AccStack catalog

This page describes what the catalog contains, how each entry was compiled from the descriptive literature and verified against it, and what the confidence levels mean. Counts refer to the data release of 13 October 2026, the release shown in the footer of every page.

1What an entry contains

An entry is one language together with what its published descriptions say about the position of word stress. The catalog has 6,107 entries: 6,107 for languages identified by a Glottolog code, one entry per language, and 0 listed under their name only because they could not be matched to a code. Where counts of languages are given below, entries without a code that share a name are counted once. Each entry gives:

• The language: its name, Glottolog code, family and area, with a link to its Glottolog page.
• Its stress type, one of: fixed stress; weight-sensitive; weight-sensitive (tentative); lexical; set by morphology; variable; tone or pitch accent; no word stress; phrase-level only; depends on segments; sources conflict (unresolved); not determinable; other (see entry).
• The rule, for fixed stress: the stressed syllable for each word length, written left to right with 1 for a stressed syllable and 0 for an unstressed one, so 10 100 1000 is initial stress in words of two, three and four syllables; where the source states secondary stresses they are marked in the same strings. 1,649 entries have such a rule string. For weight-sensitive languages the entry gives instead what counts as a heavy syllable, the edge of the word the rule counts from, and what happens when no syllable is heavy.
• The sources and the quotation: the passage of the description on which the classification rests, quoted verbatim with its page (and an English translation where the source is in another language); the form in which that source was consulted (page scan, open full text, publisher's text, search excerpt, or as quoted in another work); every work consulted for the entry; and any disagreement between descriptions and any correction to a database coding, stated in full.
• The confidence level of the classification (section 9); links to recordings (243 entries include a short audio clip, 4,175 more link to recordings in external archives); and, for 995 entries, a note on the kind of evidence the description gives for stress and on the position of affixes relative to the root; and, for 2,873 entries whose sources describe tone or pitch accent, the tone system as the description states it (section 7).

Each entry further codes, each item with its own quotation and citation where the source gives one: secondary stress, separately from the primary rule, and any later work that revises the primary rule; the type of evidence in the primary source (impressionistic transcription, a described elicitation, instrumental measurement, or a restatement of another work) and whether it presents its own data; for weight-sensitive languages what counts as heavy, whether vowel quality and geminates enter the criterion, and the stress window; for tone, lexical and morphological systems the type of prominence; the behavior of affixes and clitics; the exceptional forms the source cites; what the source says about vowel length; the variety described; and how many of the works listed for the language have been examined for a statement on stress.

2Where the languages come from

The catalog began with the fixed-stress languages of two published databases: StressTyp2 (Goedemans, Heinz & van der Hulst 2015) and the appendix of Gordon (2002). Their 394 entries (345 from StressTyp2, 49 from Gordon) were each checked against the description the database cites. WALS chapters 14A–17A (Goedemans & van der Hulst 2013) were used to identify weight-sensitive languages and as an independent coding for comparison: each entry notes whether the WALS coding agrees with the source (155 agree, 117 partly, 17 disagree) and likewise for StressTyp2 (110, 192, 28).

All other entries were compiled directly from the descriptive literature. The frame is the 7,803 spoken languages of Glottolog 5.2: its language-level entries, excluding sign languages, pidgins, artificial languages, speech registers, and Glottolog's bookkeeping and unattested entries. The aim is a complete bibliography of the phonological sources for every one of them, with every accessible source examined and the stress pattern or prosodic type (stress, tone, pitch accent) entered wherever a source allows it to be determined. Of the catalog's 6,107 languages with a Glottolog code, 6,045 fall inside the frame; the other 62 are artificial languages, pidgins, speech registers and one entry that Glottolog lists as unattested: each is listed and described as an entry, marked on its card as outside the typological frame, and enters no typological count. Sign languages have no entries, since the prominence that their descriptions call stress is made with the hands rather than the voice, although their sources remain in the bibliography; constructed languages such as Esperanto keep their entries, outside the frame, and the catalog table counts and lists them only when "include artificial languages" is switched on. The bibliography page lists 8,600 Glottolog languoids, the whole frame among them.

