Data release v4

Data for uk-tts-frontend/deploy/install.sh.

file	what
lexicon-serving.tar	the lexicon the stress API reads (2,662,501 readings; stress_lookup, active_dataset, import_run, contextual_stress_candidate), gzip CSV per table with a row-by-row digest in MANIFEST.tsv; import with wiki-stress/db/lexicon_archive.sh import
lexicon-assets.tar.gz	the API lookup tables: suffix table, dictionary defaults, counted-form nouns (unchanged from data-v1)

Changes from data-v3:

• Lexicon forms re-stressed as narrators say them (new datasets 18 and 19). The stress the audiobook miner heard on every lexicon word (11.4 million occurrences in 378 books) was compared with the lexicon; 1,575 forms are consistently heard at a vowel no dataset gives. The safe ones are applied without a manual verdict: heard 10+ times in 3+ books, 90%+ of occurrences agreeing, one reading in the lexicon, heard in lower case and under 10% as a name. Dataset 18 (audio-corrections-v1, 118 forms, exclusive) replaces a reading the narrators never use ("подоро́жі" → "по́дорожі", "ко́рисне" → "кори́сне", "пе́редам" → "переда́м", "смі́ливі" → "сміли́ві"); dataset 19 (audio-corrections-add-v1, 66 forms) adds the heard reading beside one they use rarely ("сті́ну" beside "стіну́"), and context decides. The other forms (names, rarer words, disputed readings) wait for a manual review and a later release.
• Served by default: SUPPLEMENTARY_DATASETS=5,7,9,10,12,13,16,18,19, REVIEWED_EXCLUSIVE_DATASETS=13,18 (the Compose defaults now include 16, which they had missed in data-v3).

Measured on the running service with this data: words the lexicon lacks, stress heard in 3,018 Common Voice recordings 81.88%; 200 most frequent ambiguous forms of Common Voice 95.88%; modern-text gold 97.94% (data-v3: 97.97%, one word: "ко́рисною", where both stresses are standard); lang-uk words / heteronyms 96.07% / 84.62% (data-v3: 96.13%; the six words that differ are variants: "за́грава", "зе́рно", "стіну́", "зиму́", "сокі́в").

Built from the Ukrainian Wiktionary (CC BY-SA 4.0), ukrainian-word-stress (MIT), the lang-uk heteronyms dictionary, reviewed corrections, and stress heard in recordings (word forms and their stress only; no audio or text). Checksums in SHA256SUMS.

Lexicon data v3

Data for uk-tts-frontend/deploy/install.sh.

file	what
lexicon-serving.tar	the lexicon the stress API reads (2,662,237 readings; stress_lookup, active_dataset, import_run, contextual_stress_candidate), gzip CSV per table with a row-by-row digest in MANIFEST.tsv; import with wiki-stress/db/lexicon_archive.sh import
lexicon-assets.tar.gz	the API lookup tables: suffix table, dictionary defaults, counted-form nouns (unchanged from data-v1)

Changes from data-v2:

• Dataset 16 replaces dataset 15: 99,470 out-of-lexicon word forms (was 51,298) with the stress heard in audiobook narration and YouTube speech. New in this version: 48,470 words heard only on YouTube were classified by an LLM (standard word, name, slang, surzhyk, or misheard); misheard and undecided words are left out. Served by default (SUPPLEMENTARY_DATASETS=5,7,9,10,12,13,16).

Measured on the running service with this data: words the lexicon lacks, stress heard in 3,018 Common Voice recordings 81.05% (data-v2: 80.19%); 200 most frequent ambiguous forms of Common Voice 95.88%; modern-text gold 97.97% (data-v2: 98.07%, four words differ); lang-uk words / heteronyms 96.13% / 84.62%.

Built from the Ukrainian Wiktionary (CC BY-SA 4.0), ukrainian-word-stress (MIT), the lang-uk heteronyms dictionary, reviewed corrections, and stress heard in recordings (word forms and their stress only; no audio or text). Checksums in SHA256SUMS.

Lexicon data v2

Data for uk-tts-frontend/deploy/install.sh.

file	what
lexicon-serving.tar	the lexicon the stress API reads (2,614,145 readings; stress_lookup, active_dataset, import_run, contextual_stress_candidate), gzip CSV per table with a row-by-row digest in MANIFEST.tsv; import with wiki-stress/db/lexicon_archive.sh import
lexicon-assets.tar.gz	the API lookup tables: suffix table, dictionary defaults, counted-form nouns (unchanged from data-v1)

Changes from data-v1:

• Added dataset 15: 51,298 out-of-lexicon word forms (names, rare and new words) with the stress heard in audiobook narration and YouTube speech, pooled over occurrences and checked for spelling; served by default (SUPPLEMENTARY_DATASETS=5,7,9,10,12,13,15).
• Removed the datasets the API does not serve (the superseded first import and experimental ones): 2,996,968 -> 2,614,145 readings. What the API answers is unchanged; the archive is 79 MB instead of 91 MB.

Measured on the running service with this data: 200 most frequent ambiguous forms of Common Voice 95.88%, modern-text gold 98.07%, lang-uk words / heteronyms 96.10% / 84.62%.

Built from the Ukrainian Wiktionary (CC BY-SA 4.0), ukrainian-word-stress (MIT), the lang-uk heteronyms dictionary, reviewed corrections, and stress heard in recordings (word forms and their stress only; no audio or text). Checksums in SHA256SUMS.

Lexicon data v1

Data for uk-tts-frontend/deploy/install.sh.

file	what
lexicon-serving.tar	the lexicon the stress API reads (2,996,968 readings; stress_lookup, active_dataset, import_run, contextual_stress_candidate), gzip CSV per table with a row-by-row digest in MANIFEST.tsv; import with wiki-stress/db/lexicon_archive.sh import
lexicon-assets.tar.gz	the API lookup tables: suffix table, dictionary defaults, counted-form nouns

Built from the Ukrainian Wiktionary (CC BY-SA 4.0), ukrainian-word-stress (MIT), the lang-uk heteronyms dictionary, and reviewed corrections. Checksums in SHA256SUMS.