Dataset Viewer
Auto-converted to Parquet Duplicate
word
stringlengths
1
15
ipa
stringlengths
1
16
variants
listlengths
1
1
repaired
bool
2 classes
a
æ
[ { "ipa": "æ", "count": 2644 } ]
false
aah
ææh
[ { "ipa": "ææh", "count": 2 } ]
false
aakken
æækːən
[ { "ipa": "æækːən", "count": 1 } ]
false
ababat
æβæβæθ
[ { "ipa": "æβæβæθ", "count": 10 } ]
false
abadaṛ
æβæðɑrˤ
[ { "ipa": "æβæðɑrˤ", "count": 6 } ]
false
abadmintun
æβæðmɪnθun
[ { "ipa": "æβæðmɪnθun", "count": 1 } ]
false
abadu
æβæðu
[ { "ipa": "æβæðu", "count": 2 } ]
false
abaki
æβæçi
[ { "ipa": "æβæçi", "count": 7 } ]
false
abali
æβæli
[ { "ipa": "æβæli", "count": 1 } ]
false
abalṭu
æβæltˤu
[ { "ipa": "æβæltˤu", "count": 1 } ]
false
abalṭun
æβæltˤun
[ { "ipa": "æβæltˤun", "count": 1 } ]
false
abanan
æβænæn
[ { "ipa": "æβænæn", "count": 4 } ]
false
abandu
æβænðu
[ { "ipa": "æβænðu", "count": 2 } ]
false
abanǧu
æβænd͡ʒu
[ { "ipa": "æβænd͡ʒu", "count": 3 } ]
false
abaraɣ
æβæræʁ
[ { "ipa": "æβæræʁ", "count": 2 } ]
false
abareɣ
æβærəʁ
[ { "ipa": "æβærəʁ", "count": 2 } ]
false
abargaz
æβærʝæz
[ { "ipa": "æβærʝæz", "count": 1 } ]
false
abaryir
æβærjir
[ { "ipa": "æβærjir", "count": 1 } ]
false
abatenǧal
æβæθənd͡ʒæl
[ { "ipa": "æβæθənd͡ʒæl", "count": 1 } ]
false
abayrun
æβæjrun
[ { "ipa": "æβæjrun", "count": 1 } ]
false
abazin
æβæzin
[ { "ipa": "æβæzin", "count": 2 } ]
false
abaɣur
æβæʁur
[ { "ipa": "æβæʁur", "count": 1 } ]
false
abaḍi
æβæðˤi
[ { "ipa": "æβæðˤi", "count": 1 } ]
false
abaḍni
æβæðˤni
[ { "ipa": "æβæðˤni", "count": 1 } ]
false
abaḍniw
æβæðˤniw
[ { "ipa": "æβæðˤniw", "count": 1 } ]
false
abaḥan
æβæħæn
[ { "ipa": "æβæħæn", "count": 2 } ]
false
abaṛeɣ
æβɑrˤəʁ
[ { "ipa": "æβɑrˤəʁ", "count": 4 } ]
false
abbas
æbːæs
[ { "ipa": "æbːæs", "count": 2 } ]
false
abbaɣ
æbːæʁ
[ { "ipa": "æbːæʁ", "count": 4 } ]
false
abbu
æbːu
[ { "ipa": "æbːu", "count": 1 } ]
false
abbuc
æbːuʃ
[ { "ipa": "æbːuʃ", "count": 1 } ]
false
abbuh
æbːuh
[ { "ipa": "æbːuh", "count": 7 } ]
false
abbuuuh
æbːuuuh
[ { "ipa": "æbːuuuh", "count": 1 } ]
