Which sound follows which
The 111 IPA sounds are the vertices, laid out in a spiral. Pick one of the eleven languages and a sound: the edges show which sounds follow or precede it inside a word, and the list says which never occur next to it.
How it was made
- Corpus. Spanish, English, French, German, Russian and Mandarin: the 20,000 most frequent words in the OpenSubtitles (2018) frequency lists that have a pronunciation. Latin: the 20,000 most frequent forms in ten classical authors from the Perseus Digital Library. Arabic: the 14,000 Wiktionary entries with a pronunciation (via WikiPron), frequent ones first. Swahili: the words of about a thousand Swahili Wikipedia articles, in lower case. Japanese: the common words of the EDICT dictionary. Tuyuca: the 900 distinct forms of its Intercontinental Dictionary Series list (fieldwork by Barnes and Malone). Counts are by distinct word: an edge of 300 means 300 distinct words contain that sequence.
- The corpora are not alike: Tuyuca has about twenty times fewer words (what is openly documented), Latin is literature, Arabic and Japanese are dictionaries and Swahili is an encyclopedia. Density, the “never” lists and the rare edges depend on corpus size; compare with care.
- Peninsular Spanish: rules from spelling, with the [θ]/[s] distinction, yeísmo (ll = y → [ʝ]) and [x] for j and ge/gi; [β ð ɣ] for /b d g/ except word-initially, after a nasal and /d/ after /l/; nasal assimilation; /s θ/ voiced before a voiced consonant; glides [j w]. Most proper names, English words and acronyms from the subtitles are filtered out.
- English: CMU Pronouncing Dictionary (General American), with the flap [ɾ] of water and the glottal stop [ʔ] of button. Aspiration and dark l are not marked.
- German: Wiktionary's pronunciations (german-ipa-dict). Length (aː → a), the glottal stop before a vowel ([ʔ]apfel, which the other sources never write) and syllabicity marks (-en = [n m ŋ], -el = [l]) are removed. Vocalized r is [ɐ]. Words with sounds borrowed from English or French are left out.
- Latin: restored classical pronunciation, by rule: c = [k], v = [w], qu = [k w], x = [k s], gn = [ŋ n], b before s/t = [p], ae oe au = [a j] [o j] [a w], consonantal i = [j] (doubled between vowels: maior = [majjor]); ph th ch = [p t k]. The texts do not mark vowel length and it is not modelled.
- French, Russian and Arabic: Wiktionary's pronunciations as collected by WikiPron (French and Arabic broad, Russian narrow, with its vowel reduction [ɪ ə ɐ]). The shared rules below merge what the chart does not tell apart: French nasal vowels with oral ones (ɑ̃ → ɑ), Russian palatalized consonants with plain ones (rʲ → r), Arabic emphatics with plain ones (tˤ → t). Arabic shadda and Russian long consonants count twice. Words with a sound used by fewer than 15 words of the language (loans) are left out.
- Swahili: rules from its nearly phonemic spelling: ch = [t ʃ], j = [d ʒ], ny = [ɲ], ng' = [ŋ], ng = [ŋ ɡ]; the prenasalized mb, nd, ng are two vertices. The implosive pronunciation of b d j g that some speakers have is not modelled.
- Tuyuca: the IDS's segmented phonemic forms, made phonetic with the nasal harmony of the Tucanoan languages: in a morpheme with nasal vowels, b d g and r sound [m n ŋ n] (the form “bãdã” is said [mãnã]). Then the vowel's nasality is removed like the other marks. Its “letters” are the source's graphemes. Tuyuca was chosen over Pirahã because only about 200 Pirahã words are openly available, against more than a thousand for Tuyuca.
- Mandarin: CC-CEDICT pinyin turned into IPA by rule; its “letters” are the pinyin. Tones and aspiration are removed: pinyin b and p are both [p], d and t [t], g and k [k], z and c [ts], zh and ch [ʈʂ], j and q [tɕ], so those pairs merge into one vertex. The “empty” i of zi ci si is [ɹ] and that of zhi chi shi ri [ɻ]. Glides [j w ɥ].
- Japanese: EDICT's kana reading turned into IPA by rule; its “letters” are the kana. u = [ɯ], r = [ɾ]; si chi tsu hi fu = [ɕi tɕi tsɯ çi ɸɯ]; kya = [k j a] (palatalization is not a vertex); ん assimilates ([m n ŋ ɲ], [ɴ] at the end); っ doubles the next consonant; a long vowel (ー, ou) counts as one vowel.
- In all of them: one vertex per segment. Diacritics are not chart vertices and are removed (aspiration, palatalization, pharyngealization, nasality, tone); long vowels count once and long consonants twice. What is one sound but not one chart symbol (affricates such as tʃ and ts, diphthongs such as eɪ or aɪ) is split into consecutive vertices and creates edges. The glides [j w ɥ] count as consonants. Only transitions inside a word are counted; the start and end of a word are given separately.
- A transition is rare when fewer than 3 words carry it and, in the languages with frequencies, none of them is among the 1,000 most frequent. Rare ones are drawn dashed: they are usually loans, compounds or corpus errors. “Never” means zero words in this corpus, not that it is impossible in the language.
- Letters: beside each sound go the letters (or kana, or notation) that most often write it. In English and German they come from aligning each word with its pronunciation; words that do not align cleanly do not count for the letters. When letters write a sound split into two vertices, the whole sound is given (‹ch›, part of tʃ).
- Statistics: computed on each language's directed graph, by default without the rare transitions. The search for Hamiltonian cycles and paths has a step limit; when there can be no cycle, the page gives the proof (a group of sounds with fewer possible successors or predecessors than sounds, Hall's condition), not just “not found”.
Sources
- CMU Pronouncing Dictionary, Carnegie Mellon University (BSD licence).
- FrequencyWords, by Hermit Dave, from OpenSubtitles 2018 (CC BY-SA 4.0): Spanish, English, French, German, Russian, Arabic and Chinese.
- WikiPron (CUNY), Wiktionary pronunciations for French, Russian and Arabic, and german-ipa-dict, from Wiktionary, via ipa-dict (CC BY-SA).
- CC-CEDICT (CC BY-SA 4.0) and EDICT2 by the Electronic Dictionary Research and Development Group (CC BY-SA 4.0).
- Perseus Digital Library, canonical-latinLit (CC BY-SA 4.0), and the Swahili Wikipedia (CC BY-SA 4.0).
- Intercontinental Dictionary Series, Lexibank's segmented version (CC BY 4.0): the Tuyuca list by Janet Barnes and Terry Malone.
- The counts on this page are derived from those sources and shared under the same licence.