How the tools decide.

This page describes the analysis used by Rhythm Reader and Rhythm Reader Pro. Other tools on this site use different methods.

Why this page exists

Paste a sentence into Rhythm Reader and it marks the beats. The marks look authoritative. Some rest on a named reference source. Others are engineering decisions that could reasonably have gone another way.

A reader deciding whether to trust a mark, and a researcher deciding whether to use the output as research data, both need to know which is which. The analysis keeps that distinction internally, tagging each step by how well founded it is, and this page carries the distinction out to you.

These labels describe the evidential status of a decision. The three tiers below describe when a decision is made. They are separate classifications: a single tier can contain decisions of differing evidential status.

Three tiers, in order

The analysis runs in three passes, each taking the previous one as input. Keeping them separate matters: a later pass can modify how information from an earlier pass is realized in context, and you can edit any tier without discarding the others. Two further decisions sit on top of the three, and are described after them.

Tier 1 — what the word brings

Text is first split into words, then each word is looked up in a pronouncing dictionary supplied to the engine when it loads. That dictionary is derived from the CMU Pronouncing Dictionary, a record of General American pronunciations maintained by Carnegie Mellon University. The entry gives the word’s phonemes, with a stress digit on each vowel phoneme — not each vowel letter — so the number of vowel phonemes supplies the syllable count. That much is reference-backed, in the sense that it faithfully reports what the dictionary says. The dictionary itself records pronunciations on which speakers sometimes disagree, so it is a reference, not a ground truth.

Two practical limits follow. The tools load the full dictionary as a separate file when the page opens; if that file cannot be reached they fall back to a small common-word subset, and an uncommon word then falls through to the rule-based fallback analysis rather than being looked up. The tool says so beneath the text when it is running on the subset. And the exact dictionary release in use is not recorded anywhere in the code or in the exports, which is a reproducibility gap rather than an accuracy one — see the note for researchers below.

The dictionary does not supply syllable boundaries. The tool estimates those with the Maximal Onset Principle, constrained by an explicit table of legal English onsets, plus an exception table where dictionary convention differs. Maximal onset is influential and widely used, but English syllabification has competing analyses and genuine exceptions, so the resulting division is a documented heuristic, not a uniquely correct analysis.

Two further steps are heuristic. Phonemic syllables must be mapped back onto spelling so a mark can sit on the page, and spelling does not divide the way sound does; the alignment rule is an approximation. Words absent from the dictionary get a fallback analysis driven by stress-relevant suffixes.

Words with two dictionary pronunciations get a further step. For a short, explicit list of stress-shifting heteronyms — PROduce the noun against proDUCE the verb — the tool reads the surrounding words for syntactic cues and chooses a variant. The gate is deliberately narrow. The list is fixed, and the rule is never widened into the general claim that nouns are first-stressed and verbs second-stressed, which is wrong often enough to matter. Outside that list, competing pronunciations are kept rather than resolved. The cues are a heuristic.

Each result also carries an internal confidence score, which Rhythm Reader Pro can display. The name overstates it. What the number records is how directly the tool reached the result — a single dictionary entry, competing entries that agree, competing entries that disagree, or a fallback analysis — not the probability that the displayed mark is correct. A word with one entry is not thereby more likely to be right; it is only less ambiguously derived. The code says the same thing in its own words: these are heuristic indications, not calibrated probabilities.

Tier 2 — the metrical template

Each word’s strong/weak pattern is then matched against the four recurring metrical patterns used on this site, described on the Rhythm page. Where a word carries secondary stress, both readings are kept — secondary treated as full, and secondary reduced — rather than one being chosen silently.

This tier is exploratory, tagged speculative in the code, and the reason is worth stating. Representing a word as a metrical template, and limiting the inventory to four feet, is a modelling decision made partly for teaching. A string such as weak–strong–weak is parsed across a foot boundary rather than given a fifth, word-sized foot of its own. That is defensible and it is not the only defensible choice.

