boffire/kabyle-g2p-training-data

Kabyle G2P Training Data

Phonetically-annotated Kabyle (Taqbaylit) text corpus for training Grapheme-to-Phoneme (G2P) models. Generated using the orthography2ipa rule-based phonemizer for Kabyle.

---

Dataset Overview

Property	Value
Language	Kabyle (kab) — Afro-Asiatic, Berber
Total pairs	59,462
Source	boffire/kabyle-piper-22khz
Phonemizer	orthography2ipa (dev branch)
IPA standard	Narrow transcription with Kabyle-specific allophony
License	CC0 (inherited from source)

---

Files

File	Format	Description
kab_g2p_train.tsv	TSV (tab-separated)	graphemes\tphonemes
kab_g2p_train.jsonl	JSONL	{"graphemes": "...", "phonemes": "..."}
kab_g2p_train.txt	Plain text	`graphemes

---

Phonological Features

The phoneme annotations reflect the following Kabyle-specific rules:

Consonants

Grapheme	Phoneme	Notes
ɣ	ʁ	Voiced uvular fricative
ɛ	ʕ	Voiced pharyngeal fricative
ḍ	ðˤ	Emphatic interdental
ṭ	tˤ	Emphatic dental
ṣ	sˤ	Emphatic alveolar
ẓ	zˤ	Emphatic alveolar
ṛ	rˤ	Emphatic alveolar trill
k	ç	Spirantized to palatal fricative (blocked after homorganic nasals, /l/, /n/)
g	ʝ	Spirantized to palatal fricative
b	β	Spirantized (blocked after /m/)
d	ð	Spirantized (blocked after /n/, /l/)
t	θ	Spirantized (blocked after /n/, /m/, /l/ word-finally)
x	χ	Voiceless uvular fricative
ɣɣ	qː	Morphological alternation: geminate → stop
xx	χː	Geminate uvular fricative
ww	wː	Geminate labial-velar
yy	jː	Geminate palatal approximant
tt	t͡s	Geminate dental → affricate

Vowels

Grapheme	Default	Backing / Allophony
a	æ	ɑ near emphatics/uvulars
i	i	ɪ in closed syllables
u	u	ʊ in closed syllables
e	ə	Schwa (unstressed)

Gemination

All geminate consonants are transcribed with ː (e.g., qq → qː, ll → lː, mm → mː).

Word Boundaries

The phoneme strings use spaces between phonemes. Words are separated by spaces (not | pipes, unlike VoxCommunis format).

---

Sample Pairs

Graphemes	Phonemes
Yeqqim d igellil kra yedder.	jəqːim ð iʝəlːil çræ jədːər
Bḍan.	βðˤæn
Iɛedda lawan n wuẓu.	iʕədːæ læwæn n wuzˤu
Sami ikemmel awal.	sæmi içəmːəl æwæl
Tella tin i yeẓṛan kra?	θəlːæ θin i jəzˤrˤɑn çræ
Ikad sani leḥḥuɣ.	içæð sæni ləħːuʁ
Yella ḥedd i izemren ad y-iɛiwen?	jəlːæ ħədː i izəmrən æð j iʕiwən
Bezzaf meẓẓiyet, ulamek ara tezweǧ.	βəzːæf məzˤːijəθ ulæməç æræ θəzwəd͡ʒ
Teɛṛeḍ.	θəʕrˤəðˤ
Ur tɛawen ara.	ur θʕæwən æræ

---

Usage

Loading with HuggingFace datasets

from datasets import load_dataset

ds = load_dataset("boffire/kabyle-g2p-training-data", data_files="kab_g2p_train.tsv")
# or
# ds = load_dataset("boffire/kabyle-g2p-training-data", data_files="kab_g2p_train.jsonl")

Loading as plain text

# TSV
with open("kab_g2p_train.tsv", "r", encoding="utf-8") as f:
    for line in f:
        graphemes, phonemes = line.strip().split("\t")

# JSONL
import json
with open("kab_g2p_train.jsonl", "r", encoding="utf-8") as f:
    for line in f:
        obj = json.loads(line)
        graphemes = obj["graphemes"]
        phonemes = obj["phonemes"]

Training a ByT5 G2P model

from transformers import ByT5Tokenizer, T5ForConditionalGeneration

# Prefix format for ByT5 G2P
def format_input(text):
    return f"kab: {text}"

tokenizer = ByT5Tokenizer.from_pretrained("google/byt5-small")
model = T5ForConditionalGeneration.from_pretrained("google/byt5-small")

# Example
inputs = tokenizer(format_input("Yeqqim d igellil kra yedder."), return_tensors="pt")
labels = tokenizer("jəqːim ð iʝəlːil çræ jədːər", return_tensors="pt").input_ids

# Train...

---

Limitations

Spirantization of /k/

The phonemization of the grapheme k requires special attention in Kabyle. The current implementation applies the following rules:

Context	Output	Example
Default (word-initial, intervocalic, post-consonantal except blockers)	ç	akal → æçæl
Blocked after homorganic nasals (n, m), l, f, b, s, r, ḥ, c, ɛ	k	tamellalt → θəməlːælt (final t also blocked after l)
Blocked word-finally after l	k	—

Known edge cases:

• Word-initial /t/ may or may not spirantize — this requires further verification.
• The /k/ → ʃ allophone (reported in some dialectal descriptions) is not implemented; only ç is produced.
• Loanwords with non-native clusters may produce unexpected outputs.

Users training G2P models on this data should be aware that the /k/ spirantization blocking rules are based on the validated specification from July 2026 and may not cover all dialectal variants.

---

Data Provenance

Stage	Source
Original text	Tatoeba, Common Voice, and other Kabyle corpora
Audio	Resampled to 22.05 kHz mono for Piper TTS (kabyle-piper-22khz)
Phonemization	orthography2ipa (dev branch) — rule-based Kabyle G2P with allophony, sandhi, and stress
Validation	Zero errors across 59,462 utterances

---

Citation

If you use this dataset, please cite:

@dataset{kabyle_g2p_2026,
  author    = {Mokraoui, Athmane},
  title     = {Kabyle G2P Training Data},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/boffire/kabyle-g2p-training-data}
}

---

Related Resources

Resource	Link
orthography2ipa (G2P engine)	https://github.com/TigreGotico/orthography2ipa
Source audio dataset	https://huggingface.co/datasets/boffire/kabyle-piper-22khz
Kabyle CV26 (cleaned speech)	https://huggingface.co/datasets/boffire/common-voice-scripted-speech-kab-26
Kabyle TTS (Piper)	https://huggingface.co/datasets/boffire/kabyle-piper-22khz

---

Contact

Athmane Mokraoui (boffire)
Maintainer of Kabyle NLP resources on HuggingFace.

---

Generated on 2026-07-22