false
abbuxen
æbːuχən
[ { "ipa": "æbːuχən", "count": 1 } ]
false
abdelkrim
æβðəlçrim
[ { "ipa": "æβðəlçrim", "count": 1 } ]
false
abduz
æβðuz
[ { "ipa": "æβðuz", "count": 1 } ]
false
abeckiḍ
æβəʃçiðˤ
[ { "ipa": "æβəʃçiðˤ", "count": 2 } ]
false
abeddel
æβədːəl
[ { "ipa": "æβədːəl", "count": 2 } ]
false
abehlul
æβəhlul
[ { "ipa": "æβəhlul", "count": 1 } ]
false
abelbul
æβəlβul
[ { "ipa": "æβəlβul", "count": 1 } ]
false
abelheddar
æβəlhədːær
[ { "ipa": "æβəlhədːær", "count": 1 } ]
false
abelluḍ
æβəlːuðˤ
[ { "ipa": "æβəlːuðˤ", "count": 5 } ]
false
abelqas
æβəlqɑs
[ { "ipa": "æβəlqɑs", "count": 1 } ]
false
abeqqa
æβəqːɑ
[ { "ipa": "æβəqːɑ", "count": 2 } ]
false
aberhuc
æβərhuʃ
[ { "ipa": "æβərhuʃ", "count": 1 } ]
false
aberkan
æβərçæn
[ { "ipa": "æβərçæn", "count": 42 } ]
false
abernus
æβərnus
[ { "ipa": "æβərnus", "count": 3 } ]
false
aberqac
æβərqɑʃ
[ { "ipa": "æβərqɑʃ", "count": 1 } ]
false
aberrani
æβərːæni
[ { "ipa": "æβərːæni", "count": 2 } ]
false
aberreḥ
æβərːəħ
[ { "ipa": "æβərːəħ", "count": 1 } ]
false
aberčečču
æβərt͡ʃət͡ʃːu
[ { "ipa": "æβərt͡ʃət͡ʃːu", "count": 1 } ]
false
aberǧeɣlal
æβərd͡ʒəʁlæl
[ { "ipa": "æβərd͡ʒəʁlæl", "count": 1 } ]
false
abessel
æβəsːəl
[ { "ipa": "æβəsːəl", "count": 2 } ]
false
abestuḥ
æβəsθuħ
[ { "ipa": "æβəsθuħ", "count": 5 } ]
false
abesṭuḥ
æβəstˤuħ
[ { "ipa": "æβəstˤuħ", "count": 6 } ]
false
abetṛun
æβəθrˤun
[ { "ipa": "æβəθrˤun", "count": 1 } ]
false
abeɛda
æβəʕðæ
[ { "ipa": "æβəʕðæ", "count": 3 } ]
false
abeɛɛuc
æβəʕːuʃ
[ { "ipa": "æβəʕːuʃ", "count": 1 } ]
false
abeɣli
æβəʁli
[ { "ipa": "æβəʁli", "count": 1 } ]
false
abeɣluf
æβəʁluf
[ { "ipa": "æβəʁluf", "count": 1 } ]
false
abeɣḍi
æβəʁðˤi
[ { "ipa": "æβəʁðˤi", "count": 1 } ]
false
abeḥri
æβəħri
[ { "ipa": "æβəħri", "count": 7 } ]
false
abeṛnus
æβərˤnus
[ { "ipa": "æβərˤnus", "count": 3 } ]
false
abeṛṛani
æβərˤːɑni
[ { "ipa": "æβərˤːɑni", "count": 11 } ]
false
abhim
æβhim
[ { "ipa": "æβhim", "count": 5 } ]
false
abidun
æβiðun
[ { "ipa": "æβiðun", "count": 2 } ]
