Jany-Latyn: A Latin Orthography and Comparative Testbed for the Kyrgyz Language

A design proposal, reference implementation, and open testbed, in which every spelling is computed from Cyrillic by rule

Version: 2026.4 Status: Public proposal (v1 of the orthography) and living testbed Author: Emil Madraimov · salam@qylym.com Date: September 2026 License: This document is licensed CC BY 4.0. The reference engine is MIT; the web testbed is source-available under PolyForm Noncommercial 1.0.0. See LICENSING.md in the repository. Cite as: Emil Madraimov (2026). Jany-Latyn: A Latin Orthography and Comparative Testbed for the Kyrgyz Language, version 2026.4. https://doi.org/10.5281/zenodo.22844476

Abstract

Jany-Latyn is a 28-letter Latin orthography for Kyrgyz, published with a deterministic converter and a testbed in which every contested choice is a selectable option. Its governing constraint is that every spelling be computable from Cyrillic by rule, without a loanword lexicon, so that the text of the past eighty-six years converts unattended and the converter's output is itself correct Jany-Latyn. That constraint fixes the letters on which the others depend. Kyrgyz Cyrillic writes one к and one г, and splitting either into a velar and a uvular letter (k/q, g/ğ) would require knowing which words are loans, so Jany-Latyn keeps a single k and g. The vowel then carries the back–front contrast alone, so ы is written y rather than dotless ı, which loses its distinction from i under casing and in ASCII; with y taken, the glide й is written í. The remaining mappings are ж→j, ч→ç, ш→ş, х→h, ң→ŋ, ө→ö, and ү→ü; vowel length is written by doubling, and the Russian hard and soft signs are absorbed into the glide, so that every word is a single alphabetic token. Two registers are specified together, a full form carrying the diacritics and a plain ASCII form for unmodified keyboards. The document also fixes what machines need and orthography proposals usually omit: a reverse conversion whose losses are stated rather than discovered, one search-folding key across both forms, and an alphabetical order with its collation tailoring.

About this document

This is a design proposal for version 1 of the orthography, not a research report and not a standard. It argues for a set of orthographic choices, states the reasoning behind each, and names the places where the reasoning could be wrong. The orthography has not been used at scale, taught to anyone, or reviewed as a state standard would be. There is no evaluation section: claims about tokenizer behavior, reading speed, keyboard throughput, font rendering, and search argue from documented system behavior and from the author's experience, and §23 lists the ones that could be measured and what result would count against each. No claim here should be read as an empirical finding unless a source is cited for it. Readers who want numbers should treat this document as the hypothesis and the testbed as the apparatus, which exists so that these claims can be falsified by people who disagree with them.

Table of Contents

• Part I — Jany-Latyn in Brief
  • 1. The Problem, and What Jany-Latyn Offers
  • 2. The Alphabet at a Glance
  • 3. The Rule Behind It: Design Principles
  • 4. The Letters That Shape the Rest
  • 5. What It Gives, and What It Costs
  • 6. How It Compares
  • 7. The Testbed
  • About the Author
• Part II — The Case for Each Letter
  • 8. One k and One g: Synharmonic Allophony
  • 9. y for ы: The Back Unrounded Vowel
  • 10. í for й: The Palatal Glide
  • 11. j for ж: The Voiced Affricate
  • 12. ö and ü: The Front Rounded Vowels
  • 13. ç and ş: The Sibilants
  • 14. ŋ for ң: The Velar Nasal
  • 15. Vowels, Length, and е
  • 16. Loan Signs ъ and ь, and я, ё, ю
  • 17. Alphabetical Order
  • 18. Converting Back to Cyrillic
  • 19. Two Registers: The Full Form and the Plain Form
  • 20. Letters Under Evaluation: ä, x, w
• Part III — Context
  • 21. How Jany-Latyn Relates to Existing Systems
  • 22. Sociolinguistic and Political Context
• Part IV — Open Questions and Governance
  • 23. Open Questions and Future Work
  • 24. Governance
  • 25. Conclusion
• References
  • Further reading
• Appendices
  • Appendix A. The e System
  • Appendix B. Normalization, Folding, and Collation
  • Appendix C. Collation Tailoring
  • Appendix D. Engine, Keyboard Layouts, and CLI

Part I — Jany-Latyn in Brief

Jany-Latyn is a proposed Latin alphabet for Kyrgyz: 28 letters, designed so that a computer can convert any Cyrillic Kyrgyz text into it automatically, with no list of loanwords and no one checking the result. What the converter writes is simply correct Jany-Latyn. The name is itself written in the plain form described below; in full it is Jaŋy Latyn, "New Latin".

That rule decides the two letters that shape the rest. Kyrgyz Cyrillic has one к and one г, so Jany-Latyn keeps one k and one g, with no q or ğ: splitting them requires knowing which words are borrowed (kosmos, not qosmos). With one k, only the vowel tells кы from ки, so ы needs a letter that cannot blur into i: y, not the dotless ı, which loses its difference from i in capitals, in small print, and in plain typing. With y taken, й gets its own letter: í, an i with a mark, as й is an и with one.

The rest is familiar: j for ж, as in Japarov and Jalal-Abad; ö, ü, ç, and ş, as in Turkish, Azerbaijani, Turkmen, and the Common Turkic Alphabet; ŋ, shaped like ң; and long vowels doubled, as in Cyrillic. Drop the marks and the result is the plain form, which any keyboard can type: jaŋy becomes jany.

Jany-Latyn does not call for changing the state alphabet. It offers one documented way to write the Latin Kyrgyz that people already type every day, and it already runs: a two-way converter that restores native Kyrgyz to Cyrillic exactly, keyboard layouts, and a testbed where every alternative can be tried on real text.

1. The Problem, and What Jany-Latyn Offers

Whatever happens at the state level (Part III), informal Latin transliteration is already routine. Kyrgyz speakers write Latin every day in messaging, code, social media, commerce, and diaspora communication, with no standard and considerable variation: ү appears as u, y, v, or w; ө as o, oe, or even the numeral 8; ж as j, zh, or dj; ң as n or ng; and vowel length is written with single letters, doubling, or a trailing h, depending on the writer. That variation is the problem this project addresses first, because it is a problem today regardless of state policy.

Jany-Latyn offers two things:

1. An orthography built to be typed and read now: Latin ö and ü; cedilla sibilants ç and ş; j for the voiced affricate; ŋ for the velar nasal; y for the close back unrounded vowel; doubling for phonemic length; and í for the palatal glide in every position.
2. A testbed in which each contested choice is a toggle. Swap ı for y, turn on allophonic q, write every ж as c, switch the sibilants to digraphs, substitute ñ for ŋ, or display Cyrillic ө and Kazakh ұ inline, then read the same text under each configuration and judge the result.

Two commitments shape the rest. The orthography is specified in two registers rather than one, a full form carrying the diacritics and a plain form in ASCII for unmodified keyboards, and the two are defined together so that neither is a workaround for the other (§19). And it is specified as far as machines need it: reverse conversion with its losses stated, a single search-folding key, and an alphabetical order with a collation tailoring (§17, §18, and Appendices B–C).

Above all, Jany-Latyn is designed to be derivable from Cyrillic. That constraint, more than any aesthetic preference, explains most of what follows, including the choices the author would make differently if starting from a blank page (§8).

1.1 What this system is for

Jany-Latyn has two intended uses, and they pull in slightly different directions:

• A transliteration standard for text that already exists in Cyrillic, meant to run without human intervention over corpora, archives, place names, and databases.
• A candidate orthography for writing Kyrgyz directly in Latin, whether informally today or officially at some future point.

Where the two conflict, the transliteration use wins. The reason is practical: a proposal that cannot process the existing eighty-six years of Cyrillic text automatically is a proposal that begins by discarding the corpus. This is why k is unified rather than split into k/q (§8), why the loan glide is folded into the same í used everywhere else (§15.1), and why the reverse conversion is defined to favor native Kyrgyz over Russian loanwords (§18.1).

1.2 What Jany-Latyn is and is not asking for

Jany-Latyn does not advocate a state transition. It is a tool for exploring the question and for making the informal writing that already exists more consistent. If the state question is settled in fifty or a hundred years, or never settled at all, the project still has a use: millions of Kyrgyz speakers are typing Latin today with no convention, and a documented convention is worth having independent of any decree.

That framing is not a rhetorical hedge. It changes the design. A proposal seeking legislation can assume a curriculum, enforcement, and an institution to maintain a loanword list, so it can afford the lexicon dependency (§8) and with it q. A proposal that has to work the moment someone opens a repository cannot, and has to be derivable, dictionary-free, and typable on hardware that already exists.

2. The Alphabet at a Glance

Jany-Latyn has 28 letters, in this order (§17):

A B Ç D E F G H I Í J K L M N Ŋ O Ö P R S Ş T U Ü V Y Z

a b ç d e f g h i í j k l m n ŋ o ö p r s ş t u ü v y z

Notation. Italics mark spellings. Square brackets mark pronunciation in the International Phonetic Alphabet, in which [y] is the sound of ü and [ɯ] the sound of ы, so kyz (кыз, "girl") and [kyz], the pronunciation of küz (күз, "autumn"), are different words. Cyrillic in parentheses gives the source spelling.

Fourteen letters need no comment: each is written the same way in every form, as in bala (бала, "child") and dos (дос, "friend"). The table gives the other twenty-two Cyrillic letters, each with its Jany-Latyn form, its plain form, and an example.

Long vowels are doubled, as in Cyrillic: took (тоок, "chicken"), küü (күү, "melody"). The plain form drops every mark and changes nothing else (§19).

In running text:

Cyrillic. Жаңы жылдын биринчи күнү тоого чыгып, бийик чокудан шаарды көрдүк. Үйгө кайтканда чай ичип, эс алдык.

Jany-Latyn. Jaŋy jyldyn birinçi künü toogo çygyp, biíik çokudan şaardy kördük. Üígö kaítkanda çaí içip, es aldyk.

Plain form. Jany jyldyn birinci kunu toogo cygyp, biiik cokudan saardy korduk. Uigo kaitkanda cai icip, es aldyk.

(On the first day of the new year we climbed the mountain and saw the city from a high peak. When we got home, we drank tea and rested.)

3. The Rule Behind It: Design Principles

Principle 1 is the constraint the others are tuned around. Principle 5 is described in §19 and is the part of the system most easily mistaken for a limitation.

The constraint is easiest to see through the two ways an alternative letter can differ from Jany-Latyn. A swap writes a Cyrillic letter with a different shape everywhere it occurs: ñ for ң, ı for ы, ch for ч. A converter can apply a swap on its own, and the testbed offers many of them. A split turns one Cyrillic letter into two Latin letters depending on the word: к into k or q, г into g or ğ, ж into c or j. Kyrgyz Cyrillic does not mark which words are borrowed, so a split can be applied correctly only with a list of words that does; applied by rule alone, it writes космос as qosmos. §8 calls this the lexicon dependency.

Jany-Latyn's spelling contains no splits. What the converter writes is therefore correct Jany-Latyn, with no hand-corrected second version to maintain. The same principle explains why q, ğ, and c are absent (§8, §11), and why the three letters under evaluation, ä, x, and w, which are splits, may be used by a person writing Latin directly but are never produced by the converter (§20). The one list in the system works in the other direction: an optional set of loanwords whose Cyrillic spelling cannot be restored by rule, consulted only when converting back to Cyrillic (§18.2).

4. The Letters That Shape the Rest

One k and one g. In Kyrgyz the uvular [q] and [ʁ] are positional variants of /k/ and /g/, predictable from the vowels of the word, and Kyrgyz Cyrillic has never written them apart. Writing q and ğ would be a split: the vowel rule that turns кар into qar also turns the loan космос into qosmos, and only a word list stops it. The author's own preference is q, on phonological and historical grounds; Jany-Latyn keeps one k and one g because it is defined by derivability (§8).

y for ы. With one k, only the vowel tells кы from ки, so ы needs a letter that cannot blur into i. Dotless ı differs from i by a dot that small type loses, that capitalization merges outside Turkish and Azerbaijani locales (kırgız uppercases to KIRGIZ), and that the plain form and search folding strip. y stays distinct everywhere, and it is the letter already used for ы in the country's English name (§9).

í for й. With y spent on ы, the glide needs a letter of its own. Writing it y as well would collapse тыйын into tyyyn, and plain i would turn бийик into the barcode biiik. í is i with a mark, as й is и with one, and it is written the same way in every position: toí, tyíyn, biíik (§10).

j for ж. Kyrgyz Cyrillic writes one ж for both the native affricate [dʒ] and the [ʒ] of loanwords, so one letter is needed, not two. Jany-Latyn takes j, the letter already used for ж in official romanization and on passports (Japarov, Jalal-Abad); the CTA's split into c for native words and j for loans would again need a word list (§11).