By stress type the 6,107 entries are: tone or pitch accent 1,932; fixed stress 1,604; not determinable 560; lexical 521; set by morphology 440; weight-sensitive (tentative) 401; weight-sensitive 336; no word stress 134; sources conflict (unresolved) 49; phrase-level only 48; depends on segments 46; variable 36. Fixed stress is one class; what each source says about exceptions to the pattern is coded separately (section 5).

3How entries are compiled and verified

Entries were compiled from the sources by large language models (Claude) working under the fixed protocol set out here, not by linguists. Every entry was coded and then verified independently against the same copy of the source; no entry rests on a single pass over the source. The 394 database entries were treated in the same way, starting from the description the database cites; where that description does not support the database's coding, the entry follows the description and states the correction in full (333 entries: 118 from StressTyp2, 28 from Gordon).

Coding

For each language an accessible description is located and the passage that states the stress pattern is quoted verbatim, up to 1,500 characters, with its page; a copy of the work is kept so that the quotation can be verified in that same copy. The language is then classified and the fields of the entry are filled in from the passage: the rule, what counts as heavy, the domain the source describes (the word, the phrase, or both without distinction), the exceptions the source states, and the variety the source describes.

Independent verification

A second, independent pass is given the quotation and the classification with the task of refuting them. It checks, in the same copy of the source, that the quotation stands at the cited page word for word; that the classification follows from what the passage says; that the source describes this language and not a neighbor or a differently named variety; and that the passage concerns the word rather than the phrase. Its outcome is one of three: confirmed, corrected (with the fields changed), or rejected, in which case the entry is compiled afresh.

Adjudication

Where verification did not simply confirm the coding, every disputed field was settled by a third examination of the source, with the reason recorded. The published entry is the coding as confirmed, or as settled by adjudication.

Re-verification of the whole catalog

After initial compilation the entire catalog was re-verified twice more against the sources. In the first re-verification every field of every entry was compared with the cited passage and corrections were proposed field by field; each proposed correction was itself independently checked at the cited page, and only corrections that survived that check were applied. In the second, every entry's cited source was coded afresh without sight of the existing entry and the two codings were compared; every disagreement was independently checked at the page and, for classifications and rule strings, adjudicated once more before a change was made. In the same re-verification secondary stress was coded separately from the primary rule for every entry, the exceptions each source states were classified (section 5), the entries classified as undetermined, conflicting or variable were re-examined for a decidable type, and for every language the remaining fields of the entry (stress domain, affixes and clitics, exceptional forms, vowel length, tone, variety) were completed from the sources where they allow it. For all but 12 entries the cited passage was located again in a copy of the work during re-verification.

Duplicates and conflicting descriptions

Where a source turns out to describe a variety that already has an entry, or something other than a distinct language, no separate entry is made. Where the descriptions of one language disagree, the entry follows the description that is better supported and quotes the disagreement in full (469 entries carry such a statement); if none can be preferred, the language is classified "sources conflict (unresolved)" (49 entries). Conflicting statements are never averaged into a classification. An entry whose only source is weak (a Wikipedia article, or an excerpt seen without the full text) carries a caution (282 entries).

4What is quoted and what is not

5,816 entries quote the passage on which the classification rests and 6,103 give its page or section. Quotations are verbatim, in the language of the source; 226 entries add an English translation, marked as ours. Each entry states the form in which the cited source was consulted: a page scan (2,333 entries), an open full text (2,789), the publisher's text (635), a search excerpt only (244), or the passage as quoted in another work (StressTyp2's own passages, WALS, or a secondary work: 105). For the database entries and some others (928 entries in all) the passages are given work by work under "Source passages"; 203 of them cite more than one work.

A quotation counts as first-hand only when it was taken from a copy of the work itself and independently verified in that copy at the cited page. Where no copy of the original could be obtained, the entry is coded from the passage as quoted in another work and marked second-hand. StressTyp2 is licensed CC BY-NC-ND; it is cited for its codings and its references, and its own text is not reproduced. In the public version of the catalog quotations are trimmed to the relevant sentences. Corrections and takedown requests: open an issue at github.com/accstack/issues or write to comments@accstack.org.

