Intonation

Speech Technologyalso: pitch contour

In one sentence

Intonation is the pattern of pitch across a phrase or sentence, the part of prosody that tells a listener whether they are hearing a question, a statement, or something not yet finished.

200 ms Widely cited conversational turn-taking baseline, 2026Last reviewed 31 July 2026

Definition

Intonation is the pattern of pitch across a phrase or sentence. It is the part of prosody that tells a listener whether they are hearing a question, a statement, or something still unfinished.

It carries one job that matters directly to voice agents: signaling when a turn is complete.

Intonation is a subset of prosody, specifically the pitch contour. It earns its own entry because it carries a function that voice agent design depends on: signaling turn completion.

What intonation does

  • Sentence type. A rising pitch at the end commonly signals a yes-or-no question in English, and a falling pitch signals a statement. These patterns are language-specific and do not transfer between languages.
  • Continuation. A sustained or slightly rising pitch in the middle of a sentence signals that more is coming. This is the acoustic cue people use to avoid interrupting each other.
  • Focus. Pitch prominence marks the important word.
  • Attitude. Sarcasm, doubt and enthusiasm each have their own characteristic contour.

Why it matters for turn-taking

  • Human turn-taking works partly by predicting completion from intonation. A falling contour predicts the speaker is about to finish.
  • Silence-based endpointing ignores this entirely, which is why it interrupts someone who pauses mid-sentence with their pitch still up.
  • Semantic turn detection that reads prosodic features is closer to how people actually manage turns.
  • A system that reads intonation can tell "my number is oh seven nine", said with continuation pitch, from "that's everything, thanks", said with a terminal fall.

On the output side

  • A question whose pitch fails to rise sounds like a statement, which leaves the listener unsure whether a reply is even wanted.
  • This is a common and specific synthesis failure, and it is worth testing on purpose: make the agent ask a question and listen to whether it sounds like one.

Language dependence

  • Intonation patterns vary substantially across languages, and question marking in particular differs.
  • A synthesis model trained on one language's patterns and applied to another produces speech that is intelligible and subtly wrong.

Common misconception

That a rising pitch always marks a question. English uses falling intonation for wh-questions such as "where is it", and rising intonation for statements in some dialects, so the mapping is more complicated than the rule most people were taught.

Why it matters commercially

A question that does not sound like a question is a specific, testable defect that makes an agent confusing to talk to. It is also easy to check in evaluation and rarely checked, so it survives into production more often than it should.

In voice specifically

On a screen a question mark does the whole job; the reader can see that a reply is wanted. In speech the same punctuation has to become a rising pitch, and if the synthesis flattens it, the visitor hears a statement and waits while the agent waits for an answer the visitor never realized was being asked for.

Where AsqVox fits

Turn detection quality is what a visitor actually experiences as the agent listening properly instead of cutting in, and intonation is one of the signals that makes reading a turn correctly possible.

Visual

The pitch line that tells you it is your turn

The pitch line that tells you it is your turn30relative pitch40relative pitch50relative pitch60relative pitch70relative pitch80relative pitch45relative pitchMy50relative pitchnumber47relative pitchis58relative pitchoh62relative pitchseven66relative pitchnineSource: Illustrative contour, not measured dataThe pitch stays up on "nine". Sustained means more is coming, so a silence timer that cuts in here is wrong.query types absorbed, widest and earliest firstSentence type: a rising end asks, a falling end tellsContinuation: sustained pitch means do not interrupt yetFocus: the pitch peak marks the important wordAttitude: doubt, sarcasm and enthusiasm each have a contour

The pitch line is how humans know when to speak.

The line traces the pitch of one spoken phrase, and it does not fall at the end, which is the cue that the speaker is still going. A silence-based endpointing timer, waiting 150 to 300 milliseconds of quiet, fires the same way whether the pitch was still up or had dropped, so it interrupts the number and responds to a finished sentence correctly by luck rather than by reading the turn. The pattern is language-specific: English wh-questions such as "where is it" actually fall, so a rising pitch does not always mark a question. The defect worth testing is the reverse: make the agent ask a question, and if it does not sound like one, the listener will not know a reply is wanted.

Statistics

Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.

Human conversational turn gaps average around 200 milliseconds, achieved by predicting completion rather than by measuring silence, and intonation is one of the predictive cues.

200 msindustry range

Widely cited conversational turn-taking baseline, 2026 - A reference point, not a target. It is evidence that people manage turns by prediction, and pitch contour is part of what they predict from.

Silence-based endpointing thresholds typically sit at 150 to 300 milliseconds and ignore pitch contour entirely.

150 to 300 msindustry range

Common production budgeting convention, 2026 - An engineering convention rather than a research finding. It is exactly the mechanism that misfires when someone pauses mid-sentence with their pitch still raised.

There is no published benchmark for intonation accuracy in speech synthesis, and no standard test for whether a synthesized question actually sounds like a question.

-no reliable figure

The check is manual and takes seconds: have the agent ask a question and listen. It is rarely done.

There is no public measurement of how much prosodic features improve turn detection accuracy over transcript-only approaches, despite this being an active area of implementation.

-no reliable figure

Vendors describe using prosody in turn detection; none publishes the gain it buys, so the benefit has to be judged by live demo rather than by a figure.

Examples

In practice

A voice agent asks confirmation questions that listeners systematically fail to answer, because the synthesized delivery uses statement intonation. Users hear "you would like Tuesday at three" as a statement and wait for the agent to carry on. Adding explicit question-intonation markup to the confirmation prompts clears the stall completely.

The everyday version

Intonation is the way a voice rises at the end of a question. It matters more than it sounds like it should, because if your agent asks a question without sounding like it is asking, your customer will not realize it wants an answer, and the two of them will sit there waiting for each other.

Usage

Who says it

  • Speech scientists and conversation designers.
  • It appears in synthesis-quality evaluation and in the design of turn detection.

Where it turns up

  • In SSML prosody and pitch control, and on voice evaluation checklists.

Common misuse

  • Applying the rising-pitch-means-question rule universally, when English wh-questions fall.
  • Assuming intonation patterns carry across languages, which they do not.

Questions people ask

What is intonation?

Intonation is the pattern of pitch across a phrase or sentence, the part of prosody that carries the pitch contour. It tells a listener whether they are hearing a question, a statement or something unfinished, and it signals when a speaker is about to finish, which is why it matters so much to how voice agents take turns. A rising pitch at the end commonly marks a yes-or-no question in English, and a falling pitch marks a statement.

Does a rising pitch always mean a question?

No. English uses a falling pitch for wh-questions such as "where is it", and some dialects use a rising pitch on statements, so the simple rule most people learned is incomplete. Intonation patterns are also language-specific and do not transfer, so a model trained on one language's question marking will sound subtly wrong applied to another.

Why does my voice agent ask questions that customers do not answer?

Most likely because the synthesized delivery uses statement intonation, so the question does not sound like one and the listener does not realize a reply is wanted. It is a specific and testable defect: have the agent ask a question and listen to whether the pitch rises. Adding explicit question-intonation markup to the prompts usually fixes it.

How does intonation affect turn-taking?

People predict when a speaker is finishing partly from the pitch contour: a falling pitch signals completion, a sustained or rising pitch signals more is coming. Silence-based endpointing ignores this and simply waits for a fixed gap of quiet, which is why it interrupts someone who pauses mid-sentence with their pitch still up. Turn detection that reads prosodic features is closer to how people actually manage turns.

Share this definition

Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.