5. What It Gives, and What It Costs

5.1 What it gives

• No correction step. Any Cyrillic Kyrgyz text, from an archive to a register of place names, converts in one pass, and the result is correct Jany-Latyn rather than a draft to be proofread (§3).
• Native Kyrgyz round-trips exactly. Converting back to Cyrillic restores every native word; only some loanword spellings are simplified (§18.1). Latin and Cyrillic can therefore be used side by side, which a multilingual country needs in any case (§22.6).
• Letters a Kyrgyz reader would guess. j as in Japarov and Jalal-Abad; y as in the country's English name; ö, ü, ç, and ş as in Turkish, Azerbaijani, Turkmen, and the CTA; í and ŋ echoing й and ң; long vowels doubled as in Cyrillic (§22.4).
• Fewer misreadings before anyone is taught. A reader meeting a sign reads the base letters with the values already known. Чыгуу (exit) written ÇYGUU can mislead by one letter, C for ч; the CTA-aligned ÇIĞUU adds a second, I for ы, and read with Cyrillic values it turns one of the commonest words on public signs obscene (§22.4).
• Safe in software. Capitals never change the word (KYRGYZ cannot become KIRGIZ), no apostrophe splits a word in search or selection (§16), and the plain form is also the search key, so jany finds jaŋy (§19, Appendix B).
• Easier to find, as a prediction. Spellings such as kylym and jaŋy belong to no other language, while the CTA-aligned qılım is also an Azerbaijani verb form and canım (жаным) is Turkish (§22.2). This is offered as a prediction to be tested, not as a reason for the choices (§23.2).
• Usable now. No state decision is needed, and the converter, keyboard layouts, and testbed already run (§7, Appendix D).

5.2 Known limitations

• Latin→Cyrillic conversion is lossy for Russian loanwords. Native Kyrgyz vocabulary round-trips exactly; loan-only distinctions (ц, щ, ё versus йо, ъ, ь) do not survive the round trip. This is by design and is specified precisely in §18.1.
• The plain form is one-way. Text written in ASCII cannot be mechanically restored to the full form or to Cyrillic (§19).
• ŋ is unevenly available on phones. It is reachable on the Android keyboard's long-press menu in the US locale (author-confirmed); iOS behavior is unverified. Where it is missing, the plain form writes it n (§19); ñ is available as a comparison setting (§14).
• Dialect coverage is partial. Three extended letters (ä, x, w) are offered for evaluation and none is part of Jany-Latyn (§20); other dialect features are not represented at all.
• No pedagogy exists yet. Nothing here has been tested with learners, in classrooms, or in handwriting.

6. How It Compares

Jany-Latyn agrees with BGN/PCGN, the romanization Kyrgyzstan uses for geographic names, on the letters that matter most for recognizability, j for ж and ö and ü, but it is built to be written by Kyrgyz speakers rather than transcribed for foreign readers (§21.1). It shares ö, ü, ç, and ş with the Common Turkic Alphabet agreed in Baku in 2024 and departs from it on six letters, each argued in Part II (§21.2). It shares the same four letters, and y for ы, with Syrtbai Musaev's 28-letter draft, a revival of Tynystanov's 1928 alphabet that restores the k/q split, writes ж as c, and marks the velar nasal with ņ (§21.3).

(Condensed from the table in §21.1, which adds ALA-LC and ISO 9.)

In this paper "Jany-Latyn" always means the proposed alphabet. Every other configuration the testbed can display, including a single swap such as ñ for ŋ, is a comparison setting: something to be judged against Jany-Latyn on the same text, not a variant of it. Three comparison settings recur through the paper, each a step further from Jany-Latyn. The phrase аңкыгый жигит ("a gawky young man") contains every letter they change:

The last column is the argument of §3 in one line: each step away from Jany-Latyn adds a split, and a split is written correctly only with a word list. Жигит is front-harmonic, so its г stays g even where ğ is written, and its и stays i beside the ı of the q+ğ+ı row. Real words and a running passage in each setting are in §21.2, and the alphabets side by side in §17.3.

7. The Testbed

A client-side application offers one-click presets, the Jany-Latyn baseline against the CTA-aligned setting (ı for ы, y for the glide, c for ж, q and ğ, and ñ for ң), alongside selectable options for each variable argued in Part II:

A side-by-side inspector runs any combination against real text.

A few options are coupled for convenience. Selecting ı also switches on q, ğ, and the y glide: with ы written ı, y is free, and keeping í would put ı, i, and í side by side as three letters told apart only by their marks (§17.3). Selecting q carries ğ with it, and switching back undoes each coupling. The couplings only seed options, so ı with k, q without ğ, and ı with the í glide all remain reachable by setting the option afterwards; the one restriction is that the y glide is offered only while ы is written ı, since the two would otherwise collide. The engine itself accepts any combination. The testbed also warns, without preventing it, when the ĩ glide is combined with ñ, since the same tilde would then do two unrelated jobs. Section 7 of SPEC.md lists the couplings in full.

The toggles are not a menu of equally valid house styles. Each one exists so that the counterexample can be generated rather than described: q turns kosmos into qosmos, ı shows what uppercasing does to the result, and Cyrillic ө can be switched on and the output sorted as a name list.

The project is run as a proposal that invites falsification rather than adoption: every contested choice is implemented, all source is public, the hosted testbed records which configurations readers choose so that the distribution can be published beside the arguments, and a reproducible case where the engine corrupts native Kyrgyz counts for more than agreement (§24).

About the Author

The author is a life-sciences professional and self-taught developer with university coursework in bioinformatics and philology, and thirty years of native Kyrgyz alongside fluent English and Russian. That standpoint shapes the document in two ways. Judgments about what reads naturally in Kyrgyz, what collides with what, and what people actually type are native-speaker judgments, offered as such. Judgments about tokenizers, collation, and rendering are an engineer's reasoning from documented behavior, and where they are predictions rather than measurements, they are marked as predictions and listed in §23 as things the testbed is built to check.

Part II — The Case for Each Letter

8. One `k` and One `g`: Synharmonic Allophony

Several Turkic orthographies use separate letters for the uvular stop [q] and the uvular fricative [ʁ]. In Kyrgyz these are positional allophones, predictable from the vowels of the word:

A Kyrgyz speaker reading kar produces [q] automatically; producing front [k] there would violate the language's phonotactics. Writing the two separately encodes a rule speakers already apply. Turkish makes the same choice for k and g on the same grounds, though it retains ğ for a related distinction.

The author's own preference is q. On phonological grounds it is the more honest letter, it is what most of the Turkic and Arabic-script world reaches for, and it has real historical weight in Kyrgyz specifically. Jany-Latyn nevertheless unifies k and g, and the reason is Principle 1 rather than a belief that q is wrong.

If q is written, the voiced counterpart follows: egemen has front [g] and kagyluu back [ʁ], and every system that has split the voiceless pair has split the voiced one too. Tynystanov's 1928 alphabet had q and ƣ alongside k and g [4], the CTA has q and ğ [8], and the 2021 Kazakh alphabet uses ğ-breve for ғ [5]. A q-without-ğ system is possible, since the vowels predict g just as they predict k, but nothing defends it: the rule that produces qosmos also produces ğaz, ğaloş, mağazin, and proğramma, and only a loanword lexicon stops either. The two letters differ only in which layer they tax. q is an ASCII key and survives into the plain form; ğ is a diacritic and folds back to g there. This system's position is therefore that q entails ğ. Jany-Latyn declines both, and the testbed keeps them as independent toggles so that the asymmetric configuration can still be looked at.

Three considerations keep Jany-Latyn on unified k and g:

• Derivability from Cyrillic. Kyrgyz Cyrillic has never had a separate қ [4], so q cannot be recovered from the source text by any rule that does not also know which words are loans. Note the historical nuance: the merge was not Tynystanov's doing. His Arabic reform and his Latin alphabet both distinguished the two sounds [4]. The single К arrived with the Cyrillic alphabet in 1940, two years after he was executed in the purges. Whoever made that decision and for whatever reason, eighty-six years of literacy have since been built on it.
• Loanwords. Modern Kyrgyz absorbs kosmos, karta, bank, respublika, all with plain European [k], never [q]. A harmony rule cannot tell them from native back-harmonic stems: it produces qosmos, parqta, and mixed-harmonic hybrids like Respubliqası for correct Respublikasy. There is no clean fix short of a maintained loanword lexicon. Call this the lexicon dependency: a spelling rule that cannot be applied until something knows which words are borrowed. A state-backed standard, with an institute and a budget behind it, can carry that dependency. A converter that has to run over an archive unattended cannot.
• Economy. Two fewer letters to teach and to reach on a keyboard.

The strongest counter-argument runs the other way. Because loans keep velar [k] in back-vowel words, [k] and [q] arguably now contrast in modern spoken Kyrgyz (kosmos against koş), which is the classic profile of an allophone becoming a phoneme. If that analysis is right, a future standard written by people rather than derived by machine should adopt q. Place names sharpen it: Каракол and Каракөл are different towns, one letter apart as Karakol and Karaköl and two apart as Qaraqol and Qaraköl. Jany-Latyn's position is narrower than "q is wrong": it is that q cannot be derived from Cyrillic without a lexicon, and this system is defined by derivability.

The harmony rule is also only as good as the vowels it reads. Cyrillic writes one а for both the back vowel and its fronted variant, so q mode cannot tell äka from aqa; the extended letter ä supplies what the rule is missing (§20).

The testbed's q toggle exists so that this can be checked rather than asserted. Run over contemporary news or social-media text, q mode puts the loanword distortions on screen within a paragraph.

9. `y` for `ы`: The Back Unrounded Vowel

The CTA and Turkish write ы [ɯ] as dotless ı [8]. Jany-Latyn writes y, and the testbed allows the two to be swapped. Doing so surfaces two compounding problems.

The first is casing. Dotless ı and dotted i are the lowercase forms of two different capitals, and mapping them correctly needs locale-aware case conversion; the Turkish and Azerbaijani mappings are a documented special case in the Unicode Character Database [11]. Turkish answers with a second capital, ı to I and i to İ. That holds only if every system in the chain applies those mappings, and only if the reader knows the letter. A Kyrgyz reader who has not learned Turkish has no reason to know İ exists, so a correctly cased KIRGIZ still reads as КИРГИЗ: the distinction survives the encoding and is lost in the reading. The lake makes the same point without an all-caps example, since Ысык-Көл is written Isıq-Köl and the capital that opens it is read as dotted i by anyone outside those locales.

Outside a tr_TR or az_AZ locale, which is to say in most databases, indexes, and runtimes, the distinction is lost in the encoding too, since both letters collapse to plain I. КЫРГЫЗ and КИРГИЗ then arrive as the same string, and kırgız uppercases to KIRGIZ, an exonym Kyrgyzstan left behind. The dotless configuration therefore adds a dependency on locale-correct casing everywhere the text travels. Latin y carries none: Y and I are distinct capitals under every locale and every case function, and the state's own English name, used since 1991 across treaties, passports, and diplomatic usage, is Kyrgyz Republic.

The second is that the distinction is unreliable. y and i are unmistakable at any size, so kyna and kina never collide. Under dotless ı the same pair differs by a single dot, which degrades in small type, in handwriting, and on low-resolution screens. Below the full form it disappears altogether: the plain form and search folding both reduce ı to i, which leaves kına and kina as one string in casual writing and one key in an index. With y the distinction survives everywhere. Once the vowel can no longer carry the difference, a consonant has to, which is why the testbed's ı mode switches on allophonic q. Adopting ı therefore does not cost one letter: it reopens the q question of §8.

Musaev's proposal reaches the same letter from a different direction (§21.3), and in doing so departs from its own historical base: Tynystanov wrote ы as ь and spent y on ү [4].

10. `í` for `й`: The Palatal Glide

Writing both и [i] and й [j] with plain i produces a barcode: бийик → biiik, кийим → kiiim, кийин → kiiin. Three or four vertical strokes in a row are hard to parse and impossible to tell from a long vowel. The testbed renders this mapping on demand (§7), so the claim can be judged on a page of real text rather than on three words.

The Jany-Latyn glide is í / Í in every position: aí, toí, tyíyn, biíik, kiíim. Four arguments support it.

1. A separate glide letter is forced, not chosen. The CTA writes the glide y, which it can afford because it moves the back unrounded vowel to dotless ı [8]. Jany-Latyn assigns y to that vowel (§9), so the two roles cannot share a letter. If they did, native sequences of vowel plus glide would collapse: тыйын as tyyyn, кыйык as kyyyk, айыл as ayyl. That is the same legibility failure the letter exists to fix, reappearing one vowel to the left.

