Feature: Beat Detection — Phrase-Level Stress Shape Analysis

[craigtrim]

Overview

Beat in prose is the perceived rhythmic pulse when reading aloud — the pattern of stressed and unstressed syllables across a phrase or sentence. This is distinct from aggregate stress statistics (which we already have) and refers specifically to sequential stress shape across parallel units.

This is a standalone prosodic feature and also a hard blocker for tricolon validation (#73).

---

What We Already Have

prosody/rhythm_prosody.py contains substantial foundation:

Function	What it does
_get_stress_pattern(word)	Returns per-syllable stress values [0, 1, 2] from CMU dict
_compute_stress_entropy(words)	Shannon entropy of stress patterns across full text
_compute_metrical_feet(words)	Iambic / trochaic / dactylic / anapestic ratios across full text
_compute_rhythmic_regularity(syllable_counts)	CV-based regularity score
word_stress_patterns (metadata)	Per-word stress sequences stored at document level

The CMU wrapper in prosody/pronouncing.py gives us phones_for_word() returning ARPAbet strings with stress digits (0 = unstressed, 1 = primary, 2 = secondary). All the raw materials exist.

The gap: everything above produces aggregate statistics across a full document. Beat detection requires sequential stress analysis at the phrase or member level — not "what is the overall iambic ratio" but "given this specific sequence of words or phrases, what is its stress shape, and how does it compare to adjacent members."

---

What Needs to Be Built

A compute_beat(units: list[str]) -> BeatResult function that takes a list of text units (words, phrases, or sentences) and returns:

Per-unit analysis

• Stress sequence: [1, 0, 1, 0] for each unit
• Syllabic weight: total primary-stressed syllables per unit
• Beat string: human-readable form — "DUM-da", "da-DUM", "da-da-DUM"

Cross-unit analysis (the key new capability)

• Weight sequence across units: [2, 2, 4]
• CV of weights → isocolon (low CV) vs. climactic (rising) vs. anti-climactic (falling)
• Beat shape classification:

Shape	Definition	Example
Isocolon	All units roughly equal weight (CV < 0.2)	"da-da | da-da | da-da"
Climactic	Third unit heavier than first two	"da | da | da-da-DUM"
Anti-climactic	Third unit lighter	Rare, stylistically unusual
Irregular	No clear pattern	High CV, no monotonic trend

Why this matters for AI detection

LLMs default to isocolon — they generate parallel units of roughly equal syllabic weight because it's the safest/most-learned pattern. Human writers, especially skilled ones, tend toward climactic structure — the third member carries more weight as the rhetorical culmination. This asymmetry is a discriminating signal within tricolon, not just in its detection.

---

Implementation Path

1. Reuse _get_stress_pattern(word) from rhythm_prosody.py — already does the CMU lookup
2. Add phrase-level aggregation: sum stress values across all words in a unit to get unit weight
3. Add shape classifier: compare weights across units, classify as isocolon / climactic / anti-climactic / irregular
4. Add beat string renderer: translate stress sequences to DUM/da notation for human inspection

No new dependencies required. All infrastructure is in place.

---

Output Type

BeatResult:
  units: list[BeatUnit]        # per-unit analysis
    - text: str
    - stress_sequence: list[int]
    - syllabic_weight: int
    - beat_string: str          # e.g. "da-DUM-da"
  weight_sequence: list[int]   # across units
  weight_cv: float             # isocolon indicator
  beat_shape: str              # "isocolon" | "climactic" | "anti-climactic" | "irregular"
  cmu_coverage: float          # fraction of words found in CMU dict

---

Blockers

• Blocks Feature: Tricolon Detector #73 (Tricolon Detector) — tricolon candidate validation requires beat shape to distinguish strong / weak / rejected candidates and to detect isocolon vs. climactic patterns as an AI-tell discriminator
• No other blockers — can be built now against existing infrastructure

Related

• Feature: AI Stylistic Tell Detection — Tricolon, Stacked Sentences, and AI Vocabulary Fingerprints #69 — AI Stylistic Tell Detection
• Feature: Tricolon Detector #73 — Tricolon Detector (downstream consumer)
• prosody/rhythm_prosody.py — existing stress infrastructure
• prosody/pronouncing.py — CMU dict wrapper

[craigtrim]

Theoretical Basis for the Four-Class Beat Shape System

This comment documents the academic grounding for the four beat-shape classifications (isocolon, climactic, anti-climactic, irregular) and is explicit about which parts are well-attested in literature versus which are novel computational operationalizations proposed by this project.

---

Why Exactly Four Classes

The four classes form an exhaustive logical partition over a single scalar sequence (syllabic weights across parallel units):

Shape	Condition
isocolon	Weights are approximately equal (low variance)
climactic	Weights rise toward the final member
anti-climactic	Weights fall toward the final member
irregular	Neither equal, rising, nor falling — no coherent pattern

These four exhaust all possible orderings of a monotone or flat sequence plus the residual case. There is no fifth possibility that does not collapse into one of these four.

---

Class 1: Isocolon

Definition: Parallel cola (clauses or phrases) of approximately equal syllabic length or stress weight.

Primary classical source:

Rhetorica ad Herennium, IV.20.27 (anonymous, c. 86–82 BC). Trans. Harry Caplan. Loeb Classical Library 403. Harvard University Press, 1954.

"We call Isocolon the figure comprised of cola which consist of a virtually equal number of syllables."

