taqbaylit/f5tts-kabyle-dataset

F5-TTS Kabyle Dataset

Clean, deduplicated audio-text dataset for Kabyle (Taqbaylit / Tamaziɣt) TTS fine-tuning with F5-TTS.

Statistics

Source Datasets

Preparation Pipeline

1. 1.Resampled both sources to 24 kHz mono (F5-TTS / Vocos requirement)
2. 2.Duration filter: 1.0s — 15s (removed 230 clips: 225 Tatoeba + 5 CV)
3. 3.Character standardization: corrected false friends (ε→ɛ, γ→ɣ, etc.)
4. 4.Deduplication: one instance per phrase, preferring CV over Tatoeba for speaker diversity
5. 5.Balanced: ~68% Tatoeba / ~32% CV

File Format

metadata.csv (separator |)

audio_file|text
wavs/tatoeba_033368.wav|Uɣeɣ tannumi d uzɣal.
wavs/cv_024698.wav|Tanemmirt i Mass Nagata.
wavs/tatoeba_022648.wav|Tom iteṭṭef dima deg awal.

vocab.txt

Extended vocabulary merging F5-TTS base vocab + all Kabyle characters:

• —ɛ Ɛ ɣ Ɣ č Č ǧ Ǧ
• —ṭ Ṭ ḍ Ḍ ṛ Ṛ ẓ Ẓ
• —ḥ Ḥ ṣ Ṣ ḏ Ḏ ṯ Ṯ

Usage

Load as audio dataset

from datasets import load_dataset

# Load with audio
ds = load_dataset("audiofolder", data_dir=".", split="train")

# Or load metadata only
df = pd.read_csv("metadata.csv", sep="|")

For F5-TTS fine-tuning

# Prepare F5-TTS format
python prepare_csv_wavs.py ./f5tts_kabyle_dataset ./data/kabyle_char

# Extend checkpoint
python extend_checkpoint.py \
    --checkpoint ckpts/F5TTS_Base/model_1200000.safetensors \
    --old_vocab ckpts/F5TTS_Base/vocab.txt \
    --new_vocab ./data/kabyle_char/vocab.txt \
    --output ckpts/F5TTS_Base/model_1200000_extended.safetensors

# Fine-tune
accelerate launch --mixed_precision=fp16 \
    src/f5_tts/train/train.py \
    --config-name F5TTS_Kabyle.yaml

Character Coverage

All standard Kabyle characters are present:

Dataset Quality

• —Zero duplicates (59,462 unique phrases)
• —Multi-speaker (CV provides voice diversity)
• —Clean audio (Tatoeba TTS + CV scripted speech)
• —Verified alignment (5 random samples manually checked)
• —Standardized orthography (no Greek/Cyrillic false friends)

Limitations

• —Tatoeba is single-speaker (synthetic/reading voice)
• —CV-tiny is a subset of full Common Voice 26 Kabyle
• —Average clip length is short (2.50s) — good for alignment, less for prosody
• —No speaker labels in metadata (for multi-speaker conditioning)

Future Work

• —[ ] Add speaker embeddings for multi-speaker TTS
• —[ ] Integrate full CV26 dataset (~23h additional)
• —[ ] Add phonetic transcription (IPA) for better pronunciation
• —[ ] Create train/dev/test splits with speaker disjointness

Citation

If you use this dataset, please cite:

@dataset{f5tts_kabyle_2026,
  author = {MOKRAOUI, Athmane (boffire)},
  title = {F5-TTS Kabyle Dataset},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/boffire/f5tts-kabyle-dataset}
}

And the source datasets:

• —Tatoeba Project
• —Mozilla Common Voice

License

• —Tatoeba portion: CC BY 4.0
• —Common Voice portion: Mixed (see original dataset licenses per clip)
• —This compilation: CC BY 4.0

Contact

• —Author: Athmane MOKRAOUI (boffire / ButterflyOfFire)
• —GitHub: athmanemokraoui
• —HuggingFace: boffire

Prepared for F5-TTS fine-tuning on Google Colab T4 / Kaggle T4 (16 GB VRAM)