boffire/tatoeba-kabyle-audio

Tatoeba Kabyle Audio Dataset

A clean, standardized audio-text dataset for Kabyle (Taqbaylit) automatic speech recognition, extracted from the Tatoeba Project and rigorously orthographically corrected.

Dataset Description

This dataset contains 47,789 Kabyle sentences with audio recordings (~25.78 hours total) sourced from Tatoeba. All transcriptions have been standardized to use correct Kabyle Latin characters, replacing visually similar false friends from Greek, Cyrillic, and other Latin scripts that commonly contaminate Kabyle text corpora.

Character Standardization

Kabyle uses the following Latin characters in its standard orthography:

• ɛ (U+025B) / Ɛ (U+0190) — open-mid front unrounded vowel
• ɣ (U+0263) / Ɣ (U+0194) — voiced velar fricative
• č (U+010D) / Č (U+010C) — voiceless postalveolar affricate
• ǧ (U+01E7) / Ǧ (U+01E6) — voiced postalveolar affricate
• ḥ (U+1E25), ḍ (U+1E0D), ṛ (U+1E5B), ṭ (U+1E6D), ṣ (U+1E63), ẓ (U+1E93) — emphatic/pharyngeal consonants

These are frequently confused with visually similar characters in raw web data. This dataset applies the following corrections:

False Friend	Correct Kabyle	Unicode	Origin
ε (Greek epsilon)	ɛ	U+025B	Greek
Σ (Greek sigma)	Ɛ	U+0190	Greek
γ (Greek gamma)	ɣ	U+0263	Greek
Γ (Greek Gamma)	Ɣ	U+0194	Greek
Ԑ (Cyrillic rev. Ze)	Ɛ	U+0190	Cyrillic
ԑ (Cyrillic rev. ze)	ɛ	U+025B	Cyrillic
З (Cyrillic Ze)	Ɛ	U+0190	Cyrillic
з (Cyrillic ze)	ɛ	U+025B	Cyrillic
Ǝ (Latin reversed E)	Ɛ	U+0190	Latin
ǝ (Latin turned e)	ɛ	U+025B	Latin
ż / Ż (Polish)	ẓ / Ẓ	U+1E93 / U+1E92	Latin
ṙ / Ṙ (Irish)	ṛ / Ṛ	U+1E5B / U+1E5A	Latin
ṫ / Ṫ	ṭ / Ṭ	U+1E6D / U+1E6C	Latin
ṡ / Ṡ (Irish)	ṣ / Ṣ	U+1E63 / U+1E62	Latin
ḋ / Ḋ (Irish)	ḍ / Ḍ	U+1E0D / U+1E0C	Latin
ḣ / Ḣ	ḥ / Ḥ	U+1E25 / U+1E24	Latin
ċ / Ċ (Maltese)	č / Č	U+010D / U+010C	Latin
ć / Ć (Polish/Croatian)	č / Č	U+010D / U+010C	Latin
ç / Ç (French)	č / Č	U+010D / U+010C	Latin
ĉ / Ĉ (Esperanto)	č / Č	U+010D / U+010C	Latin
ġ / Ġ (Maltese)	ǧ / Ǧ	U+01E7 / U+01E6	Latin
ǥ / Ǥ (Skolt Sami)	ǧ / Ǧ	U+01E7 / U+01E6	Latin
ğ / Ğ (Turkish)	ǧ / Ǧ	U+01E7 / U+01E6	Latin
ĝ / Ĝ (Esperanto)	ǧ / Ǧ	U+01E7 / U+01E6	Latin

Additionally, Unicode NFC normalization is applied to resolve precomposed vs. decomposed forms (e.g., z + combining dot below → ẓ).

3,643 sentences (7.62%) required at least one character fix.

Audio Statistics

Metric	Value
Total clips	47,789
Total duration	25.78 hours
Sampling rate	16 kHz mono
Mean duration	1.94 seconds
Median duration	1.82 seconds
Min duration	0.73 seconds
Max duration	12.65 seconds

The audio consists of short, read-aloud sentences typical of the Tatoeba project. The tight duration distribution (most clips are 1–3 seconds) makes this dataset well-suited for Wav2Vec 2.0 CTC training with minimal memory overhead.

File Sizes

Split	Size
train	~807 MB
validation	~43 MB
Total download	~803 MB
Total dataset	~851 MB

Splits

Split	Examples
train	45,399
validation	2,390

Features

• sentence_id: Tatoeba sentence ID
• text_raw: Original transcription from Tatoeba (before cleaning)
• text: Cleaned, standardized Kabyle transcription
• audio: Audio waveform (16 kHz, mono)
• character_fixed: Boolean flag indicating whether a false-friend fix was applied

Usage

from datasets import load_dataset

ds = load_dataset("boffire/tatoeba-kabyle-audio")
sample = ds["train"][0]
print(sample["text"])
# → standardized Kabyle sentence

Citation

If you use this dataset, please cite the Tatoeba Project:

@misc{tatoeba,
  title = {Tatoeba: Collection of sentences and translations},
  howpublished = {\url{https://tatoeba.org}},
}

License

Audio files and sentences from Tatoeba are licensed under CC BY 4.0 (or compatible individual licenses as noted in the original Tatoeba metadata).

Acknowledgments

• Tatoeba Project contributors
• Dataset prepared by boffire