Can one word be an utterance?

Yes. A word such as “Thanks” or “Yes” can function as a contribution in context.

Is every sentence a separate utterance?

Not under every segmentation scheme. A continuous contribution can contain several sentences, while a dataset can use different boundaries.

Does an utterance always have one intent?

No. A single contribution can perform several functions, such as acknowledging an answer and making a new request.

AI Glossary / Natural language and speech / Intermediate

Utterance

Definition

An utterance is a particular instance or segment of language produced in communication, commonly a stretch of speech by a speaker in context. It may be a complete sentence, several sentences, or a short expression such as “Yes.” Speech datasets and dialogue systems define utterance boundaries according to their own documented segmentation conventions.

What it means

A sentence is commonly discussed as a grammatical structure. An utterance concerns language as used on a particular occasion. The same sentence spoken twice produces two different utterance events, possibly with different intentions or reference points.
An utterance need not be grammatically complete. “Tomorrow morning,” in answer to a scheduling question, can communicate a useful response without being a full sentence.
The term is also used operationally in conversational systems for a user’s input or an annotated segment. A project should explain whether that unit follows pauses, speaker changes, dialogue functions, or another rule rather than assuming one universal definition.

How it works

For recorded speech, identify the speaker and the boundaries of the segment using the annotation scheme. Pauses and speaker changes can provide evidence, but the treatment of hesitation, overlap, and self-correction requires explicit rules.
Associate the segment with relevant information such as its transcript, timing, speaker label, or conversational context. A transcript records words but may omit delivery cues that matter to interpretation.
For dialogue understanding, interpret the utterance in relation to preceding contributions and the current task. Short expressions often depend on that context.
When using utterances as machine-learning examples, preserve the distinction between how the audio was segmented and what communicative function the segment performs. Those units do not always align one to one.

Three practical examples

A speech corpus stores a short recorded contribution with its transcript and start and end times.
A scheduling assistant interprets “Friday” as an answer to a date question by using the preceding exchange.
A conversation analyst distinguishes a brief acknowledgment from a longer contribution, even when both are meaningful utterances.

Think of it this way

Think of a contribution made at a particular moment in a conversation. Its words matter, but so do who said them and what they respond to. The analogy is limited because annotation projects can split a contribution into multiple segments, and an utterance does not always coincide with a whole conversational turn.

Worked example

Consider this invented exchange: Mira asks, “Which room should we use?” Theo replies, “The small one.” Mira then asks, “At nine?” and Theo says, “Yes.”
Treat each listed contribution as one utterance for this exercise. There are four utterances, but not four independent full-sentence descriptions. “The small one” depends on a known set of rooms, and “Yes” confirms the immediately preceding time proposal in this context.
If Theo had said “Yes” after a different question, the same word could communicate a different answer. The word sequence alone does not supply what is being accepted.
Now imagine a pause inside Mira’s first question. A pause-based audio segmentation system might split it, while the exercise’s contribution-based scheme keeps it together. That difference should be described as a segmentation choice, not silently treated as a change in the number of conversational intentions.

Illustrative case study

In a fictional project, Juniper Booking trains a dialogue assistant using short user inputs. Several examples contain only “That one,” but their intended selections depend on options shown in the preceding system message.
The team preserves the relevant conversation context and records how utterances were segmented. It also distinguishes a speaker’s correction from a new request when preparing evaluation cases.
The imagined case shows why isolated text snippets can be insufficient. It does not suggest every model needs the entire conversation history or that a particular silence threshold defines an utterance correctly for every task.

Common misconceptions and mistakes

An utterance is not necessarily a complete grammatical sentence.
It is not automatically the same as a turn, which concerns participation in the conversation, or a dialogue act, which describes communicative function.
A pause does not always mean the speaker has finished a thought or surrendered the turn.
The same words on different occasions can have different referents and functions.
A transcript without speaker, timing, or context can lose information important to interpreting an utterance, even when every transcribed word is correct.

Business use

Clear utterance definitions improve speech-data preparation, conversational analytics, and evaluation. Teams can compare results more meaningfully when they know what counts as one input example or one processed contribution.
Document segmentation rules, treatment of overlap, and whether context accompanies each item. Test fragmentary answers and corrections as well as polished full sentences. When reporting performance per utterance, explain the unit so a change in segmentation is not mistaken for a change in model quality. Preserve enough context for reviewers to understand ambiguous short responses.

Check your understanding

A user says “Tomorrow” after being asked “When should I remind you?” Is it unusable because it lacks a subject and verb?

Answer: No. It can be a meaningful utterance answering the question. The system still needs the relevant date context and any missing reminder details before acting appropriately.

Related terms

Learn next

Compare sentences, turns, utterances, and dialogue acts using a short conversation. Mark the segmentation convention before counting units. Then examine short answers, overlapping speech, pauses, and self-corrections to see why different research or product tasks may choose different boundaries.

Sources

Jurafsky and Martin, Speech and Language Processing, Chapter 26, Conversation and Its Structure: https://web.stanford.edu/~jurafsky/slp3/26.pdf Text Encoding Initiative, archived P5 definition of the utterance element: https://www.tei-c.org/Vault/P5/4.10.1/doc/tei-p5-doc/en/html/ref-u.html EAGLES spoken-language resources glossary, discussion of utterance terminology: https://www.lancaster.ac.uk/fass/projects/eagles/GLOSSAR1.htm Independent source review: Chapter 26’s conversation, endpointing and grounding passages and the EAGLES terminology discussion were read. The archived TEI URL failed; the official current utterance definition was consulted instead: https://www.tei-c.org/release/doc/tei-p5-doc/en/html/ref-u.html

Previous: Utility theory Next: V-cycle