5Fixed stress and exceptions

Earlier releases of the catalog distinguished two sets of fixed-stress languages: a strict set (a single quantity-insensitive rule for the word, stated without exceptions; 290 languages) and a dominant set (Gordon's (2002) broader criterion: the stated pattern is the predominant one; 489 languages). The catalog now codes, for each of its 1,604 fixed-stress entries (1,581 in the typological frame), the kinds of exception its source states, and leaves the choice of criterion to the user. The exception classes are: qualification only (words such as usually or generally with no exception named), loanwords or names, compounds or reduplications, phrase or utterance conditioning, secondary stress only, native lexical items, morphological classes or affixes, clitics or particles, and other conditions (weight, free variation, dialect). On the catalog page the filter Fixed stress: exceptions allowed keeps a language in the selection only if every exception its source states is of a kind the user allows. Three presets are provided, counted over the languages in the typological frame (section 2): exceptionless (absolute), no exception or qualification of any kind (596 languages, 60 distinct rule strings); exceptionless (core vocabulary), exceptions confined to loanwords, compounds, phrase-level conditioning or secondary stress, or a qualification only, and no native lexical, morphological, clitic or other exception (937 languages, 80 rule strings); and dominant pattern (Gordon 2002), any quantity-insensitive positional rule stated as the predominant pattern (1,581 languages, 111 rule strings). An entry for which verification finds no quantity-insensitive positional rule at all, but a weight-sensitive, lexical or morphological system or none that can be determined, falls under no preset and says why; in this release every fixed-stress entry states a positional rule. The two sets of the previous edition remain available as a single filter, "Previous edition (18 Sep 2026)".

Among fixed-stress entries the sources state: no exception of any kind for 609; a qualification only for 131; loanwords or names 199; compounds or reduplications 131; phrase or utterance conditioning 154; secondary stress only 108; native lexical items 475; morphological classes or affixes 134; clitics or particles 57; other conditions 308.

The database codings were checked against their own sources, beginning with the 43 languages whose codings mattered most for the typological counts drawn from the catalog. For 17 the database coding did not match the source (a rule projected to word lengths the source never describes, a secondary-stress pattern the source contradicts on the cited page, a wrong syllable), for 18 the coding depends on which of two published descriptions one follows, and for 8 it was confirmed. Where the two databases both cover a language, their independent codings agree for 177 of 193. Each correction was then applied individually, with the page that decides it; 333 entries carry a database correction stated in full (118 on entries from StressTyp2, 28 on entries from Gordon 2002). A database coding that departs from the description it cites is recorded as a correction of the database, under "Note on earlier databases"; only a published description that disagrees with another published description counts as a dissenting description (469 entries).

Membership of the fixed-stress sets is decided on the sources alone. All 632 earlier decisions to exclude a language or a coding from these sets, or from the populations behind a published count, were reviewed against the written criteria; four languages that had been excluded on theoretical rather than documentary grounds (Dehu, Yimas, Southern Paiute, Timucua) were reinstated, and every exclusion that rested on the sources themselves was upheld.

The 394 fixed-stress entries from the databases were also checked for the domain of the prominence their sources describe: 362 describe the word, 18 state one rule for word and phrase together, 7 describe the phrase, and 7 give a position without stating the domain. A language whose source states the rule for word and phrase together is listed as a dominant pattern, not as an exceptionless one, and its entry says so.

6Weight-sensitive languages

Weight-sensitive languages were identified from StressTyp2, from WALS and from the descriptive literature, and re-examined in their sources. An entry for a weight-sensitive language states, in the source's own terms, what counts as a heavy syllable (a long vowel, a closed syllable, both, a full as against a reduced vowel, or something else), the edge of the word the rule counts from, where stress falls when no syllable is heavy, whether the domain is the word, and what the source says about exceptions. Two classifications are used: weight-sensitive where the source states the rule without qualification (336 entries) and weight-sensitive (tentative) where the source states it as a tendency, with qualifications, or with exceptions it does not list (401 entries).

For 200 entries the stated rule was also tested against the example words the source itself gives in its stress section (typically 20 words, up to 164): the rule was applied to each word with its syllable weights as marked by the source, and weights and stress were independently re-derived from the source. For 76 entries every example fits the rule, for 75 most do (at least 80 percent), for 46 the examples contradict the rule as stated, and 3 give no examples. The result is shown on the entry as "Examples check" and can be filtered in the catalog.