This is the earliest surviving Latin text to use the word isocolon and to define it explicitly in terms of syllable equality. The Greek antecedent is parisosis in Aristotle's Rhetoric, III.9 (trans. W. Rhys Roberts; Loeb Classical Library 193, Harvard University Press, 1926), where it is defined as "making the two members of a period equal in length."

Modern scholarly treatments:

• Corbett, Edward P. J., and Robert J. Connors. Classical Rhetoric for the Modern Student. 4th ed. Oxford University Press, 1999. (ISBN 978-0-19-511542-0.) Isocolon treated under "Schemes of Construction."
• Lanham, Richard A. A Handlist of Rhetorical Terms. 2nd ed. University of California Press, 1991. (ISBN 978-0-520-07669-3.) Entry: isocolon.
• Burton, Gideon O. Silva Rhetoricae. Brigham Young University, 1996–2007. https://rhetoric.byu.edu/Figures/I/isocolon.htm. Cites Ad Herennium 4.20.27 directly.

Computational relevance:

• Kühn, Ramona, Jelena Mitrović, and Michael Granitzer. "Computational Approaches to the Detection of Lesser-Known Rhetorical Figures: A Systematic Survey and Research Challenges." ACM Computing Surveys 58, no. 2, Article 42 (2025). DOI: 10.1145/3744554. arXiv:2406.16674. Isocolon is among the 24 figures surveyed; no standalone NLP detection paper specifically for isocolon exists as of the survey date — this project's operationalization is therefore genuinely novel.

---

Class 2: Climactic (gradatio / κλῖμαξ)

Definition: Parallel members where each successive unit carries greater rhetorical or syllabic weight than the one before it, culminating in the final member.

Primary classical source:

Quintilian. Institutio Oratoria, IX.3.54. Trans. H. E. Butler. Loeb Classical Library 127. Harvard University Press, 1922.

"Gradation, which the Greeks call climax, necessitates a more obvious and less natural application of art and should therefore be more sparingly employed."

Quintilian's full treatment (IX.3.54–57) defines gradatio as a figure that "repeats what has already been said and, before passing to a new point, dwells on those which precede," producing a chain of ascending intensity. The canonical example is Julius Caesar's Veni, vidi, vici — three members in which the final carries the greatest rhetorical weight.

Modern scholarly treatments:

• Corbett & Connors (1999), as above. Gradatio / climax treated under "Schemes of Repetition."
• Lanham (1991), as above. Entry: gradatio, climax.
• Burton, Gideon O. Silva Rhetoricae. https://rhetoric.byu.edu/Figures/G/gradatio.htm.

---

Class 3: Anti-climactic (bathos / descending order)

Definition: Parallel members where the final unit carries less weight than those preceding it — the structural inversion of gradatio.

Honest statement on classical sourcing: Anti-climax as a positively named rhetorical figure does not appear in classical Greek or Latin rhetorical texts with the standing of isocolon or gradatio. Quintilian and Aristotle treat descending order implicitly as a failure or inversion of climax, not as a distinct named figure.

The term anticlimax enters the scholarly lexicon primarily through:

Pope, Alexander. Peri Bathous, or the Art of Sinking in Poetry. 1727. (A parody of Longinus's On the Sublime, coining "bathos" as the comic descent from the elevated to the trivial.)

Modern rhetorical handbooks treat anti-climax as a recognized structural figure:

• Lanham (1991), as above. Entry: anticlimax.
• Corbett & Connors (1999), as above.

For this project: Anti-climactic classification is well-grounded as the logical inverse of climax and is recognized in modern rhetorical scholarship. Its absence from classical texts as a named figure reflects the ancients' aesthetic prejudice against it (Quintilian remarks that descending order is "stylistically unusual"), not a gap in the conceptual system.

---

Class 4: Irregular

Definition: Stress weight across parallel units shows no coherent monotone or flat pattern; high variance with no directional trend.

Source: This class is a computational catch-all. No classical rhetorical figure corresponds to it directly — the ancients classified deviations by their specific failure mode rather than by a single "irregular" label. The Irregular class is proposed by this project as a necessary residual category to complete the exhaustive partition.

---

What Is Not from the Literature

The following elements of the proposed implementation are novel computational operationalizations not found in prior published work, and should be treated as empirical hypotheses to be validated:

1. CV < 0.2 as the isocolon threshold. The Rhetorica ad Herennium defines isocolon as cola of "virtually equal" syllable count; it gives no numeric threshold. The CV = 0.2 cutoff is asserted by this project and must be empirically validated against known examples.

2. Syllabic weight defined as count of primary-stressed syllables (CMU digit == 1). Tepperman and Narayanan (ICASSP 2005, DOI: 10.1109/ICASSP.2005.1415269) demonstrate automatic syllabic stress detection at 87–89% accuracy using prosodic features. This project applies a simplified version (CMU dict lookup) rather than acoustic modeling — appropriate for text-only analysis but not attested in prior stylometry literature as the specific operationalization of "syllabic weight."

---

Summary

Element	Status	Primary Source
Isocolon as a rhetorical category	Attested, 86 BC	Rhet. ad Her. IV.20.27
Climax / gradatio as a rhetorical category	Attested, c. 95 AD	Quintilian IX.3.54–57
Anti-climax as a rhetorical category	Attested, modern	Lanham (1991); Pope (1727)
Irregular as a residual category	Novel / catch-all	This project
CV < 0.2 threshold for isocolon	Novel, unvalidated	This project
Primary-stress count as syllabic weight	Novel operationalization	This project