Are boundary tones the same as lexical tones?

No. Lexical tone participates in distinguishing words in tonal languages; the boundary tones discussed here concern prosodic phrase edges.

Can punctuation supply the labels?

It can offer clues for speech generation, but written punctuation does not fully determine a recorded speaker's intonation.

Does every language use H% and L% identically?

No. Annotation systems need language-specific analysis.

AI Glossary / Natural language and speech / Advanced

Boundary tones

Definition

Boundary tones are tonal events associated with the edges of prosodic phrases in speech. In the English ToBI framework discussed here, H% and L% mark high and low boundary tones at the end of an intonational phrase. They help describe how an utterance is organized and interpreted beyond its sequence of words.

What it means

The same words can sound finished, continuing, or inviting a response depending partly on their ending. Boundary tones capture one component of that difference. They concern the organization of spoken language, rather than the dictionary meaning of a word or a punctuation mark in a transcript.
English ToBI separates boundary tones from pitch accents on prominent syllables and from phrase accents. Consequently, L-H% contains a low phrase accent, L-, followed by a high boundary tone, H%. Treating that whole combination as a single elementary tone loses useful information. Other languages and varieties require appropriate annotation conventions.

How it works

A speech analyst listens to a recording, examines its fundamental-frequency contour, and identifies prosodic phrasing. The analyst then assigns tonal labels using the chosen transcription system. Fundamental frequency describes vocal-fold vibration; measured values require interpretation in the speaker's pitch range and surrounding speech.
In the English analysis used here, an intermediate phrase has a phrase accent, such as L- or H-. An intonational phrase contains one or more intermediate phrases and ends with a boundary tone. Thus the final combination can be L-L%, L-H%, H-H%, or H-L%. These labels describe tonal structure; they are not a lookup table that infallibly converts pitch into intention.
For AI, an annotation model might predict these labels from audio, while a speech generator might use prosodic targets to shape an output contour. Neither task requires assuming that every sentence has only one phrase.

Three practical examples

A navigation voice uses phrasing to distinguish the first part of a multi-part instruction from its completion. Designers assess whether listeners wait for the next clause rather than acting prematurely.
A conversational assistant considers a user's final pitch pattern together with words, timing, and dialogue history when deciding whether a response is expected. Pitch alone does not settle that decision.
A language-learning tool displays two recordings of the same text and highlights differences near phrase endings. It explains the relevant variety and avoids presenting one intonation pattern as universally correct English.

Think of it this way

Think of boundary tones as part of the way a musician finishes a musical phrase. An ending can create closure or leave an expectation of continuation. The analogy concerns phrasing: speech tones are not fixed musical notes, and listeners interpret them together with language and context.

Worked example

Consider an original, hypothetical annotation exercise with four phrase endings. No recordings or measured pitch tracks are supplied; the labels are stipulated to demonstrate how the notation is read.
Ending A is L-L%. Ending B is L-H%. Ending C is H-H%. Ending D is H-L%.
Read the component before the percent-marked tone as the phrase accent. A and B share L-, whereas C and D share H-. Now read the boundary component. A and D share L%; B and C share H%.
There are therefore four combinations but only two boundary-tone categories in this exercise. A system reporting all four as different boundary-tone labels has mixed two annotation decisions into one label inventory. That combined inventory can be useful, provided it is named and evaluated as a combination.
Suppose a model predicts A correctly, predicts L-L% for B, predicts L-H% for C, and predicts H-L% for D. It gets three of four phrase accents correct: only C's phrase accent is wrong. It also gets three of four boundary tones correct: only B's boundary tone is wrong. Each component accuracy is 75%, but exact combination accuracy is only 2/4, or 50%.
This calculation shows why a report must identify its unit of evaluation. It establishes no performance claim about real speech, and it does not assign communicative intentions to these invented cases.

Illustrative case study

Fictional example: ParcelVoice tests a delivery assistant that sounds unfinished at the end of some notifications. Its team initially asks a model to classify four ending combinations but calls the result boundary-tone accuracy. Reviewers discover that phrase-accent errors and boundary-tone errors have been merged without explanation.
The team separates the two components in its evaluation, retains exact combination accuracy, and listens to complete messages in their dialogue contexts. It also tests whether users understand when a message has ended. The revised process connects annotation errors with the actual listening task; the fictional scenario does not claim a measured improvement or that one low ending always guarantees closure.

Common misconceptions and mistakes

A rising ending does not always mean a question, and a falling ending does not always mean a statement. Context and language variety matter. A boundary tone is also not equivalent to a pause: a recording can contain several cues to phrasing.
High and low labels should not be assigned using one fixed frequency threshold for every speaker. A pitch-tracking error can imitate a rise or fall. Finally, correct tonal labels do not demonstrate emotion recognition, sincerity detection, or knowledge of a speaker's private mental state.

Business use

Boundary-tone analysis can support voice-interface testing, spoken-content quality assurance, and pronunciation feedback. Define the target language variety, label inventory, and intended listening outcome before collecting evaluation data. Keep component accuracy, combination accuracy, and user-task results distinct. An improvement in one measure may leave the others unchanged, so assess complete utterances as well as isolated endings.

Check your understanding

In L-H%, which component is the boundary tone?

Answer: H%; L- is the phrase accent.

Why can both component accuracies be 75% while combination accuracy is 50%?

Answer: Their errors occur on different examples, leaving only two examples with both components correct.

Related terms

Prosody
Pitch accent
Fundamental frequency
Speech synthesis
intonation
phrase accent
dialogue systems
ToBI

Learn next

Learn to distinguish pitch accents from phrase-edge tones, then compare full recordings with their annotations. Next, study how prosodic phrasing interacts with turn-taking and how annotators resolve disagreements without treating written punctuation as ground truth.

Sources

Jurafsky and Martin, Speech and Language Processing, speech production and perception chapter: https://web.stanford.edu/~jurafsky/slp3/15.pdf Hirschberg, Communication and Prosody: Functional Aspects of Prosody, especially section 2.1: https://www.isca-archive.org/diapro_1999/hirschberg99_diapro.pdf Beckman and Ayers Elam, Guidelines for ToBI Labelling, scope and introductory guidance: https://www.ling.ohio-state.edu/research/phonetics/E_ToBI/

Previous: Boundary set Next: Bounded cost search