About the catalog

The catalog records what published descriptions say about word stress and tone for 6,080 languages, one entry per language, keyed by Glottolog code. Each entry gives the stress type, the stressed syllables at each word length the source states, the exceptions and domain of the rule, and the tone system where the source describes one. The catalog and its bibliography can be read entry by entry or downloaded from the catalog page.

A five-minute film, Where the Stress Falls, tells the story behind the catalog.

The entries were compiled by a large language model (Claude), which read the sources under written instructions, quoted the passage that decides each entry and coded it. A companion bibliography lists the works known on each language and what each is known to say about stress or tone.

Where the entries come from

The languages are the spoken languages that Glottolog (version 5.2) lists at language level, minus sign languages, pidgins, argots, ritual and artificial languages, and bookkeeping languoids: 7,803 languages, the frame. Candidate sources are every grammar, grammar sketch and phonology in Glottolog's bibliography and every other work there whose title names phonology, phonetics, sounds, syllables, tone, pitch, prosody, stress or accent, 54,586 works on 5,468 languages.

The entries come from readings made in August and September 2026 by language models of Anthropic's Claude family; the first few hundred, for languages that earlier databases code with fixed stress, were read against the sources those databases cite. The authors wrote the coding rules and the confidence scale; the models applied them.

Each entry rests on two readings. In the first, the model found an accessible source, searched it for the terms for stress, accent, prominence and tone, quoted the deciding passage verbatim with the page, and coded it. The instructions say: "Never invent a quotation or a page: every quote comes from text you opened." In the second, made with the first report in hand and under instruction to refute it, the model opened the work at the cited page, checked that the passage is there verbatim, read the surrounding section, confirmed or corrected the reading, and checked that the work describes this language. If the two readings disagreed, a third reading decided the case field by field.

What each entry records

An evidence grade says how the deciding passage was read: a page image (2,333 entries), the publisher's text (635), an open repository or author's copy (2,789), a search-engine snippet with the work unopened (244), or a quote carried from StressTyp2, WALS or a later work (105). The rule gives one digit per syllable at each word length the source states, so 10 010 0010 is penultimate stress in words of two to four syllables. Other fields cover secondary stress, exceptions, weight, domain, affixes and clitics, vowel length and tone, each with the source's own sentence. Where the source describes tone or pitch accent, the entry records the type of system, the number of tones if stated, and the passage and page (2,843 entries).

Every work an entry cites is one node, with editions and citation variants merged, and the confidence line counts those nodes: works by the same first author, or a work and its restatement, count once. Three of the five levels are set by rule: very confident needs the entry's own source confirmed in a copy we read and at least one other independent first-hand work stating the same rule, with no upheld contradiction and no better-graded known work unread; fairly confident is one confirmed first-hand source standing alone, or a source whose statement is itself second-hand or was read in a search excerpt; not confident means no statement of the pattern was confirmed in a copy we read. Moderately confident and disputed come only from a judged reading of the sources, which every entry with an upheld contradiction or a database coding that agrees only in part receives. Where the best-graded known work is unread, the entry says so and names it. When descriptions disagree, no compromise is coded: the dissenting source and its words are shown beside the preferred classification.

How the entries were checked

• The second reading corrected about one entry in three in some field, not necessarily the stress type. After compilation the catalog was read twice more against the cited sources.
• At 1,883 languages, every entry was re-read against its sources and the language's other descriptions: the quotation was confirmed for 1,754, a different classification was proposed for 171, and for 12 no source on stress appears to exist.
• Tone statements were checked twice. The first check found an error in 184 of 2,862 entries; after those corrections the second found 163 of 2,853, most in the three fields filled from the passage's own words (relation to stress, downstep, tone-bearing unit). Both sets of corrections are in this release.
• Where the reading differs from an earlier database's coding of the same language, the entry says so, for the user to weigh.
• Every open work not yet consulted for its language, 7,462 works, was read; works on frame languages are tagged as confirming an entry's coding (2,302), refining it (2,063) or contradicting it (386).
• Twenty-six fixed-stress entries have a pattern no other entry shares. Of 25 re-read, 14 are as their sources state, 1 is phrase-level, 3 are tendencies and 7 turn on loosely stated secondary stresses.

What has not been done

A catalog of this size is beyond what a linguist or a small group could compile by hand in any reasonable time, and not every entry has been checked by hand. Every entry was made to the same standard, and each shows its citation, the quoted passage and the reason for the ruling, so that readers can judge the interpretation for themselves and corrections are easy to make. All readings, reviews and confidence judgments were made by model readers. A work marked silent may treat prominence under a heading the search terms missed. For 300 entries the work was not opened and the quotation is a search-engine snippet; 149 carry a quote from a database or secondary work not yet found in the primary source. Digital copies behind a sign-in, a paywall or a dead link were cataloged but not read. Five rule strings identified for correction are kept as coded. A first-hand statement was found for 154 frame languages that have no entry yet.

Corrections are welcome; the contributing page says what helps and how to send it.

Sources and copyright

Every source was read from a copy we were entitled to read. Most were openly accessible online; the rest were read through university library holdings. The catalog quotes short passages with their citations, which is ordinary scholarly quotation, and it does not reproduce or distribute the works themselves. If you hold rights in a work quoted here and believe a quotation goes beyond fair use, write to comments@accstack.org and we will look at it.

Licence and reuse

The catalog data (codings, rule strings, confidence lines and the compilers' prose) and the bibliography records are released under CC BY 4.0: reuse them with attribution to AccStack. The quoted passages remain their authors' copyright and are quoted with attribution; the material carried from StressTyp2 keeps that database's own terms (CC BY-NC-ND).

How to cite

Schwartz, Matthew D., Kevin M. Ryan, Matthew K. Gordon, Joe Pater and Larry M. Hyman. 2026. AccStack: Agentically Compiled Catalog of Stress and Tone. [date accessed] https://accstack.org/catalog/ (release of 1 October 2026). The paper describing the catalog, AccStack: word stress and tone in the world's languages, with sources (Schwartz, Ryan, Gordon & Pater, 2026, preliminary), can be cited alongside it.