2. The acute is chosen for its relationship to й, not for its meaning elsewhere. й is и with a mark above it; í is i with a mark above it. A reader literate in Kyrgyz Cyrillic already treats that pair as related but distinct, and i/í carries the relationship across intact.

3. It preserves lexical roots. Kyrgyz is agglutinative and stems must stay recognizable. Кийим (clothing) comes transparently from кий- (kií-, to wear); тийиштүү comes from тий- (tií-). Contracting these to kiim and tiishtüü truncates the verb stem and splits related forms arbitrarily: biíi (its dance) would keep the glide while biíik (tall) lost it.

4. It makes the glide exceptionless. Every Cyrillic й becomes í in every position, with no positional rules and no lookups.

The barcode argument applies to the full form only. In the plain form í drops to i and biíik becomes biiik again (§19). This is a real cost of the two-register design, not something the design escapes.

Why the acute, and not another mark

In short: of the marks that can sit on i, the acute is the lightest and the most widely available, and within this system it has no competing job, since Kyrgyz stress is unwritten and length is doubled. i is the right base because [j] is the semivocalic counterpart of [i], and because a marked i falls back to i rather than to y when the mark is stripped. The one real conflict, with stress marks in teaching materials, is narrow and has a better remedy than changing the glide.

Argument 2 leaves an objection standing. The acute has associations elsewhere, marking length in Czech, Slovak, Hungarian, and Irish and stress in Spanish, and a foreign association is what ruled out the macron in §12.1. The difference is what a Kyrgyz reader brings to the page. Kyrgyz stress is regular and never written, and length is written by doubling, so within this system the acute has no competing job; the macron's usual value, length, is a live category in Kyrgyz.

Two questions hide inside the choice, and they are better taken apart: which mark, and which letter it sits on.

Of the marks that can sit on i, a diaeresis (ï) would collide visually with the fronting marks on ö and ü. A circumflex (î) is available, but in Turkish it marks length in Arabic loanwords (millî, tarihî), which is the objection that ruled out the macron, arriving from the nearest Latin Turkic tradition. A caron (ǐ) is heavier at text sizes and sits outside Latin-1. Two alternatives are close enough to ship as testbed options rather than to dismiss (§7). The breve ĭ has the best claim of them: it is the literal analogue of the Cyrillic letter, which is и under a breve, and it is what ALA-LC chose for that reason [8]. It loses on weight and reach, being heavier than an acute at reading sizes, outside Latin-1, and absent from most layouts and long-press menus, and those are differences of degree, so the testbed renders whole texts in it and lets the comparison be made on a page. The tilde ĩ is weaker still and carries an interaction: a reader who also selects ñ for the velar nasal puts the same mark on two letters doing unrelated jobs, which the testbed warns about. í (U+00ED) is the lightest of the set, and it is in Latin-1, on Spanish, Portuguese, Irish, and Icelandic layouts, and on standard mobile long-press menus.

The base is a separate question, and the systems compared in §21.1 answer it four ways: j in ISO 9 and Tynystanov's 1928 alphabet, y in BGN/PCGN and the CTA, ý in Turkmen [20], and a marked i in ALA-LC. The j and y choices map the sound onto a letter that already carries /j/ in some Latin tradition, j from German and Slavic spelling and y from English and Turkish. Turkmen is the instructive one, because it starts where this orthography starts: it has no dotless ı, so y holds the back unrounded vowel [ɯ] and the glide is written ý, and one syllable can carry both, as ýyl (year) does.

Two things argue for i over y as the base, and neither depends on Cyrillic being the source script. The first is that [j] is the semivocalic counterpart of [i], the same articulation without a syllable of its own, which is why transcription can write it as a non-syllabic i, and why ko-i-on and ty-i-yn, said fast enough, arrive at koíon and tyíyn. The glide keeps that identity wherever it lands, including after o and y. The semivocalic counterpart of [ɯ] is not [j] but [ɰ], a velar approximant Kyrgyz has no use for, so ý would assert a kinship the phonetics denies. Cyrillic happens to agree, since й is и with a mark and alternates with и inside the same roots (кий-, кийим).

The second is what happens when the mark comes off. The plain form and search folding both strip diacritics, so í drops to i while ý would drop to y. Under í, тыйын and айыл reach ASCII as tyiyn and aiyl. Under ý they would arrive as tyyyn and ayyl, the collapse argument 1 introduced a separate glide letter to prevent, returning by the back door. Sorting adds a third consideration: under the ordering principle of §17.1 a marked y would be taught and filed at the end of the alphabet, away from the letter it alternates with.

The one genuine conflict is with stress notation. Teaching materials, dictionaries, and foreign-language textbooks mark stressed syllables with an acute (a-tá, o-ro-mó), and on i that mark and the glide would be the same character. The conflict is narrower than it first appears: it arises only when stress falls on /i/, leaving every other vowel free to take a stress acute unambiguously, and Kyrgyz stress is largely predictable and word-final, so marking it is a device for exceptions and loans in a small body of specialized material rather than a feature of ordinary text. Weighed by frequency, the glide occurs in running text constantly and the stress mark does not. The resolution is to give pedagogy a device that does not compete with any letter: bolding the stressed syllable, or the IPA stress mark before it (bi-ˈíik), which has the further advantage of working identically across the whole vowel inventory. §23.6 lists the convention as work still to be written.

Kazakhstan's 2018 alphabet is the cautionary case (§21.5), and its lesson is not that acutes fail: an acute scattered across six or seven letters reads as clutter where a single systematic use does not. í is one letter with one job.

11. `j` for `ж`: The Voiced Affricate

Native initial and medial ж is a voiced postalveolar affricate [dʒ], the sound of English j in journey, not the Russian-mediated fricative [ʒ] that zh implies. Jany-Latyn writes j: jakşy, jaŋy, jol, jigit.

This is not a novel choice for Kyrgyz. The BGN/PCGN romanization, which the Kyrgyz government adopted for rendering geographic names in Latin script [8], also writes ж as j, and the result is visible on Kyrgyz passports and in international coverage: Japarov, Jalal-Abad. Against that, zh costs an extra keystroke on one of the highest-frequency consonants in the language and tells an international reader the wrong thing.

The divergence from the CTA, which assigns c to this sound and j to [ʒ] [8], is deliberate and is argued below.

j where the CTA uses c. In the CTA, following Turkish, c is [dʒ] and j is the [ʒ] of loanwords. Jany-Latyn has one letter to place rather than two, since Kyrgyz Cyrillic writes a single ж for both sounds, and it places it on j. To a reader without Turkish literacy, c is not a neutral symbol: it suggests [s] to a French-influenced reader and [k] to an English one. j is read as something close to [dʒ] by nearly every Latin-literate reader who has not been trained on Turkish, and it is already the letter used for ж in Kyrgyz official romanization. For a system aimed at diaspora typing on non-Turkish keyboards and at international legibility, c is a false friend.

The split also carries the lexicon dependency of §8: the CTA writes native жазуу as cazuu and the loan журнал as jurnal, and Kyrgyz Cyrillic writes one ж for both. With q and ğ, the CTA-aligned setting needs a loanword list at three points in its consonants, where a single j, k, and g need none.

This is also a divergence from the Kyrgyz Latin tradition, not just from a modern committee. Tynystanov's alphabet used j for the glide й and c for the affricate ж (replaced by ƶ in the 1938 revision) [4], the same assignment the CTA and Musaev use. This orthography accepts the break: ƶ reads as a modified z, which signals [dʒ] no better than c does, and it remains a rare Latin Extended-B character with thin font coverage.

12. `ö` and `ü`: The Front Rounded Vowels

12.1 `ü` for `ү`, and the alternatives

The Jany-Latyn letter for ү [y] is ü / Ü, matching Turkish, Azerbaijani, and Turkmen usage for the same vowel, and matching the 2021 Kazakh Latin alphabet, which also selected ü-umlaut for Kazakh ү [5]. Length is handled separately by doubling: short ü in kün (day), long üü in küü (melody).

The alternatives are paired, not independent. The testbed offers four configurations of the front rounded vowels, and three of them are the same experiment run with different glyphs:

This pairing matters for reading the results. Every alternative to ü ships alongside Cyrillic ө, so choosing ұ or ū does not trade one letter for another: it also accepts the script-mixing costs set out in §12.2. The costs compound instead of sitting side by side, and kөrūū against körüü is what that compounding looks like on the page.

The macron ū was tested and set aside. In Latvian and in Māori the macron marks vowel length, and length is the value a reader is likeliest to bring to it. Kyrgyz has phonemic length, so ū for a vowel quality offers a Kyrgyz reader an available and wrong reading. The internal clash is the decisive one: length in this system is written by doubling, so the long form of ū would be ūū, which reads as length marked twice. The argument is specific to Kyrgyz rather than universal: the 2021 Kazakh alphabet uses ū-macron for Kazakh ұ [5], which works there because Kazakh does not carry the same length contrast.

Kazakh ұ is retained as a transitional display option because Cyrillic-adjacent glyphs come up in script-planning discussion as a bridge for Cyrillic-literate readers. Two problems come with it. Its narrow crossbar is the feature that distinguishes it, and crossbars are what degrade first at small sizes and under aggressive anti-aliasing, at which point it approaches Latin y, a letter that already carries its own phonemic load here. And unlike ü, which drops cleanly to ASCII u, ұ has no ASCII fallback at all, so it fails in exactly the environments the plain form exists for.

12.2 `ö` for `ө`, and the cost of script-mixing

The Jany-Latyn letter for ө [œ] is ö / Ö. Cyrillic ө is retained in the testbed as a display mode for transitional and educational material, not as a candidate for Jany-Latyn.

The author finds inline Cyrillic ө perfectly legible, which is why the case against it cannot rest on legibility. Keeping one Cyrillic letter inside otherwise-Latin text carries three costs, and all three hold however legible a reader finds the glyph:

• Glyph metrics. Fonts without Cyrillic coverage substitute another face for that one glyph; fonts with it still draw their Latin and Cyrillic as separate designs, so ө commonly sits at a smaller x-height than the Latin lowercase beside it. Either way the word carries a letter that does not match its line, and the effect changes with every font stack the text passes through. This is the least speculative of the three costs, and the cheapest to check (§23.4).
• Collation. Under the Unicode Collation Algorithm, Latin and Cyrillic sort in different weight groups [6], so mixed strings sort unpredictably against pure-Latin ones. Alphabetizing a name list becomes locale-dependent in a way it need not be.
• Tokenization (predicted, not measured). SentencePiece splits input at Unicode script boundaries by default, so kөl may be segmented into three pieces rather than one [7]. Byte-level BPE tokenizers whose pre-tokenizer groups all letters regardless of script will not necessarily behave this way, and ö and ө occupy the same two bytes in UTF-8, so the effect depends on the tokenizer. §23.1 states how to test it.

Standard ö avoids all three without asking anything of the reader that ü does not already ask.

13. `ç` and `ş`: The Sibilants

The Jany-Latyn sibilants are ç and ş: çaí, şaar. Digraphs were considered and set aside for the full form because Kyrgyz morphology produces collisions with them.

The agentive suffix -чы/-чи on a stem ending in ш is the everyday case: баш + -чы → башчы gives the boundary-obscuring bashchy under digraphs, against the clean başçy with cedillas. The s+h sequence in personal names is the other: Исхак (Is-hak) transliterates via digraph to ishak, which is also how ишак would come out, a Russian and Uzbek word for "donkey" that most Kyrgyz speakers recognize, even though the native word is эшек. Cedillas keep them apart: ishak against işak.

Russian щ merges into ş (ящик → íaşik, борщ → borş) instead of producing a trigraph. Spoken Kyrgyz realizes it as [ʃ], and şç would collide in reverse with native başçy.

The digraphs are not discarded; they are relocated. The reverse converter accepts legacy ch, sh, and sch so that older text and casual writing convert correctly, and the testbed can render whole texts in digraph mode for comparison.

14. `ŋ` for `ң`: The Velar Nasal

ŋ / Ŋ (U+014B / U+014A) marks the single phoneme [ŋ] and keeps it distinct from a real n+g cluster: күнгө (stem kün + dative -gö) stays küngö, while жаңы is jaŋy, not the ambiguous jangy.