7Tone

For the 2,873 entries whose classification or sources involve tone or pitch accent (tone languages, languages with tone beside stress or with a pitch accent, and languages whose descriptions raise tone without settling it), the work that describes the language's tone was read for what it states about the tone system: the kind of system (level tones, contour tones, both, a restricted or marginal contrast, a pitch accent, or tone beside stress); the number of contrastive tones and the names the source gives them; the unit that bears tone; whether tone belongs to each syllable, to the stem or root, or to the word as a melody; whether the source states downstep or floating tones; what it says about the relation between tone and stress; and the kind of evidence it offers (instrumental measurements, transcription, a described elicitation, tonal alternations, or the word of another work). The statement was taken from the primary source cited in the entry where that source treats tone, otherwise from the work among those consulted that does, and is quoted with its page, with a translation where the source is not in English (704 entries). The quotation and every item were then checked independently at the cited page and confirmed, corrected or struck; only items that stood are shown, and an item the source does not state is marked "not stated" rather than inferred. In this release 2,306 of these entries state a number of tones (1,001 two, 531 three, 570 four or more), 1,347 state the tone-bearing unit, 481 state downstep and 548 floating tones, and 476 rest on instrumental measurements; by system, level tones 1,054; tone beside stress 501; level and contour tones 480; not determinable from the source 261; no tone according to the source 228; pitch accent 199; a restricted or marginal contrast 139; contour tones only 11. The tone fields are shown on the entry under "Tone" and can be filtered in the catalog; they report what the descriptions say and are not a tonal analysis, and they leave the classification of the stress system unchanged.

A second reading covered the languages whose stress sources say nothing about tone. For every catalogued language whose entry recorded no statement on tone, and for every language of the frame without an entry, the works on tone found for the bibliography (section 8) that could be opened were read for a statement on the tone system of that language; the statement was quoted, coded as above and checked independently at the cited page, and a reader who judged that the statement concerns a group of languages or another variety rather than this language said so, in which case it is listed but not used. Where the statement stands and concerns this language, an existing entry gains the tone block and its note on tone follows what the work states, including an explicit statement that the language has no tone (182 entries; 111 of them state that there is no tone); a language of the frame that had no entry because no statement on its stress was available enters the catalog as a new entry when the work states its tone system or states that it has none (299 new entries: 272 with a tone system, 27 stating no tone), classified "tone or pitch accent" or, where the same work states a rule for stress that was verified, by that rule, and otherwise "not determinable" for stress. Statements read as concerning a group or another variety (101) and readings that leave the system open (53) add no entry.

8The bibliography

The bibliography page lists, for 8,600 languages, every work that describes or could describe the language's word stress or tone: 151,702 works in all, from Glottolog 5.1's bibliography for each language (grammars, sketches, phonologies, and any work whose title names phonology, phonetics, sounds, syllables, tone, pitch, prosody, stress or accent), from the sources cited in the catalog, and from works found during compilation. For each work the page notes how a copy can be obtained: online with open access (88,922 works) or restricted access (22,689); in print, with a research-library copy located (8,815); in print, with no library copy located (9,989); or not yet located (30,417).

Every work with an accessible copy was examined for a statement about stress; a statement, where present, was quoted with its page, related to the language's entry as agreeing with it, adding detail (weight, secondary stress, exceptions) or conflicting with it, and independently verified at the cited page. In the current release 4,278 works examined agree with the catalog's classification, 4,944 add detail and 543 conflict with it; 4,939 works supplied a stress statement for a language's entry, 11,108 mention stress without stating a rule, and 46,186 say nothing about stress. A further 29,410 works are likely to describe stress but have not yet been obtained. Conflicting statements are shown as such on both pages; none has been averaged into a classification.

The bibliography also lists works on tone. For every language of the frame the descriptive literature was searched for works that describe its tone system or state that it has none: 22,686 works were added for 4,983 languages (14,103 not previously listed, 8,954 already listed and now examined for tone), 8,571 of them with an open-access copy. Where an open copy could be read, the passage that states the tone system was quoted with its page and verified independently at the cited page: 9,093 works state the tone system of their language (4,481 with the quotation verified), 1,753 mention tone without a quotable statement, 2,285 were examined and say nothing about tone, and 9,926 have not yet been obtained. Each language's page groups these works under "Describes the tone system" and marks them "states tone"; a work read for tone only does not count as examined for stress.

