Should a system always remove “um” and “uh”?

No. The decision depends on whether the output is verbatim, edited for readability, or used for another analysis.

Is a self-correction a speech-recognition error?

No. The speaker may genuinely have said both versions; the recognizer's task is separate from interpreting the repair.

Why is real-time detection difficult?

Later words can reveal that an earlier phrase was abandoned or replaced. A system processing only the current prefix may not yet have that evidence.

AI Glossary / Natural language and speech / Intermediate

Disfluency

Definition

A disfluency is an interruption, hesitation, repetition, or revision in the flow of speech. Examples include filled pauses such as “um,” unfinished words, and self-corrections. In speech-processing work, disfluencies are ordinary features of spontaneous language that a system may need to preserve, interpret, or remove according to the task.

What it means

People often plan and revise while speaking. A speaker might begin with one word, pause, and replace it with another. The resulting signal contains both the abandoned material and the intended continuation.
That creates a distinction between a faithful record of what was spoken and an edited version of the intended message. Both can be useful, but they should not be silently treated as the same artifact.
Disfluency does not by itself establish confusion, dishonesty, incompetence, or a medical condition. A speech application should interpret the utterance in context and avoid turning ordinary conversational patterns into unsupported judgments about the speaker.

How it works

For a self-repair, annotation often distinguishes the material being replaced, called the reparandum, an optional intervening expression such as “I mean,” and the repair that supplies the replacement.
A detection system can use words, acoustic cues, or their combination to identify these structures. In streaming settings, it may need later evidence before recognizing that an earlier phrase is being replaced.
The output policy depends on purpose. A verbatim transcript may retain the full sequence. A readable note may remove some hesitation material while preserving the final correction. In either case, the system must avoid deleting meaningful repetition or changing the speaker's intended content.

Three practical examples

A meeting-note tool interprets “Send it to the finance team - sorry, the design team” as a correction of the recipient.
A captioning workflow retains filled pauses when a faithful rendering of the spoken exchange is required.
A voice assistant delays committing to a slot value when the speaker begins a repair such as “Make that...” and then supplies a replacement.

Think of it this way

Think of speech as a document being drafted aloud, with some edits occurring before the audience has finished hearing the sentence. The analogy explains abandoned and replacement material. It is limited because spoken hesitations also serve interactional purposes, and they are not simply written typing mistakes rendered as sound.

Worked example

Consider this invented utterance: “Send the blue - uh, I mean green - folder to Maya.” A useful repair analysis identifies “blue” as the material being replaced, “uh, I mean” as intervening hesitation and editing material, and “green” as the repair.
A readable task note could become “Send the green folder to Maya.” A verbatim transcript would instead retain the original sequence, with suitable punctuation or annotation. These outputs serve different purposes.
Now imagine a system that acts as soon as it hears “Send the blue.” It has committed before the repair arrives and may select the wrong folder. A streaming interpretation needs a policy for provisional values and corrections; the example does not imply one universal waiting duration.
Finally, compare “It is very, very important.” Repetition here may be deliberate emphasis. Removing repeated words automatically would lose meaning. The correct handling depends on function and context, not merely on matching adjacent tokens.

Illustrative case study

Fictional case: NoteClear offers both raw transcripts and edited meeting notes. Its first cleanup rule removes all repeated words and every occurrence of “I mean.” Reviewers find cases where this hides a correction or weakens an emphasized warning.
The team replaces the blanket rule with task-specific annotation and review. Edited notes keep the final corrected recipient or value, while the raw transcript remains available for checking what was actually said.
Evaluation includes repairs, repetitions, fragments, and fluent emphasis. This fictional case illustrates a product distinction and a review method; it does not claim measured accuracy or suggest that automated cleanup can always recover a speaker's intention.

Common misconceptions and mistakes

Disfluencies are not all meaningless noise. They can reveal that a speaker is repairing or continuing an utterance.
Do not assume that every repetition is accidental or that every pause has the same function. Annotation definitions differ, and some studies focus on only a subset of phenomena.
Avoid evaluating a cleaned transcript against a verbatim reference without aligning the intended policy. Also distinguish a speaker's self-correction from an error introduced by speech recognition.

Business use

Disfluency handling affects captions, meeting notes, voice interfaces, and speech analytics. The central business decision is what kind of representation the user needs: a faithful record, a readable summary, or an executable interpretation of a request.
Teams should state that choice and test whether cleanup changes names, amounts, dates, recipients, or qualifications. A smoother-looking transcript is not necessarily more accurate. Keeping provenance between raw speech, transcript, and edited output helps reviewers diagnose where meaning was lost.

Check your understanding

A transcript reads “Book the large - sorry, small - room.” A cleanup system deletes only “sorry” and leaves both room sizes. Has it resolved the repair?

Answer: No. It has removed an editing cue without identifying which value was replaced. An edited interpretation should preserve the corrected choice, while a verbatim transcript can retain the full sequence.

Related terms

Filled pause
speech repair
reparandum
interregnum
repair
speech recognition
transcript cleanup
incremental processing

Learn next

Annotate several spoken corrections by marking abandoned material, optional editing phrases, and replacements. Produce both a verbatim and an edited version, then check whether names, quantities, and intended actions remain faithful to the audio.

Sources

Sources: Julian Hough and David Schlangen, Joint, Incremental Disfluency Detection and Utterance Segmentation from Speech: https://aclanthology.org/E17-1031.pdf ; Alan Bell and colleagues, Effects of disfluencies, predictability, and utterance position on word form variation in English conversation: https://web.stanford.edu/~jurafsky/jasa03.pdf Hough and Schlangen Section3.1 was independently read for the tripartite repair structure and the need for later evidence in incremental detection. The original blue-to-green example correctly preserves the final value; deliberate emphasis is separately distinguished.

Previous: Discriminator Next: Disjoint pattern databases