Eng was chosen for continuity of shape. Cyrillic ң (en with descender) and Tynystanov's 1928 Latin ꞑ (n with descender) [4] both carry the n below the baseline, and ŋ (eng) hooks its right leg to the same effect. Among the available options it is the only one with continuity in both the Cyrillic and the 1928 Latin traditions. The CTA's ñ (n with tilde) puts its mark above the letter instead, and for readers with any Spanish exposure it carries a specific wrong reading, the [ɲ] of piñata and mañana.

ŋ also has real costs:

• Mobile availability is uneven. It is reachable by long-pressing n on the Android keyboard in the US locale; iOS and non-US locales have not been checked, and ñ is the option reliably present everywhere.
• Capital Ŋ has two different designs in circulation, an enlarged lowercase form and an N-based form, so all-caps text can look inconsistent across fonts.
• It sits outside Latin-1 and Windows-1252, where ñ is present.

Turkmen wrote this sound ꞑ, ñ, and ng before settling on ň, the caron rather than the tilde [20]; Kazakhstan moved from Ŋ to Ñ within 2021, and the Uzbek draft keeps the digraph ng [9]. The question is open across the family. Jany-Latyn keeps ŋ on the derivation and false-friend arguments, and offers ñ as a comparison setting in the testbed for anyone who weighs mobile typing more heavily. The keyboard layouts (Appendix D) make ŋ a single keystroke on desktop, and the plain form writes it n (jaŋy → jany), which is what people already type.

15. Vowels, Length, and `е`

Kyrgyz has eight basic vowels, organized by front/back and rounded/unrounded harmony:

Six of these contrast short and long: аа, ээ, оо, өө, уу, үү. The two high unrounded vowels, ы and и, have no long counterparts [4]. Length is written by doubling throughout.

The vowels whose Latin letters required a decision are argued in §9 (y) and §12 (ö and ü), and the glide in §10; length and е are covered below.

15.1 Phonemic length, and the `е` problem

Length is written by doubling: ток [tok] (sated) against тоок (chicken), Ala-Too, aalam, jeek, kuu, küü.

The alternative was a macron over a single vowel, on the Latvian and Māori model, which would give tōk for took (тоок, "chicken") and Ala-Tō for Ala-Too. It is the more compact of the two notations, and it is what a reader of those orthographies would expect. Doubling is preferred to it for four reasons: combining marks are stripped or normalized inconsistently by search and database layers; doubling needs no dead key or long-press, only the same key twice; Kyrgyz Cyrillic already writes length by doubling, as does the Kyrgyz Arabic orthography [4], so the convention is inherited rather than invented; and two of the six vowels that lengthen already carry a mark, so long öö and üü would need the stacked ȫ and ǖ, which sit outside Latin-1, are thin on font coverage, and cannot be typed without a dead key.

Kyrgyz long vowels, sometimes described as an archaism, are largely a secondary development from lost intervocalic consonants (тоо from an older form with a velar) rather than Proto-Turkic length of the kind Turkmen preserves [26]. Either way the contrast is live in modern Kyrgyz, which is all the orthography needs.

The one irregularity that must be resolved explicitly is е. Kyrgyz Cyrillic never doubles it; long /eː/ is written ээ in every position (ээги, керээз, кээде, жээк), which Jany-Latyn renders as it looks: ee (eegi, kereez, keede, jeek). But е has a second job in Russian loans, where it represents [je] word-initially (Европа, Енисей) and after a vowel (переезд, проект). Jany-Latyn handles that with the glide letter it already has: loan [je] is íe (Íevropa, pereíezd, proíekt), native word-initial [e] stays e (ene, el), and post-consonantal [e] stays e (kel, mektep). Post-vocalic е in a loan that is not iotated in Kyrgyz speech, as in поэт and аэропорт, is written plain: poet, aeroport. Three graphemes (e, ee, and íe) cover every position. What they do on the way back to Cyrillic is specified in §18.1.

16. Loan Signs `ъ` and `ь`, and `я`, `ё`, `ю`

The hard and soft signs mark hiatus and palatalization that Jany-Latyn captures with the glide it already has. A separating sign before an iotated vowel is absorbed into í (объект → obíekt, семья → semía, разъезд → razíezd). A coda soft sign, which Kyrgyz phonology does not realize as palatalization in natural speech, is dropped (июль → iíul, апрель → aprel, роль → rol). Iotated vowels decompose consistently: я → ía, ё → ío, ю → íu (saíakat, koíon, aíuu), while non-iotated hiatus stays plain (диаграмма → diagramma).

The result is that every word written in Jany-Latyn is a continuous alphabetic string matching \p{L}+, with no internal apostrophes. That property has concrete consequences: search indexes and naive tokenizers do not fracture the word at a punctuation mark, and double-clicking selects all of it [10]. Uzbek, which uses apostrophes in oʻ and gʻ, offers a live demonstration of what the alternative costs in practice, since the mark is routinely replaced by whichever quote character the keyboard produces.

This property does not, however, make URLs ASCII-clean. í, ŋ, ö, ü, ç, and ş are all non-ASCII and percent-encode in a URL path (ŋ becomes %C5%8B). What the \p{L}+ property gives is a single, unbroken token; what produces clean slugs is the plain form (§19), which is the correct layer for slug generation: jaŋylyktar → janylyktar.

17. Alphabetical Order

This section states the alphabetical order of Jany-Latyn, the sequence a reader learns; §17.4 shows what it changes in sorting, and Appendix C gives the rules software needs.

17.1 The ordering principle

One question decides most of the order: where a modified letter goes. Kyrgyz Cyrillic and the Latin Turkic orthographies answer it the same way, placing it immediately after the letter it modifies rather than at the end of the alphabet. Kyrgyz Cyrillic places ң after н, ө after о, and ү after у; Turkish and Azerbaijani place ç after c, ş after s, and ö and ü after o and u. Jany-Latyn follows both.

17.2 The Jany-Latyn alphabet (28 letters)

A B Ç D E F G H I Í J K L M N Ŋ O Ö P R S Ş T U Ü V Y Z

a b ç d e f g h i í j k l m n ŋ o ö p r s ş t u ü v y z

Three consequences of the principle are visible in it. Í follows I, mirroring the и/й adjacency a Cyrillic-literate reader already knows. Ŋ follows N, as ң follows н. Y sits in its Latin position near the end rather than where ы falls in Cyrillic, matching international expectation for the letter's shape.

The Common Turkic Alphabet follows the same principle. Its 34-letter order places Ğ after G, X after H, Ñ after N, Ö after O, Ū and Ü after U, Ə after E, and Q immediately after K [17], each letter beside the one it pairs with rather than at the end or in inherited Latin sequence. The q rows below place Q after K for the same reason.

C, Q, W, and X are not members. Jany-Latyn never emits them: ч is ç, ц is ts, and the rest belong to other configurations or to foreign spellings.

17.3 The alphabets side by side

The alphabets are given in case pairs because the pairs differ: in the last three rows I is the capital of dotless ı, and dotted i takes İ.

In the q+ğ+ı row, writing ы as ı frees y, and the glide takes it: the argument for í in §10 is that y is otherwise spent, and here it is not. Y returns to its ordinary Latin place at the end. Keeping í as well is possible and the testbed offers it, at the cost of an alphabet carrying I, İ, and Í as three letters told apart only by the marks above them.

The last row is the Baku alphabet itself [17], so the others can be read against it. The row above it is what a Kyrgyz adaptation would use: the same alphabet without Ə, Ū, and X, for the reasons given in §21.2. Neither has W.

17.4 Sorting: what the tailoring changes

Alphabetical order has to be handed to software as a tailoring: a short list of adjustments to the default Unicode sort order [6], which knows nothing about Kyrgyz and files ç as a variant of c. Appendix C gives that list; this section shows what it changes for a reader who will never look at the syntax.

Two of those rows hold a result that readers often find counterintuitive. Because ö and ü count as letters in their own right and not as accented forms of o and u, every word with the base vowel in that position sorts before every word with the modified one: köl follows konok rather than sitting beside kol. That is what the Cyrillic order already does with о/ө and у/ү, so a Kyrgyz reader finds it familiar and a search implementer often does not.

The last row is the practical cost of the q+ğ+ı setting for anyone maintaining a dictionary or an index: it is not one letter's position but a reshuffle of every entry containing the language's most frequent vowel.

Two larger movements fall outside the table, which compares Latin configurations rather than Latin against Cyrillic. Because ç anchors to c, Чычкан moves from the twenty-eighth letter of the Cyrillic alphabet to the third of the Latin one, so a name index does not shift on conversion, it re-sections. A word whose first letter becomes a multi-character grapheme changes section outright: Екатеринбург is filed as Íekaterinburg, under Í and away from every other E entry.

18. Converting Back to Cyrillic

Cyrillic→Latin conversion is deterministic and rule-based: every Jany-Latyn form is computed from position and vowel context, with no lexicon. Latin→Cyrillic conversion is also deterministic, and it restores native Kyrgyz vocabulary exactly, but it is not lossless for loanwords. The Cyrillic→Latin mapping is many-to-one, since several Cyrillic distinctions that exist only in Russian loanwords are deliberately merged, and a merged distinction cannot be recovered.

§18.1 and §18.2 are for anyone judging the orthography: what the converter does with loanwords, and why. What an implementation has to reproduce is specified in Appendices A–C, and the alphabetical order in §17.

18.1 The native-priority rule

Where two Cyrillic spellings map to the same Latin string, reverse conversion restores the one that occurs in native Kyrgyz vocabulary. Distinctions that exist only in Russian loanwords are the ones allowed to be lost.

This single rule decides every ambiguous case, and it follows from Principle 1: the corpus that must survive automatic conversion is the Kyrgyz one.

Which losses appear in the table is partly downstream of a question §23.6 leaves open: whether loanwords are spelled from their Russian source or from Kyrgyz pronunciation. poet → поет and aeroport → аеропорт are losses under the current approximating treatment and would not be under a source-faithful one.

Only ts, post-vocalic íe, and legacy sh touch native text; the other rows decide only which loanword spelling is lost. ts matters most, because the conditional puts т+с across a morpheme boundary on every stem ending in т, so sending ts back to ц would corrupt an open class of native forms.

18.2 The optional loan-restoration list

Losing loanword spellings is acceptable for corpus work and unacceptable for, say, converting a legal text. The reference engine therefore ships a small, explicitly documented loan-restoration list: the known loanwords whose Cyrillic form cannot be derived by rule (semía → семья, statía → статья, Ishak → Исхак, and similar), layered on top of the rule-based core and switchable.

The list is a lookup table, and the claim made for the architecture is narrower than "no dictionaries anywhere": the core is complete and dictionary-free on its own, and the list only ever restores loan spellings the core has already flagged as lossy. It never affects native vocabulary and never changes Cyrillic→Latin output.

The list introduces an asymmetry of its own: a bypass that restores ishak to Исхак will get the Russian loan ишак wrong in digraph mode. The trade is deliberate, since the proper name is the one that appears in documents.

The same rule sets where the list stops. Асхат converts to Ashat and comes back as Ашат, and no entry can repair it: in digraph input ashat is also native ашат, which §18.1 restores first. Исхак was admissible because the word it displaces, ишак, is borrowed. The list can hold a name whose competitor is a loan and not one whose competitor is Kyrgyz, which is a firmer limit than the general caution that exception lists cost something.

19. Two Registers: The Full Form and the Plain Form

Jany-Latyn defines two registers together rather than one script with a workaround attached.

The full form — ö, ü, í, ŋ, ç, ş — is the complete form: print, dictionaries, official text, the testbed's default output, and anywhere the orthography is presented as itself.

The plain form is ASCII, for unmodified keyboards:

ö → o, ü → u, í → i, ŋ → n, ç → c, ş → s

Every letter loses its diacritic and nothing else. Çaí becomes cai, şaar becomes saar, jaŋy becomes jany, biíik becomes biiik.

The case for mapping ç → ch and ş → sh instead is real: the digraphs signal the sound better to a reader, and they are what people type today (chai, jakshy). Two considerations outweigh it here. First, coherence: if every other letter loses a mark, the sibilants should too, and the plain form then teaches the shape of the full one instead of competing with it: the reader who types cai is one diacritic away from çaí, while the reader who types chai is spelling a different word shape. Second, folding: standard accent-stripping already maps ç→c and ş→s, so with this mapping the plain form is the folded form and search folding is nearly free (Appendix B). With digraphs, the two forms fold to different keys and every index needs custom handling.

The digraph fallback is transitional, not a second house style; the project's own early drafts used ch/sh as the reference. The direction of travel is toward the strip mapping, where a reader who has internalized ç writes çöírö casually as coiro and reads it as the same word.