Tier 3 — rhythm in context

Only now does the sentence matter. Words are grouped into phrases, and within each phrase the syllable stream is fitted to the four feet. Lexical stress supplies weighted preferences at this stage; it does not force every word to behave like a complete foot. A phrase may open or close on an incomplete foot — a single weak pickup, or an unmatched syllable at the end — because ordinary prose does not arrive in tidy metrical parcels. At this tier feet never cross a phrase boundary. These foot-fitting rules are modelling constraints used by the analyser, not established properties of prose rhythm.

Phrase boundaries come in two kinds, and the distinction matters more than it looks. Punctuation sets the outer boundary. Inside it, a heuristic proposes further boundaries: a new phrase opens at a function word that follows at least one content word, with a cap of four words on how long a phrase may run. These inner boundaries are editable candidates. They are a useful starting point, not a validated predictor of where a particular reader would break. The Phrasing page describes what those boundaries do to meaning.

Function words are demoted at this tier, which is why the and will come out weak in the BAND will PLAY. In a context-neutral reading, function words are often prosodically weak, though contrast can make them prominent — as the Emphasis and focus page shows with He CAN. That broad tendency is well supported (Selkirk, 1996); the particular word list and demotion rule used here are heuristics.

The demotion is graded rather than flat. Determiners resist taking a beat strongly; prepositions, particles and negation barely resist at all. Among content monosyllables, a noun resists losing its beat more than a finite verb does. That is what lets the MOUSE ran UP the CLOCK and do NOT go gentle come out right without special-casing either phrase — and it is why the engine can decide which member of a collision should yield rather than always demoting the left one. These weights were tuned by hand on development examples. They are engineering quantities, not measured psycholinguistic ones.

Two further pressures act at this tier. English avoids adjacent strong syllables (a clash) and long runs of weak ones (a lapse), so the fitting carries a cost for both, and a beat may be moved rather than deleted — the retraction that turns TennesSEE plus AIR into TENnessee AIR. The tendency is the rhythm rule, and it is reference-backed (Liberman & Prince, 1977). The specific costs, and the judgement of which member of a clash yields, are heuristic. Rhythm Reader Pro exposes them as switches, so the clash pressure can be raised, lowered, or removed entirely — the quickest way to see what it was doing.

Last, one beat in each phrase is marked as the strongest: the nuclear accent, which in a neutral reading falls on the last content word. It is recorded on a layer of its own, separate from the beats, so a syllable can carry phrase prominence without sitting on a regular metrical beat. The nuclear stress tendency is reference-backed (Chomsky & Halle, 1968; Liberman & Prince, 1977). That it should be applied at all to a sentence whose context is unknown is a heuristic, which is why the rule can be switched off.

After the tiers — meter, and texts the tools already know

Two further decisions sit on top of the three tiers.

The first is a meter pass. Once the beats are settled, the tool parses each punctuation-bounded phrase into feet a second time, looking for a pattern that prevails across the passage, and keeps near-equal alternative scansions rather than hiding them. This second parse works to the coarser, punctuation-bounded unit, so its feet can cross the inner boundaries the tool estimated at Tier 3. The two layers are easy to confuse, because both draw feet. Before either, the passage is classified as prose or as metrical verse — from the regularity found in it, unless you say which — and that verdict changes the costs used throughout. The same sentence can therefore be analysed differently depending on the company it keeps. This layer is exploratory.

The second is a table of known readings. A short list of familiar verse and of stimuli used in the 2026 study carries a stored conventional reading — ALL the king’s HORses and ALL the king’s MEN — and when text you paste matches one, that reading is offered and selected by default, with its meter handed to the meter pass. The intent is straightforward: for a rhyme most readers already share, the conventional scansion is the more useful starting point. The automatic analysis is still computed and remains available as an alternative reading, and the table is never generalized to unrelated prose. It is worth knowing about all the same, since for those particular texts the display begins from a stored reading rather than a derived one.

What it will not tell you

Three limits are worth stating outright, because they bound what the tools can be used for.