false
abiskwi
æβɪsçwi
[ { "ipa": "æβɪsçwi", "count": 1 } ]
false
ablaḍ
æβlæðˤ
[ { "ipa": "æβlæðˤ", "count": 4 } ]
false
ablug
æβluʝ
[ { "ipa": "æβluʝ", "count": 2 } ]
false
ablul
æβlul
[ { "ipa": "æβlul", "count": 2 } ]
false
abninan
æβninæn
[ { "ipa": "æβninæn", "count": 1 } ]
false
abraḥ
æβræħ
[ { "ipa": "æβræħ", "count": 1 } ]
false
abrid
æβrið
[ { "ipa": "æβrið", "count": 197 } ]
false
abriki
æβriçi
[ { "ipa": "æβriçi", "count": 1 } ]
false
abrizili
æβrizili
[ { "ipa": "æβrizili", "count": 1 } ]
false
abriḍ
æβriðˤ
[ { "ipa": "æβriðˤ", "count": 1 } ]
false
absis
æβsis
[ { "ipa": "æβsis", "count": 1 } ]
false
abubul
æβuβul
[ { "ipa": "æβuβul", "count": 6 } ]
false
abuchiḍ
æβʊʃhiðˤ
[ { "ipa": "æβʊʃhiðˤ", "count": 1 } ]
false
abudali
æβuðæli
[ { "ipa": "æβuðæli", "count": 2 } ]
false
abudid
æβuðið
[ { "ipa": "æβuðið", "count": 1 } ]
false
abug
æβuʝ
[ { "ipa": "æβuʝ", "count": 1 } ]
false
abugatu
æβuʝæθu
[ { "ipa": "æβuʝæθu", "count": 1 } ]
false
abugaṭu
æβuʝɑtˤu
[ { "ipa": "æβuʝɑtˤu", "count": 21 } ]
false
abuhal
æβuhæl
[ { "ipa": "æβuhæl", "count": 2 } ]
false
abujad
æβuʒæð
[ { "ipa": "æβuʒæð", "count": 5 } ]
false
abujed
æβuʒəð
[ { "ipa": "æβuʒəð", "count": 1 } ]
false
abzim
æβzim
[ { "ipa": "æβzim", "count": 1 } ]
false
abɣas
æβʁæs
[ { "ipa": "æβʁæs", "count": 3 } ]
false
abḍan
æβðˤæn
[ { "ipa": "æβðˤæn", "count": 1 } ]
false
abṛaham
æβrˤɑhæm
[ { "ipa": "æβrˤɑhæm", "count": 1 } ]
false
abṛakan
æβrˤɑçæn
[ { "ipa": "æβrˤɑçæn", "count": 1 } ]
false
abṛasli
æβrˤɑsli
[ { "ipa": "æβrˤɑsli", "count": 1 } ]
false
abṛik
æβrˤiç
[ { "ipa": "æβrˤiç", "count": 1 } ]
false
acacfal
æʃæʃfæl
[ { "ipa": "æʃæʃfæl", "count": 1 } ]
false
acamar
æʃæmær
[ { "ipa": "æʃæmær", "count": 2 } ]
false
acampanzi
æʃæmpænzi
[ { "ipa": "æʃæmpænzi", "count": 1 } ]
false
acangal
æʃænʝæl
[ { "ipa": "æʃænʝæl", "count": 2 } ]
false
acapu
æʃæpu
[ { "ipa": "æʃæpu", "count": 1 } ]
false
acapun
æʃæpun
[ { "ipa": "æʃæpun", "count": 11 } ]
false
End of preview. Expand in Data Studio