The cost is a merger. Stripping ş to s merges it with native s: баш and бас both become bas, аш and ас both become as. Digraph fallback avoids that merger but reintroduces the sh ≈ с+х collision and lengthens the word. This is a genuine trade, so the testbed implements both fallbacks and lets a reader compare them over real text. The right way to settle it is to count how many distinct words each mapping merges across a corpus (§23.3).

The plain form already has mergers anyway, and always did: кол (hand) and көл (lake) both surface as kol; жаңы (new) and жаны (his soul) both as jany. This matches how Kyrgyz speakers already text, and it resolves the way informal writing always resolves ambiguity: from context, in real time.

Two properties of the plain form limit what can be built on it:

• The plain form is one-way. biiik cannot be mechanically restored to biíik, and cai cannot be restored to çaí or to чай. ASCII text is a destination, not an intermediate form. Conversion tools should accept it as input on a best-effort basis and should not claim round-tripping.
• It is a departure from current habit, not a description of it. ö→o, ü→u, í→i, and ŋ→n match what people type now. ç→c and ş→s do not; chai and jakshy are the live forms. This mapping is an attempt to pull habit toward the full form, which is a normative choice and should not be presented as a descriptive one.

20. Letters Under Evaluation: `ä`, `x`, `w`

Three sound differences are heard in Kyrgyz speech that Kyrgyz Cyrillic does not write. Latin has letters for all three and they are cheap to type, but they are not equally strong: ä has a case for permanent status, and x and w, argued below, do not. The testbed offers them as an evaluation set, off by default.

None of the three is derivable from Cyrillic. Kyrgyz Cyrillic writes one а, one х, and one в, so the converter cannot produce ä, x, or w from source text without knowing which words are loans. They are therefore compose-mode letters: available when a person writes Latin directly, never emitted by automatic conversion. The reverse converter accepts them as input and folds each back to the base letter (ä → а, x → х, w → в), which is the §18.1 native-priority rule applied to a new case: the distinction survives in Latin and is lost on the way back, exactly as loan ц and щ are. Search folding unifies them the same way, so the lost distinction does not fragment an index.

ä is the one that does phonological work. The other two refine the spelling of loanwords. ä marks harmony class, and harmony class is what the q rule reads: äka, ükä, and källä are front-harmonic and take velar [k], while aqa and qalla are back-harmonic and take uvular [q]. Without ä the rule cannot tell these apart; with it, the rule has the information it needs. The experiment follows directly: run southern text through q mode with and without ä, and measure whether the second configuration is more accurate.

x is as weak a case as w, and for a more basic reason: the difference it would write is not a contrast between words. Which fricative a speaker uses in haram, halal, or Buhara depends on who is speaking and where, not on the word, so no pair of Kyrgyz words is told apart by [h] against [x]. The loanwords that carry these sounds came from Arabic, Persian, Chagatai, and Russian by different routes, and Kyrgyz reshaped the source consonant as it took each one in. Sometimes it dropped it, as in маала (also маале) from mahalla. Sometimes it made it velar, as in махабат from muhabbat. Sometimes it hardened it into the stop к, as in Ыктыяр against Ихтияр, Ыскак against Исхак, and Аскат against Асхат. Most of these words and names are current in more than one form, and a spelling rule has no business choosing among them. It does not need to: where a fricative survives, Jany-Latyn's h writes it whichever way the speaker says it, as Cyrillic х has for eighty-six years, and where it became к or disappeared, there is nothing left to mark.

x also carries a cost the other two do not. In Latin text of any kind, x already has a ubiquitous international value (Linux, LAX, X Factor, taxi), and a Kyrgyz reader meets those strings constantly. Adopting x for the velar fricative would give one letter two readings that no rule separates, and would drag search folding with it. So the split marks a difference Kyrgyz speakers do not use to tell words apart, and it spends a letter whose international value is worth keeping. The testbed keeps x in the evaluation set so that it can be tried against real text, not because the question is still open.

w fails the same way, because intervocalic [w] is largely the predictable realization of /b/ rather than a contrast, and Principle 3 says predictable allophones do not get letters, the same reasoning that unified k in §8. If w earns a place, loanwords will be what earns it.

The limiting principle for anything added to this set: the sound must distinguish words for its speakers, Cyrillic must be unable to represent it, and it must not be predictable from context. ä passes all three, and is the strongest of the set. x fails the first, since its [h] and [x] vary by speaker and region rather than by word. w fails the third.

Part III — Context

A Latin orthography for Kyrgyz (Кыргыз тили) sits where phonology, typographic engineering, and everyday digital practice meet, and the question is not hypothetical. In September 2024, at a meeting in Baku, the Turkic World Common Alphabet Commission, working under the Turkic Academy and the Organization of Turkic States, agreed on a proposed 34-letter Common Turkic Alphabet (CTA) [1][2]. Kyrgyzstan is the one OTS member that has not decided to move away from Cyrillic [2], and the Kyrgyz linguist who represented the country at those talks, Syrtbai Musaev, has said a 28-letter draft Latin alphabet for Kyrgyz is ready and awaiting a political decision by the president and parliament [2][3]. President Sadyr Japarov's stated position is that it is premature to discuss the transition and that development of the state language should continue in Cyrillic [3].

21. How Jany-Latyn Relates to Existing Systems

21.1 Existing romanizations of Kyrgyz

There is no generally accepted romanization of Kyrgyz; for geographic names the government adopted BGN/PCGN [8]. Four systems are in circulation, and Jany-Latyn relates to each of them differently.

(BGN/PCGN, ALA-LC, ISO 9, and CTA columns after [8].)

BGN/PCGN is the system with official standing in Kyrgyzstan, and Jany-Latyn agrees with it on the two letters that matter most for recognizability: j for ж and ö/ü for the front rounded vowels. It is a transcription for foreign readers rather than an orthography, and it shows: it writes both й and ы as y, so тыйын becomes tyyyn, and its digraphs ch, sh, kh, ng collide at morpheme boundaries. Its strength is that it needs no letters beyond ö and ü; its weakness is that it was never meant to be written by Kyrgyz speakers.

ALA-LC is a library cataloging standard and optimizes for reversibility with combining marks (n͡g, u̇, ĭ). It is the only system that shares Jany-Latyn's instinct about the glide, writing й as ĭ, a mark on i. But its combining diacritics normalize inconsistently and cannot be typed on an ordinary keyboard.

ISO 9 is a mechanical Cyrillic-to-Latin mapping designed to be script-agnostic, not language-specific. It writes й as j and ж as ž, which is precisely backwards from the perspective of a language where ж is [dʒ]. It round-trips perfectly and reads badly, which is a reasonable trade for a bibliographic standard and not for an orthography.

The CTA is treated in §21.2.

Jany-Latyn's differences from all four follow from a different goal. These systems transcribe Kyrgyz into Latin for people who read something else. Jany-Latyn is meant to be read intuitively by native Kyrgyz speakers.

21.2 The Common Turkic Alphabet

The 34-letter CTA agreed in Baku in September 2024 [1][2] is intended as a shared baseline across the Turkic world, with each country adapting it nationally [3]. Jany-Latyn shares the CTA's functional core wherever Kyrgyz phonology supports it (ö, ü, ç, and ş all match). Six letters differ, and each is argued in Part II: й (§10), ж (§11), ы (§9), к and г (§8), and ң (§14).

í where the CTA uses y. The CTA can spend y on the glide because it moves ы to dotless ı; Jany-Latyn has already spent y on ы (§10).

Jany-Latyn starts from current Kyrgyz usage and takes most of the convergence the CTA offers. Three of the CTA's 34 letters have no work to do in Kyrgyz: Ə is a vowel outside the inventory, Ū is the Kazakh ұ, and X is redundant where х is written h. Of the remaining 31, Jany-Latyn takes 26, ç, ş, ö, and ü among them (§17.3). It declines four, c (§11) and q and ğ (§8) for the lexicon dependency and ı for what casing and small type do to it (§9), and writes ŋ where the CTA has ñ (§14); writing ы as y also removes any need for the dotted capital İ. With ŋ in place of ñ and with í added, the one letter Jany-Latyn has that the CTA lacks in any form, the alphabet comes to 28 letters, and the difference is a set of decisions rather than a shortfall. The Kazakh choices were weighed the same way, and ū or ұ and a retained Cyrillic ө were set aside for the shared ü and ö (§12.1). Every declined letter stays in the testbed, whose CTA-aligned preset takes all of them to reach the full 31, so partial or wholesale adoption is a setting rather than a fork, and can be judged against the same text as Jany-Latyn.

Beyond letters, the argument for not adopting the CTA wholesale is structural. Kyrgyz has phonemic vowel length that Turkish lacks; Kyrgyz ж is an affricate where Turkish j is a fricative; Kazakh underwent sibilant shifts Kyrgyz did not. An orthography tuned to its own phonology will agree with a pan-Turkic standard where the phonology agrees and part from it where it does not, which is the pattern above. The CTA itself anticipates this: a participant at Baku noted that a language needing fewer than all 34 letters can adapt the alphabet to itself [3].

There is also a cost that is not linguistic: spelling nearly identical to Turkish, Azerbaijani, or Uzbek removes a signal machines use to tell a smaller language apart. §22 takes that up, together with the case for convergence.

The same words in each

Native vocabulary only, so that nothing here depends on how a converter treats loanwords. Whether q is written is a lexicon question, not an alphabet question.

And in running text:

Cyrillic. Адамдардын бир-бири менен алмашкан каттары, эл аралык иммуногистохимия же башка техникалык журналдарда чыккан изилдөө макалалар, элге белгилүү болгон ырлар жана чыгармалар — баары бир алфавитте жазылат.

Jany-Latyn. Adamdardyn bir-biri menen almaşkan kattary, el aralyk immunogistohimiía je başka tehnikalyk jurnaldarda çykkan izildöö makalalar, elge belgilüü bolgon yrlar jana çygarmalar — baary bir alfavitte jazylat.

q+ğ. Adamdardyn bir-biri menen almaşqan qattary, el aralyq immunogistohimiía je başqa tehnikalyq jurnaldarda çyqqan izildöö maqalalar, elge belgilüü bolğon yrlar jana çyğarmalar — baary bir alfavitte jazylat.

q+ğ+ı. Adamdardın bir-biri menen almaşqan qattarı, el aralıq immunogistohimiya je başqa tehnikalıq jurnaldarda çıqqan izildöö maqalalar, elge belgilüü bolğon ırlar jana çığarmalar — baarı bir alfavitte jazılat.

CTA-aligned. Adamdardın bir-biri menen almaşqan qattarı, el aralıq immunogistohimiya ce başqa tehnikalıq jurnaldarda çıqqan izildöö maqalalar, elge belgilüü bolğon ırlar cana çığarmalar — baarı bir alfavitte cazılat.

The CTA-aligned row is written as a person with a lexicon would write it: native ж as c (ce, cana, cazılat) and loanword ж as j (jurnaldarda), native back-harmonic г as ğ (bolğon, çığarmalar) while the loan keeps plain g. The testbed reproduces every other feature of the row. What it cannot produce is jurnaldarda in a c-mode text, which writes every ж as c and gives curnaldarda; only knowing which words are Russian loans closes that gap.

The passage is built to make the configurations differ visibly rather than to read elegantly, and two words in it repay following across the rows.

• Техникалык shows k against q inside a single word: in the q rows it is tehnikalyq, the loan stem's velar k beside the native suffix's uvular q. A rule that reads the nearest vowel happens to get it right, as it does for the plain g of immunogistohimiía beside native bolğon, but nothing tells the rule that tehnika is a loan, which is why the same heuristic also produces qosmos and parq.
• Immunogistohimiía shows what the systems do with the vocabulary the language keeps acquiring: h for х, ía for the final я, doubled consonants intact, and a word long enough that legibility and sorting both have to hold across it.

21.3 Musaev's historical revival

Syrtbai Musaev, director of the linguistics institute at I. Arabaev Kyrgyz State University and one of Kyrgyzstan's representatives at the 2024 Baku talks, proposes a 28-letter Latin alphabet built on Tynystanov's project, arguing that it preserves the phonemic structure of Kyrgyz completely and therefore raises no linguistic problems [3].

The two proposals agree on more than they disagree on. Both use ö, ü, ç, ş, and both use y for ы [3]. The divergences follow from his being a revival and this being a fresh design: his alphabet restores the k/q split that Tynystanov's had, writes ж as c after Tynystanov and the CTA, and marks the velar nasal with ņ instead of eng. Both mark the same base letter for the same reason, though not by the same device: ņ is n with a cedilla attached beneath, where ŋ and Tynystanov's ꞑ carry the stroke of the letter itself below the baseline. ŋ also has the better font and IPA support of the two.

