in spontaneous speaking” (2002), original author-hosted paper: https://web.stanford.edu/~clark/2000s/Clark%2C%20H.H.%20_%20Fox%20Tree%2C%20J.E.%20_Using%20uh%20and%20um%20in%20spontaneous%20speaking_%202002.pdf. Hough and Schlangen, original work on incremental disfluency detection: https://aclanthology.org/E17-1031.pdf. The distinction between edit terms and repair structure was consulted. Independent review checked Clark and Fox Tree’s original delay-signaling proposal and its contextual qualifications.

Previous: FIFO queue Next: Filter