9The confidence levels

Each language with a Glottolog code is rated on a five-level scale:

very confident

several independent first-hand descriptions agree

fairly confident

one first-hand description, or minor disagreement with a clear resolution

moderately confident

descriptions disagree; one analysis is better supported

disputed

specialists disagree in print

not confident

no first-hand statement has been confirmed; the classification is provisional

The level is derived from the evidence in the entry and in the bibliography: whether the entry's own quotation is first-hand and verified, and how many independent works agree with, add detail to, or conflict with the classification. Works count as independent only when their first authors differ (by surname, so that a fieldwork team publishing under one name counts once) and neither cites the other for its stress facts: a second work by the author of the entry's source, or a work that restates it, is named in the entry ("further works by the same author agree") but does not add to the count.

• Very confident applies when the entry's own verified first-hand source and at least one other independent first-hand work support the classification, and no work examined conflicts with it.
• Fairly confident applies when a single verified first-hand source stands alone; when several sources support the classification and the only conflicting statements could not themselves be verified; or when other works support an entry whose own quotation could not be verified first-hand.
• Moderately confident and disputed are assigned case by case rather than by rule: any language with a verified conflicting statement, a noted conflict between its sources, or a classification of "sources conflict", "not determinable" or "variable" is rated by weighing its sources against the five definitions, and that rating is independently reviewed; where rating and review differ, the lower level is published.
• Not confident applies when the entry has no verified first-hand statement: the quotation is second-hand, the source could not be obtained or is unclear on the point, or no usable statement about word stress exists.

In the current release the entries stand at: very confident 1,107; fairly confident 3,943; moderately confident 385; disputed 84; not confident 588; not yet rated 0 (entries added since the last rating, and the entries without a Glottolog code). Each entry's "Why" line gives the reason for its rating. By default the catalog shows the very confident and fairly confident languages; the other levels can be selected in the Confidence filter.

10Limitations

All coding, verification and adjudication described above was carried out by large language models (Claude) following this protocol; no entry has been reviewed by a linguist, and verification means one model run checking another against the same copy of the source. The catalog therefore reports what the published descriptions say, with the page cited so that every statement can be checked, and makes no independent judgment about the languages themselves. The kinds of error this method is known to make are:

• a rule projected to word lengths the source never describes, or read into a source that only states a tendency;
• a coding inherited from a database that its own cited passage contradicts;
• word-level and phrase-level prominence stated as one rule and coded as word stress;
• a description matched to the wrong language or variety (known cases are flagged under "Which variety");
• a quotation taken from another work or from a search excerpt that the original does not bear out;
• errors in scanned text, wrong page numbers, or errors in our translations;
• in the bibliography, the same work listed twice because two lists cite it differently, and access classes not yet verified.

To report a mistake, open an issue at github.com/accstack/issues or write to comments@accstack.org, naming the language and, if possible, the description and page that settle the matter. Corrections that change a classification quote the source, and the superseded classification remains visible on the entry.

11Versions and data

The catalog and the bibliography are generated from a single data release; the footer of each page shows the release date. On both pages "Download this list as CSV" saves the languages currently selected, with quotations, citations, conflicting statements and confidence levels; the bibliography's CSV includes, for each language, a link to its entry.

StressTyp2 (Goedemans, Heinz & van der Hulst 2015; CC BY-NC-ND) is cited for its codings and references and its text is not reproduced. Gordon (2002), A factorial typology of quantity insensitive stress, Natural Language and Linguistic Theory 20: 491–552, is cited for its appendix. WALS chapters 14A–17A: Goedemans & van der Hulst (2013), in Dryer & Haspelmath (eds.), The World Atlas of Language Structures Online. Works are listed from Glottolog 5.1's bibliography. Quotations from descriptions are short extracts from the cited pages, reproduced with full attribution so that each classification can be checked. The audio clips come from openly licensed collections and are redistributed under their own licenses with attribution, listed clip by clip in the public repository.

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).

AccStack catalog, data release of 13 October 2026. Maintained by Matthew Schwartz, Kevin Ryan, Matthew Gordon, Joe Pater and Larry Hyman. Corrections and takedown requests: open an issue at github.com/accstack/issues or write to comments@accstack.org.