The more consequential difference is not linguistic. Musaev's proposal has a plausible route to becoming law that this project neither has nor claims. Jany-Latyn's contribution runs the other way. It exists as running software today, engineered against problems a 1920s alphabet was never built to answer: deterministic Cyrillic round-tripping, tokenization, search folding, and a plain ASCII form. The two proposals are not competing for the same job.

21.4 Other domestic proposals

Musaev's is not the only Kyrgyz Latin project. A 30-letter alphabet published at qyrgyz.com, developed with input from historian Tynchtykbek Chorotegin and others, likewise starts from the 1930s Latin alphabet but modernizes its rare letters, replacing Ƣ with Ğ and Ь with Y, and normalizing Ꞑ, Ɵ, and Y to Ñ, Ö, and Ü [12].

Jany-Latyn agrees with it on Y for ы, Ö, and Ü: three points on which three separate projects converged independently. It diverges on Ğ and Ñ, for a different reason in each case. Ğ survives in a revival because Tynystanov's alphabet had Ƣ; it cannot survive Principle 1. Ñ is the closer call, because both projects are modernizing the same letter, Tynystanov's Ꞑ, and disagree about what should replace it: Ñ moves the mark above the letter as a tilde, while Ŋ keeps the descending stroke. What separates the two conclusions is the starting point rather than the phonology, which both projects read the same way.

21.5 What Kazakhstan and Uzbekistan demonstrate

Kazakhstan is the most useful comparison available, because it shows what happens to an alphabet between decree and daily use. The 2017 decree introduced an apostrophe-based alphabet that was widely mocked; in February 2018 a second decree replaced the apostrophes with acute accents [13]; the acute version was itself criticized, in part over the acutes and the digraphs it retained [14]; and in 2021 a further revision settled on 31 letters using umlauts, a macron, a breve, and a cedilla — ä, ö, ü, ū, ğ, ş [5]. Even within 2021 the velar nasal changed from Ŋ to Ñ between the January and April versions [9]. The completion date has moved from 2025 to 2031 [15]. Uzbekistan began its transition in 1993 and it is still not complete [3], with a 2021 draft revision still retaining the digraph ng [9].

Three lessons carry into this document. Letter-level decisions that look settled get reopened, which is an argument for shipping a testbed rather than a decree. Diacritic systems are judged in public on how they look in bulk, and the phonological logic rarely enters the argument. And the gap between adopting an alphabet and writing in it is measured in decades.

Kyrgyzstan has its own version of the third lesson, and it is closer to home: an orthography reform adopted in 2004 was rolled back after the public did not take it up [3].

22. Sociolinguistic and Political Context

22.1 The politics of leaving Cyrillic

Kyrgyz has changed alphabet twice in a century, from Arabic to Latin in 1928 and from Latin to Cyrillic in 1940. The next change, if it comes, is political in origin, not academic. Musaev's summary is that alphabets and orthographies are adopted by resolution of the Jogorku Kenesh and require a presidential decree, and that scholars may only propose [3]. In April 2023, after the chairman of the National Commission on the State Language told parliament that scholars and the public were ready to move to Latin if the political decision were made, the president publicly reprimanded him and stated that the discussion was premature and that development of the state language should continue in Cyrillic [3]. Days later, Russia suspended dairy imports from Kyrgyzstan [18]. However inevitable the move may look to Kyrgyz speakers, it is read as a turn away from Moscow, and the answer arrives as trade policy.

22.2 What a shared Turkic standard costs a smaller language

Moscow's reading of a Latin transition is argued in public constantly; the cost of the alternative is not. Adopting the Common Turkic Alphabet wholesale aligns Kyrgyz with the largest written standard in the family, and the smaller language pays for that alignment in whatever tells it apart from its neighbours on the page.

Cyrillic marks that difference in two ways. Against Turkish and the other Latin-script Turkic languages, the script is itself the label. Against Kazakh, which shares the script, the letters do it: Kazakh writes ә, і, қ, ғ, ұ, and һ, and Kyrgyz writes none of them and nothing Kazakh lacks [4][15], so Kyrgyz is marked only by what is absent. A Latin alphabet taken from a shared template removes the first signal for every Latin-script neighbour, and leaves the second to whichever letters each language keeps of its own.

Short strings are where the loss shows. Titles, names, hashtags, metadata, and queries carry almost no signal, so whatever resolves them falls back on prior probability, and prior probability follows corpus size, which follows the number of people writing the language. Kyrgyz has an order of magnitude fewer speakers than Turkish and fewer than Kazakh, so it is the smaller party in both comparisons this section makes. Jany-Latyn writes кылым (century) as kylym, a string that belongs to no other language; the CTA-aligned setting writes it qılım, which in Azerbaijani is a form of the verb qılmaq (to do). Searched on Google in September 2026, qılım returned that verb, where the Cyrillic кылым returns the Kyrgyz word. A song title containing жаным faces the same fork: janym is Kyrgyz, while canım is Turkish.

Longer text is safer, by a margin that depends on distance. Against Turkish nothing is at risk: a Kyrgyz paragraph stays identifiable whatever the alphabet, because vocabulary and morphology diverge further than spelling does. Against Kazakh the margin is narrower, since the languages are close enough that vocabulary separates them less reliably, and under a shared template a text has to run longer before it identifies itself. That is a fact about tools, not a complaint about kinship. In both cases the effect concentrates in short strings and in automated labeling; §23.2 sets out how to measure it, and why only the comparison with Turkish can be run yet.

The loss compounds where training corpora are built, because a crawl has to be sorted by language before any of it can be used, and whatever does the sorting has only the text to go on. Cyrillic Kyrgyz is separable from Turkish at a glance and from Kazakh by its letters; Latin Kyrgyz shaped like its neighbours is filed as one of them, subtracted from a corpus that is already thin and added as noise to a larger one. A language that is hard to find is inconvenient; a language mislabeled at the point of collection stops accumulating.

Speech recognition and synthesis draw from the same buckets, and where the signal is weakest the error is already audible. Kyrgyz Cyrillic, marked against Kazakh only by the letters it does not use, is thin enough to be taken for Kazakh today: given Kyrgyz lyrics in September 2026, Suno V6-mini sang them to the author with Kazakh values, ж as the fricative [ʒ] in place of the Kyrgyz affricate [dʒ], and the velar and uvular readings of к and г alongside it. A corpus mislabeled at collection returns as an assistant that handles Kyrgyz as an approximation of a neighbour, and children meet that earliest, learning to read against whatever the tools around them treat as correct.

Any move to Latin gives up a signal Cyrillic supplies for free, and the choice among Latin orthographies decides how much of it survives. No Turkic Latin orthography in current use writes ŋ, and none writes í, so jaŋy and kylym carry their own language tag where cañı and qılım do not. That property was not a design goal. It follows from choices made on other grounds, for the glide, the velar nasal, and ы, so it is offered as a prediction to be tested and never as a reason for those choices.

22.3 Related languages, shared letters

Kyrgyz needs no elaboration to be distinct from Turkish: the two sit in different branches of the family and are not mutually intelligible (Kloss's Abstand [19]). Spelling alone does not close that distance. Kazakh is nearer on every measure — the same branch of the family, a shared inherited core, centuries of contact — so less of what separates the two in writing comes from the languages themselves and more of it from their separate standards, of which the alphabet is one part. That is why the margin for identification is narrower there. In either case, a distinction the orthography declines to record loses one of the supports that hold it in place, and over a long enough run that can reach speech.

Kyrgyz has already shared a script with a dominant language for most of a century, and the useful comparison is what that contact did and did not do. Russian norms entered Kyrgyz Cyrillic wherever Russian words did and stopped there, which is why ц, щ, ъ, ь, and iotated е are loan-only letters in Kyrgyz and why a converter can afford to lose them. The contact was not a one-sided loss: it enriched the lexicon heavily, and the rules that arrived with the loanwords stayed in the loan layer, because Russian had no cognates in native Kyrgyz for its rules to act on.

Turkish is not in that position. It shares with Kyrgyz the inherited core of the Turkic lexicon [26], and its conventions are the ones the CTA prescribes, so a written standard shared with it brings rules that apply to native words directly — and the benefit such a standard offers, easier borrowing across the family and a larger pool of shared technical vocabulary, arrives on the same channel as that risk. Where the channel has been open between closely related languages, the cases run one way. The 1933 Soviet reforms moved Ukrainian and Belarusian spelling toward Russian, removing the Ukrainian letter ґ until 1990 [22][23]. Galician written to Spanish conventions carries a standing argument about Castilianization [24], and Scots written in English spelling reads as English misspelled [25].

The everyday mechanism is correction. Spellcheckers and query normalization trained on the larger language would correct the smaller one into it, a model asked for Kyrgyz would fill its gaps with Turkish where the orthography invites it, and writers would normalize word forms that look like errors against a language they read daily. The drift this predicts is in lexicon and written norm first, and in speech only by the route above.

These are tendencies, not certainties. How far a dominant written standard travels depends on more than the kinship between the two languages: on how long the contact runs, on what schooling, media, and policy do in the meantime, and on how much text and tooling each side brings. This document measures none of that. Orthography is one force among several, and none of them decides on its own [28].

22.4 Familiarity and adoption

An alphabet is adopted by people who did not ask for it, and they meet its letters before they meet its arguments. Kazakhstan replaced two alphabets that were judged in public on how they looked, and Kyrgyzstan reversed its own 2004 reform when readers did not take it up (§21.5). A new letter costs little to design and a great deal to teach, and the cost falls on every reader at once.

Measured against what a Kyrgyz reader already knows, several CTA letters are new: c for ж, dotless ı with dotted İ, and ğ, which has no counterpart in any Kyrgyz writing in use since 1940. Each is familiar to a reader of Turkish, which is the familiarity the CTA was built around.

Signage and casual writing are where that is met first, by readers who have had no instruction at all, and a reader meeting an unfamiliar orthography takes in the base letters before the diacritics and reads them with the values already known. In the CTA-aligned setting жарыя (announcement) is written CARIYA, which a reader who has not learned that c is [dʒ] sounds out as кария or сария; neither is the word on the sign, and nothing in the string signals the error. Чыгуу (exit) becomes ÇIĞUU, which that reader takes as CIGUU, the spelling the plain form writes outright. Read with the Cyrillic values of C and I, one of the commonest words on public signs turns obscene. Jany-Latyn's ÇYGUU can mislead by one letter, C for ч; the CTA spelling adds a second, I for ы.

These costs are transitional, and end once the letters are taught; that is the answer a state-backed standard can give. The transition is long, measured in decades in both neighbouring states. An orthography is judged during that interim, and orthography development treats reader acceptance as a design criterion in its own right [27].

Jany-Latyn draws instead on the two kinds of familiarity a Kyrgyz reader already has. One is the Cyrillic they read: í is i with a mark as й is и with one, ŋ keeps the descending stroke of ң, and length is doubled as it is in Cyrillic. The other is the Latin they already type and see: j in the passport spelling Japarov and the place name Jalal-Abad, y in the country's English name, and a plain form that matches how people text now. The two follow a generational line, since readers educated in Cyrillic know the first and readers who grew up on phone keyboards know the second, and the letters are chosen to be recognizable from both sides.

ŋ is where the two part ways. Its familiarity is of shape, to a reader of ң, and less of keyboards: ŋ is on some long-press menus, ñ on nearly all of them. That is why ñ is offered as a comparison setting (§14). Everywhere else, the letter a Kyrgyz reader would guess is the letter the orthography writes.

22.5 The case for convergence, and the position taken

Convergence is the point of the CTA, and mutual legibility with the Turkic languages already written in Latin is a real benefit; it is one of three reasons a Latin alphabet is worth considering here at all. The other two come with any Latin orthography, whatever it aligns itself with: Latin is the script international readers already read, and the script that keyboards, encodings, and text tooling are built around. Only the first is bought by conformity to the CTA, which makes this a question of how much conformity rather than whether Latin.

With Kazakh the benefit is not new. The two already read each other's Cyrillic, and both sent representatives to the Baku talks that agreed on the Common Turkic Alphabet. Kazakhstan has since decreed and dated a Latin transition of its own (§21.5); Kyrgyzstan has not decided to switch. What a transition settles is whether the link survives. If both build on a common Latin base, it is kept; if one moves and the other does not, it lasts only as long as the readers who know both scripts; if both move but diverge further than their Cyrillic did, it thins. Jany-Latyn aims to keep the two languages legible to each other in Latin as they are in Cyrillic, and its engine and testbed can be forked to run the same evaluation for Kazakh. How near the two draw to each other is a matter between sister languages that share their inheritance in both directions. Conventions that come from a third language are a separate question. The CTA's c and ı follow Turkish usage, and they are worth adopting where they hold up against the language-shift risk of §22.3 and the arguments of this document, not by default.