KabG2P

A grapheme-to-phoneme pronunciation dictionary for Kabyle (Taqbaylit, kab, Latin script), from the AƔBALU project.

25,634 Kabyle word–IPA pairs recovered by aligning 292,921 tokens across 59,462 sentence pairs at a 99.53% alignment rate, with a 0% ambiguity rate across the entire vocabulary. Every attested word has exactly one IPA reading. It is the phonetics layer underlying agbalu/Matoub-TTS and agbalu/Fadhma-300M, and the reference target for any Kabyle G2P model.

from datasets import load_dataset

ds = load_dataset("agbalu/KabG2P", split="train")

print(ds[4])
# {
#   "word":     "ababat",
#   "ipa":      "æβæβæθ",
#   "variants": [{"ipa": "æβæβæθ", "count": 10}],
#   "repaired": False
# }

Data

split rows source tokens
train 25,634 292,921

Eight entries whose headword falls outside the Kabyle writing system (3d, androïd, mp3, muḥ€nd, rosé, supermarché, xelleṣ̣, ṭeyyeb‟) are present in the source lexicon with correct IPA but are excluded from this release; G2P benchmarking on non-Kabyle words is not meaningful.

Schema

field type description
word str Canonical Kabyle Latin spelling, NFC-normalised
ipa str IPA transcription — the majority pronunciation across aligned tokens
variants list[dict] Frequency-sorted alternative readings: [{"ipa": "...", "count": N}, ...]
repaired bool true where this project restored a character the upstream source dropped

IPA Phoneme Inventory

42 phoneme symbols in the source; 38 appear in 3 or more entries. The full inventory of symbols observed across 292,921 aligned tokens:

category symbols
Vowels æ ɑ ə i u
Stops (plain) b d ɡ k t p
Stops (emphatic)
Fricatives (plain) β ð ʝ ç θ f s z ʃ ʒ x χ ʁ ħ ʕ h
Fricatives (emphatic) ðˤ
Affricates t͡ʃ
Nasals m n ɲ ŋ
Liquids r l
Glides j w
Length mark ː

Phonological Rules

Three conditioned rules are applied and verified against the aligned data:

1. Spirantization by gemination. Singleton b d g k t ḍ → fricatives β ð ʝ ç θ ðˤ; geminates → stops b d ɡ k t dˤ. Count agreement is exact for t, b, and k across 292,921 tokens; off by one for g.

2. Vowel backing beside emphatics and uvulars. aɑ when adjacent to ḍ ṣ ṭ ẓ ṛ q ɣ x. The rule is 92.44% accurate against an 87.19% always-æ baseline — +5.25 pp.

3. i/u laxing is not applied. The only recoverable conditioning environment (closed syllable) scored 75.59% against a 74.99% baseline — below the threshold for inclusion. Attested words carry their verified allophones from this dictionary; out-of-vocabulary words receive the tense vowel.

Benchmark Baseline

The AƔBALU rule-based G2P system achieves 78.18% exact-match rate on this dictionary without any dictionary lookup — purely from spelling rules. This is the floor any learned G2P model must exceed, and the dictionary itself is the training target and ceiling.

Repaired Entries

The upstream generator drops any character it has no rule for, silently producing a shorter IPA string. This release repairs 199 entries, flagged by repaired: true:

character affected words repair
o 128 bobββ repaired to βoβ
ţ U+0163 73 aţanææn repaired to ætːæn (geminate, from corpus evidence)
é, ï 3 left as-is (outside the writing system)

🔴 ţ is attested Kabyle. It is the Dallet-tradition spelling for the spirantised t, occurring 21,058 times in AƔBALU-Text v1 and in 1,407 word types. It is repaired to the geminate on corpus evidence. No audio confirms this mapping — ţ does not appear in Common Voice Kabyle — and the repair is disclosed rather than hidden.

Orthography

All headwords are normalised to canonical Kabyle Latin script under normaliser 1.3.0+rules1.0.0 (81 rules). 199 entries carried Greek homoglyphs (ε γ Σ Γ Ԑ) in the upstream source and were corrected to canonical ɛ ɣ Ɛ Ɣ Ɛ. The repaired field records every affected entry.

Source

field value
Upstream boffire/kabyle-g2p-training-data
Upstream licence CC0-1.0
Sentences aligned 59,462 (59,185 successfully aligned)
Alignment rate 99.53%
Ambiguity rate 0.00%
Normaliser version 1.3.0+rules1.0.0
Repaired entries 199

Citation

@misc{agbalu_kabg2p,
  title  = {KabG2P: a Kabyle grapheme-to-phoneme pronunciation dictionary},
  author = {AƔBALU},
  year   = {2026},
  url    = {https://huggingface.co/datasets/agbalu/KabG2P}
}

Please also cite the upstream source: boffire/kabyle-g2p-training-data.

Licence

CC0-1.0. The upstream source (boffire/kabyle-g2p-training-data) is CC0-1.0. The repairs and normalisation applied by this project are released under the same licence.

Part of AƔBALU, a Kabyle corpus and model collection.

Downloads last month
19