Emphasis is not modelled. This is the important one, and the code says so outright: contrastive stress is not modelled. The analysis has no representation of what the reader already knows, so it cannot know that Where is MY hamster? differs from WHERE is my hamster?. It assigns one context-free default analysis — which is not the same as a neutral one, since the default itself encodes choices. The nuclear accent described above is not an exception to this: it says where prominence falls by default position, not what the sentence is about. Everything on the Emphasis and focus page describes work the tools do not do, which is precisely why the marks are editable by hand.

Orthographic syllabification has no single right answer. Where to place a hyphen in spelling is a matter of convention, and dictionaries disagree with each other. The tools follow one convention. Disagreement with it is not necessarily an error in either direction.

A proposal is not a ground truth. Where more than one metrical parse fits nearly as well, the analysis keeps the alternatives rather than hiding them. A displayed pattern is the best available parse under these rules, not the reading. Every beat can be overridden, and the disagreement is often the interesting part.

For researchers

Several properties matter if you are considering using the output in a study. Every analysis records its source and a confidence score, so dictionary-backed marks can be separated from fallback ones and from your own edits. The three tiers can be edited independently — lexical stress, template choice, and phrase boundaries each have their own edit operations, and the analysis reflows afterwards rather than being discarded. Rhythm Reader Pro exports annotated plain text, four CSVs (syllable table, prosody profile, metrical lexicon, and a session log), and a full JSON document carrying the analyser’s document version, which dictionary was loaded, and the configuration in force.

With research mode on, the instrument also logs every edit as a process measure — what changed, on which tier, the latency since the previous action, the confidence of the value being overridden, and whether the edit reverses an earlier one — and reports a Prosodic Divergence Index summarising how far the current annotation has moved from the automatic default. That index is a proposed measure, tagged speculative in the code, and its per-tier components carry more information than the composite, since a single number cannot say which representation the reader disagreed with. Two cautions the instrument states itself: the latencies are click-level browser timings, informative about ordering and deliberation but not eye-tracking-grade; and logging is always visibly indicated, never covert.

What is not yet published is a versioned technical specification: the exact dictionary release, the tokenisation rules, the suffix and exception tables, the confidence table, the metrical fitting weights and tie-breaking, the clash and lapse costs, the role-graded resistance weights, the phrase-length cap, the full function-word list, the heteronym list and its syntactic cues, the prose/verse classifier and its thresholds, the contents of the known-readings table, and the export schema. The prose above is deliberately readable; it is not sufficient to reproduce an output.

The dictionary is the sharpest instance. Both the full dictionary and the common-word subset shipped alongside it are built from the CMU Pronouncing Dictionary, but the build recipe points at an unpinned branch, so two builds made on different days may not contain the same entries. The JSON export records which of the two was loaded and the analyser’s document version, but no dictionary release identifier and no analyser build identifier is written into the shipped files or into any export. Until those are recorded in the exports themselves, treat any figure taken from these tools as tied to the copy that produced it, and archive that copy alongside your data.

What does not yet exist is a validation study of the tools themselves. The published work behind this site tested the effects of typographically marked stress and rhythm on readers’ prosodic judgments and sensitivity (Gross, Winegard, & Plotkowski, 2018; Gross, Roldan, Kiessel, Esselink, & Dunlap, 2026). Whether this analyser’s output agrees with trained human coders is a separate question, and an open one. Treat the output as a well-documented proposal to be checked, not as a validated instrument.

Try it

Rhythm Reader shows the result of all three tiers with every syllable tappable. Rhythm Reader Pro exposes the tiers separately, shows phrase structure directly, lets you move boundaries to watch Tier 3 recompute, names the feet, marks the nuclear beat, and lets you switch the clash and nuclear rules off — which is the quickest way to see what each of them was contributing.

The most informative thing you can do is disagree with it. Where your inner voice and the analysis part company, one of you is responding to something the other is not — and on this site that gap is the subject, not a defect.

References and further reading

The five components are covered in Learn; the claim beneath them is Implicit prosody.