Declining parts of the standard (§21.2) would not make Kyrgyz the exception, because no Turkic state has adopted it whole. Turkmen writes the glide ý for the same structural reason this orthography writes í, and the velar nasal ň where the CTA has ñ [20]; Uzbekistan's draft keeps the digraph ng; and Kazakhstan revised its own letters twice inside 2021. A shared Latin base has never required shared marks above it: the Latin-script languages of Europe, Germanic, Romance, Slavic, Baltic, and Uralic among them, build on the same twenty-six letters and agree on no common set of diacritics above them, and divergent spelling between neighbours is the ordinary state of affairs [21]. The closest precedent is at home. Kyrgyz and Kazakh Cyrillic share a base, each keeps letters of its own, and the two stay legible to each other while remaining distinct on the page, Kazakh by the letters it adds and Kyrgyz by their absence. That relationship is worth keeping in mind for a Latin transition, though it is not a condition on one.

The Common Turkic Alphabet could accommodate every distinction argued in this document. The price is more letters, a maintained loanword lexicon and an institution to keep it, and the teaching cost of shapes no Kyrgyz reader has met. A state-backed standard can pay the first two. The third falls on readers no matter who pays for the rest. Jany-Latyn is built to need none of the first two, and as little of the third as it can.

A language is worth writing in the form it already uses, and a regional standard is worth joining as far as it keeps the smaller language distinguishable, since convergence pursued far enough leaves fewer distinct written languages rather than more legible ones. Those are positions, not empirical results, and where that limit lies is for the tests of §23.2 to measure. Neither is a defence of purity: a large share of everyday Kyrgyz vocabulary is Russian in origin, and whatever follows Cyrillic will bring borrowings of its own. The concern is narrower: a change that could make the language harder to find and easier to file as another, a loss that shows first in machine layers, where it is least visible. The other burden falls in the open, on readers asked to learn an alphabet and conventions they did not choose, and it applies to the letters argued here as much as to any others. Whether either burden is worth accepting, and which way a transition goes, is for the Kyrgyz to decide.

22.6 Costs and constraints

• Cost and capacity. The recurring domestic objection to Latinization is that it is expensive and that education has more urgent needs. The objection is serious and is not addressed by better letter design.
• A multilingual state. Kyrgyzstan is not monolingual. Russian has official status, and Uzbek, Dungan, and other communities write in Cyrillic. A Kyrgyz Latin transition would not move those languages, so Cyrillic would remain in daily use whatever happened to Kyrgyz. Any realistic plan is a plan for coexistence, not replacement, which is also an argument for a converter that runs in both directions.
• Geopolitics. Moscow's reaction on one side and the pull of a shared Turkic standard on the other both weigh on a state-level transition. Neither constrains a tool that needs no decision to be useful.
• Generational asymmetry. Readers educated in Cyrillic and readers raised on phone keyboards start from different letters, which is what the two registers are for: the full form for print and the classroom, the plain form for the phone.

Part IV — Open Questions and Governance

23. Open Questions and Future Work

Some of these are experiments the testbed is built to support; the rest is specification work not yet done. Each experiment states the claim, the test that would settle it, and the result that would count against it.

23.1 Tokenization

Claim: monolithic Latin encoding tokenizes more efficiently than mixed-script text, and comparably to or better than Cyrillic.

Test: take a fixed corpus (Kyrgyz Wikipedia, or FLORES-200 kir_Cyrl for a clean parallel set), render it in Cyrillic, Jany-Latyn, the CTA-aligned setting, and a mixed-script configuration, and measure tokens per word under several tokenizers.

What would refute it: Latin Kyrgyz tokenizing worse than Cyrillic in widely used models. This is a plausible outcome, since Cyrillic Kyrgyz is far better represented in training data than any Latin Kyrgyz, and it would need to be reported if found.

23.2 Language identification and retrieval

Claim: orthographic divergence from the Latin-script Turkic standards keeps Kyrgyz identifiable to machines, and the effect grows as the text gets shorter (§22.2).

Scope: of the two comparisons in §22.2, only the one with Turkish, Azerbaijani, and Uzbek can be tested now, because those languages are written in Latin today. Latin Kazakh, the nearer case (§22.3), exists in decrees and specimen charts and will not be written in quantity before its transition ends in 2031 (§21.5). Transliterating a Cyrillic Kazakh corpus would test the transliterator rather than the language, so the Kazakh comparison stays an argument, and only the third test below reaches it.

Tests: four, in increasing order of cost.

1. Collision census. Count the Kyrgyz word types identical to a Turkish, Azerbaijani, or Uzbek word type under Jany-Latyn and the CTA-aligned setting. It needs only word lists and measures the mechanism directly.
2. Identification against string length. Run an off-the-shelf identifier such as fastText's lid.176 or GlotLID over the same Kyrgyz text in both configurations at 1, 3, 5, 20, and 100 tokens. The prediction is that the configurations agree on paragraphs and separate as strings shorten.
3. Black-box observation. Give the same text in each configuration to deployed systems — a web search engine, an identifier, a speech or song synthesizer — and record which language each one answers in. This is the weakest evidence in this document: the systems are opaque and change without notice, so each observation carries its date and whatever version the system exposes, and says nothing about why the system answered as it did. The Google search and the Suno rendering in §22.2 are observations of this kind.
4. Generative intrusion. Prompt a model for Kyrgyz prose in each configuration and count Turkish and Kazakh lexical items per thousand words, excluding the vocabulary Kyrgyz and Kazakh inherit in common (§22.3).

What would refute it: a negligible collision rate, or identification holding up at three tokens in both configurations, and either would need to be reported if found. The Kazakh comparison has no refutation condition until Latin Kazakh text exists.

23.3 Plain-form mergers

Claim: the words the strip fallback merges (baş with bas, çaí with cai) are few enough in real text that its coherence and folding advantages outweigh them (§19).

Test: count the distinct word types merged by each ASCII fallback (c/s versus ch/sh) over a corpus, and report both numbers alongside the round-trip loss rate by word class.

What would refute it: the strip mapping collapsing a materially larger share of word types, or collapsing frequent pairs that context does not separate.

23.4 Rendering

Claim: inline Cyrillic ө is visually inconsistent with surrounding Latin text, independent of whether the font covers Cyrillic (§12.2).

Test: render kөl and köl across a matrix of common UI, web, and serif fonts, and record for the mid-word glyph the x-height ratio against the adjacent Latin lowercase, the stroke weight, and any baseline offset. Report which fonts substitute and which merely mismatch.

What would refute it: metrics within a few percent of the Latin lowercase across the widely used stacks, with substitution confined to fonts that genuinely lack Cyrillic.

23.5 Input

Claim: the full form can be typed at ordinary speed, with the custom keymap on a desktop layout and with long-presses on an unmodified phone keyboard.

Test: keystrokes and time per 1,000 characters on a standard layout with and without the custom keymap, and on unmodified iOS and Android keyboards, which is also where the ŋ availability of §14 gets checked beyond the one locale it is confirmed in.

What would refute it: a sustained speed penalty against digraph spellings, or letters that cannot be reached at all on a common phone keyboard, which would make ñ the better letter for a future version (§14).

23.6 Specification work not yet done

• A BCP 47 language tag convention (ky-Latn, possibly with a registered variant subtag) so text can declare which system it uses.
• A CLDR transliteration rule set, which would make the mapping available to any CLDR-based platform.
• A stress-notation convention for teaching materials and dictionaries, which must not be the acute, since the acute is the glide (§10).
• Orthographic rules beyond the letter level: hyphenation, abbreviations, the treatment of Russian personal names and surnames, and a policy for whether loanwords are spelled from their Russian source or from Kyrgyz pronunciation.
• Capitalization beyond the core rule. Multi-character graphemes (ía, ío, íu, ts, and the digraph fallbacks) take the case pattern of the source word, so sentence case gives Ía and all-caps gives ÍA; that much is settled. Hyphenated names are settled as well: each part between the hyphens takes its own case, so Ысык-Көл → Ysyk-Köl while Кызыл-кыя → Kyzyl-kyía keeps the lowercase second part the source wrote. What remains is narrower and largely stylistic: acronyms and initialisms carrying suffixes (ЖОЖдо), and camel-case identifiers.
• Whether search folding should fold q to k (Appendix B).
• A machine-readable version marker for converted text (§24), so a document can declare which engine version produced it.
• A documentation-conformance check that infers each example's direction and configuration from its context, rather than from the curated list it relies on now.
• Handwriting and pedagogy: nothing here has been tested with learners, and cedillas and acutes behave differently in handwriting than in type.

24. Governance

This is version 1 of the orthography, published as a proposal rather than a standard, and the project is structured so that disagreement produces changes rather than forks-by-frustration:

• Every contested choice is a toggle in the testbed, with every alternative implemented in code.
• The hosted testbed records which configuration a reader selects, once the reader has answered its consent prompt, so the distribution of preferences across real use can be published alongside the arguments. A reader who declines is counted without cookies or any identifier stored on the device. A self-hosted or classroom deployment records nothing; the instrumentation belongs to the hosted instance, not to the software.
• All of the source is public, under three grants: the conversion engine is MIT and may be used for any purpose; the testbed application is source-available under a noncommercial license, so it can be read, modified, and deployed in a classroom but not sold; this document and the specification are CC BY. Issues, edge cases, and corpus counterexamples are accepted on GitHub, and anyone may review the mapping rules in the specification, propose changes, or fork.
• Versions are numbered and changes are logged. A machine-readable version marker traveling with converted text, so that a document can declare which engine version produced it and be reconverted safely, is a natural next step but is not yet implemented; it is listed in §23.6.

The intended failure mode is falsification, not adoption pressure. A reproducible case where the engine corrupts native Kyrgyz is worth more to this project than agreement.

25. Conclusion

Every choice in this document answers to one constraint ahead of the others: a Jany-Latyn form has to be computable from the Cyrillic that already exists, or the orthography begins by discarding the corpus it is meant to carry. Within that constraint, an orthography succeeds when it balances phonological fidelity, typographic clarity, and typing ergonomics. Today it must also survive contact with search engines, tokenizers, and unmodified keyboards. Those are the standards on which this work should be judged: whether the cedillas and the acute sit inside a line of running Kyrgyz without disturbing it, whether ŋ stays distinct from n at reading size, whether a native speaker meets any ambiguity, and whether the same text survives a round trip, an index, and a phone keyboard. Those questions are better settled on a page of real prose than on a curated example, which is what the testbed is for.

A century after Tynystanov's Latin alphabet, the same historical thread runs through Musaev's draft, and that draft waits on a political decision that may not come soon [3]. This project does not need that decision to arrive. It is built to be useful the moment someone opens the repository, on the terms Kyrgyz speakers have already been writing Latin for years: informally, inconsistently, and constantly.

Linguists, engineers, and policymakers are invited to test these choices rather than take them on trust: open the testbed, read the mapping rules in docs/SPEC.md, and file an issue with any case the engine gets wrong.

Zamanbap, erkin jana tabigyí jazuu egemen elibizge sanariptik doorunda kyzmat kylsyn. (May a modern, open, and natural orthography serve our independent people in the digital era.)

References

[1] "Adoption of Latin-Based Common Turkic Alphabet." The Times of Central Asia, 12 September 2024. https://timesca.com/adoption-of-latin-based-common-turkic-alphabet/

[2] "Turkic States Agree On Common Latin Alphabet, But Kyrgyzstan Happy With Its Cyrillic Script." RFE/RL, 3 October 2024. https://www.rferl.org/a/common-turkic-alphabet-kyrgyz-kazakh-uzbek-turkmen-latin-cyrillic/33137392.html

[3] Жангазиев, Максат. "Нужен ли тюркоязычным странам общий алфавит?" Азаттык Азия (RFE/RL), 25 September 2024. https://www.azattyqasia.org/a/33134603.html — includes a reproduction of Musaev's proposed alphabet chart. Kyrgyz original: https://www.azattyk.org/a/33125096.html

[4] "Kyrgyz alphabets." Wikipedia. https://en.wikipedia.org/wiki/Kyrgyz_alphabets — for the 1928–1938/1940 Latin alphabet correspondence table and the vowel length inventory. Primary source cited there: Сумарокова, О.Л. (2021), Кыргызский алфавит: долгий путь к кириллице, KRSU.

[5] "New version of Latin-based Kazakh alphabet presented at government." Kazinform, 2021. https://www.inform.kz/en/new-version-of-latin-based-kazakh-alphabet-presented-at-government_a3746549

[6] Unicode Technical Standard #10: Unicode Collation Algorithm. https://www.unicode.org/reports/tr10/

[7] Kudo, T. & Richardson, J. (2018). "SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing." EMNLP 2018 (system demonstrations).

[8] "Romanization of Kyrgyz." Wikipedia. https://en.wikipedia.org/wiki/Romanization_of_Kyrgyz — comparative table of ALA-LC, BGN/PCGN, ISO 9, and Common Turkic systems, and government adoption of BGN/PCGN for geographic names. Primary sources cited there: BGN romanization of Kyrgyz (US Board on Geographic Names, 2017); UNGEGN report, 2016; ISO 9:1995.

[9] "What will be the Kyrgyz Latin alphabet." qyrgyz.com. https://qyrgyz.com/post/path-of-kyrgyz-latin-script — for the January-to-April 2021 Kazakh change from Ŋ to Ñ and the 2021 Uzbek draft.

[10] Unicode Standard Annex #29: Unicode Text Segmentation. https://www.unicode.org/reports/tr29/

[11] Unicode Character Database, SpecialCasing.txt (Turkic i/ı case mappings). https://www.unicode.org/Public/UCD/latest/ucd/SpecialCasing.txt

[12] "Кыргызская латиница / Кыргыз тилинин латын алфавити." qyrgyz.com. https://www.qyrgyz.com/kyrgyzskaya-latinitsa

[13] "Kazakhstan Changes Its Alphabet. Again!" Eurasianet, 20 February 2018. https://eurasianet.org/kazakhstan-changes-its-alphabet-again

[14] Bazarbayeva, Z.M. & Chukayeva, T.K. (2021). "The new Kazakh latinized script." Bulletin of L.N. Gumilyov Eurasian National University, Philology Series 3(136). https://bulphil.enu.kz/index.php/main/article/download/291/143/419

[15] "Kazakh alphabets." Wikipedia. https://en.wikipedia.org/wiki/Kazakh_alphabets — 2017 decree and the 2021 postponement of the completion date to 2031.

[16] Unicode Standard Annex #15: Unicode Normalization Forms. https://www.unicode.org/reports/tr15/

[17] "Declaration on the Common Turkic Alphabet." Turkic Academy. https://www.turkicacademy.org/en/smi/declaration-common-turkic-alphabet — for the 34-letter order agreed at Baku. See also "Common Turkic Alphabet", Wikipedia: https://en.wikipedia.org/wiki/Common_Turkic_alphabet

[18] "Russia Suspends Dairy Products From Kyrgyzstan After Calls In Bishkek To Drop Cyrillic Script." RFE/RL's Kyrgyz Service, 21 April 2023. https://www.rferl.org/a/russia-kyrgyzstan-dairy-products-banned-cyrillic-latin/32373802.html

[19] Kloss, H. (1967). "'Abstand Languages' and 'Ausbau Languages'." Anthropological Linguistics 9(7), 29–41.

[20] "Turkmen alphabet." Wikipedia. https://en.wikipedia.org/wiki/Turkmen_alphabet — for the current Latin letters ý [j] and ň [ŋ], and for the earlier ꞑ, ñ, and ng spellings of the velar nasal.

[21] Sebba, M. (2007). Spelling and Society: The Culture and Politics of Orthography Around the World. Cambridge University Press — the standard treatment of orthography as a social rather than purely linguistic object.

[22] Goletiani, Liana (2017). "The Letter that was Missing. On the Return of the Letter Ґ to the Ukrainian Legal Language." Studi Slavistici 14(1). https://doi.org/10.13128/Studi_Slavis-21938

[23] Kiyosawa, Shiori (2021). "Divergence and Conflict between the Two Norms of the Modern Belarusian Literary Language" [in Japanese, with English abstract]. Slavic Studies 68: 1–43. Slavic-Eurasian Research Center, Hokkaido University. http://hdl.handle.net/2115/84289

[24] Ramallo, Fernando (2015). "The Standardization of Galician." Sociolinguistica 29(1). https://doi.org/10.1515/soci-2015-0006

[25] Bann, Jennifer & Corbett, John (2015). Spelling Scots. Edinburgh University Press. https://doi.org/10.3366/edinburgh/9780748643059.001.0001

[26] Johanson, L. & Csató, É.Á. (eds.) (2021). The Turkic Languages, 2nd ed. Routledge. https://doi.org/10.4324/9781003243809 — for the comparative Turkic picture, including the shared inherited lexicon and the origin of Kyrgyz long vowels.

[27] Cahill, M. & Rice, K. (eds.) (2014). Developing Orthographies for Unwritten Languages. SIL International — on acceptability to readers as a criterion in orthography design, alongside phonemic fit.

[28] UNESCO Ad Hoc Expert Group on Endangered Languages (2003). Language Vitality and Endangerment. Document submitted to the International Expert Meeting on the UNESCO Programme Safeguarding of Endangered Languages, Paris, 10–12 March 2003. https://ich.unesco.org/doc/src/00120-EN.pdf — the nine evaluative factors, which the document presents as guidelines to be used together rather than as a ranking, and the Factor 6 scale ("Materials for Language Education and Literacy"), on which an orthography figures only as established, practical, or absent.

Further reading

These are not cited above but bear directly on the arguments made here, and are recommended to anyone extending this work.

• Kara, D.S. Kyrgyz. Lincom Europa, 2003.
• Unseth, P. (2005). "Sociolinguistic parallels between choosing scripts and languages." Written Language & Literacy 8(1).
• Venezky, R. (2004). "In search of the perfect orthography." Written Language & Literacy 7(2).
• Fierman, W. (2009). "Identity, symbolism, and the politics of language in Central Asia." Europe-Asia Studies 61(7).
• "Prospects of Kyrgyz language transition to the Latin script based on experience of Kazakhstan and Uzbekistan." Language Problems and Language Planning (2025). https://www.jbe-platform.com/content/journals/10.1075/lplp.25003.tok
• Petrov, A. et al. (2023). "Language Model Tokenizers Introduce Unfairness Between Languages." NeurIPS 2023.
• Kargaran, A.H., Imani, A., Yvon, F. & Schütze, H. (2023). "GlotLID: Language Identification for Low-Resource Languages." Findings of EMNLP 2023 — the identifier named in §23.2, built for languages in Kyrgyz's position.
• Washington, J.N., Salimzyanov, I. & Tyers, F.M. (2014). "Finite-state morphological transducers for three Kypchak languages." LREC 2014 — relevant to building morphological tooling on top of this orthography.

Appendices

Appendix A. The `e` System

The э/е/glide system is the least intuitive part of the orthography, so the algorithm is stated in full:

1. Forward. Word-initial or post-vocalic е in an iotated loan → íe. Doubled ээ → ee. Word-initial э → e. Post-consonantal е → e. Post-vocalic э in a non-iotated loan → e.
2. Reverse. ee → ээ. Word-initial e → э. Post-consonantal e → е. Post-vocalic e → е. íe per the table in §18.1.
3. Glide. í → й everywhere except in the sequences resolved above. Legacy iy/iyi input (biyik) is accepted on reverse conversion without being Jany-Latyn output.
4. Iotation. ía → я, ío → ё, íu → ю; hiatus without a glide stays plain (diagramma → диаграмма).

SPEC.md states the same rules in its sections 2 (rule 4) and 4.3.

Appendix B. Normalization, Folding, and Collation

The computational requirements an orthography proposal usually omits, and that anyone implementing this one meets in the first afternoon:

• Normalization. The converter's output is Normalization Form C (NFC) [16], the Unicode form that writes a letter and its mark as one codepoint wherever a combined one exists. í, ö, and ü all have such codepoints, and input that arrives split into letter plus mark should be normalized on the way in. ŋ has no decomposition at all, which matters for the next point.
• Search folding. A single foldKey() maps either form to one index key: ö→o, ü→u, í→i, ç→c, ş→s, ŋ→n, then casefold, which is the Unicode operation behind case-insensitive matching. For five of the six full-form letters this is what standard accent-stripping already does, because ç, ş, ö, ü, and í split into a base letter plus a mark under Normalization Form D (NFD), the decomposed counterpart of NFC; ä folds to a the same way. ŋ is the exception among the full-form letters and must be handled explicitly, since no amount of Unicode normalization will fold it to n. So are ñ→n and ı→i, for the configurations that use them.
• Two folds are configuration-dependent. When the compose-mode letters of §20 are enabled, x→h and w→v unify the distinctions that reverse conversion deliberately loses, so a search for haram matches xaram and tav matches taw. When they are not enabled, x and w are ordinary Latin letters in foreign names and must fold to themselves: Linux, LAX, and X Factor are not Kyrgyz words with a velar fricative, and folding them to linuh and lah would be wrong in an index that holds mixed-language content. The fold follows the configuration; it is not a property of the letters.
• The uvular letters fold unevenly. ğ needs no rule of its own: it decomposes to g plus a breve under NFD and strips exactly as ç and ş do, so bolğon and bolgon land on one key. q is a base letter with nothing to strip, so it folds to itself, and qyz and kyz are two keys. An index holding text from both Jany-Latyn and the settings with q needs an explicit q→k rule to match them; whether that rule belongs in foldKey itself, at the cost of merging a distinction the configuration deliberately writes, is an open question.
• Form equivalence holds for the strip fallback. With Jany-Latyn's ASCII mapping, the plain form is the folded form, so çaí and cai land on the same key. Under the transitional digraph fallback the plain form is chai, which folds to a different key unless the index also applies ch→c and sh→s. Any deployment that indexes digraph-style text needs that extra pair of rules; the strip fallback needs nothing.
• Collation. Under the untailored order of the Unicode Collation Algorithm [6], ç sorts as a variant of c and ş as a variant of s, which is not the Kyrgyz alphabetical order. A locale for Jany-Latyn needs a collation tailoring, in the way Turkish tailors ı/i.

Section 5 of SPEC.md gives the folding procedure step by step, with its options.

Appendix C. Collation Tailoring

The rules are given in the rule syntax of International Components for Unicode (ICU), which the Common Locale Data Repository (CLDR), PostgreSQL's ICU collations, Java, and PyICU all accept. & names the letter being anchored to, < marks a primary difference (a distinct letter of the alphabet), and <<< marks a tertiary one (the same letter differing only in case). Read the first line below as: place ç immediately after c as a letter in its own right, and treat Ç as that same letter in upper case. Every other line works the same way.

# Jany-Latyn
&c < ç <<< Ç
&i < í <<< Í
&n < ŋ <<< Ŋ
&o < ö <<< Ö
&s < ş <<< Ş
&u < ü <<< Ü

# Added in the settings with q and ğ
&g < ğ <<< Ğ
&k < q <<< Q

# Added with the ñ setting, which the CTA-aligned setting also uses
&ŋ < ñ <<< Ñ

# Added with the breve or tilde glide options
&í < ĭ <<< Ĭ < ĩ <<< Ĩ

# Added when the extended letters of §20 are enabled
&a < ä <<< Ä

# Replaces the &i rule in the settings with dotless ı; the trailing í applies
# only where the setting keeps the acute glide, since with the y glide
# there is no í to place and Y keeps its default Latin position
&h < ı <<< I < i <<< İ < í <<< Í

Anchoring ç to c instead of placing it after b is deliberate. C is not a letter of the Jany-Latyn alphabet, but it occurs constantly in foreign names, so anchoring ç to it keeps the two sorting adjacently instead of scattering unfamiliar spellings through the list. A letter that substitutes for another (ñ for ŋ, ĭ or ĩ for í) takes the position of the letter it replaces, sorting immediately after it. Because no configuration uses both letters of such a pair, a list sorts identically whichever one is selected. Letters outside the alphabet that are not anchored (w, x, and q or c where the configuration excludes them) keep their default Latin positions. Alphabet membership and collation coverage are different questions: a teaching alphabet may omit a letter that a sort must still place.

Sorting is defined on the full form; plain-form text is ASCII and sorts as such, so ş and s interleave there. Collation and the search folding of Appendix B pull in opposite directions, one separating what the other merges, and a system needs both.

Implementation note

Intl.Collator accepts no custom tailorings, so a JavaScript implementation cannot consume the rules above directly. The reference engine ships a compare(a, b) built from an explicit rank table per configuration, pinned against a checked-in expected ordering; the ICU rules are published for every other platform.

Appendix D. Engine, Keyboard Layouts, and CLI

The conversion engine is pure TypeScript with no runtime dependencies, and runs unmodified on Node.js, in browsers, and on edge workers. Keyboard layouts for macOS and Linux put ö, ü, í, ŋ, ç, and ş on single modified keystrokes, and a command-line tool converts files and streams for corpus work.