boffire/kabyle-ljspeech-22khz

Kabyle LJSpeech TTS Dataset (22kHz)

A high-quality, LJSpeech-formatted dataset designed for training Text-to-Speech (TTS) models in Kabyle (Taqbaylit) kab, an Amazigh language spoken primarily in northern Algeria and among the kabyle diaspora worldwide.

This dataset contains 59,462 utterances (approximately 40 hours of audio) resampled to 22,050 Hz, 16-bit mono WAV format, making it immediately ready for training VITS-based models (e.g., phoonnx, Piper TTS, or sherpa-onnx).

Dataset Summary

• Language: Kabyle (kab)
• Total Utterances: 59,462
• Estimated Duration: ~38–45 hours
• Audio Format: 22050 Hz, 16-bit PCM, Mono WAV
• Text Format: Raw Kabyle Latin script (Berber Latin alphabet), ready for Grapheme-to-Phoneme (G2P) conversion.
• Sources: Aggregated from Mozilla Common Voice and Tatoeba.

Dataset Structure

(max 10,000 files per directory), the audio files are organized into 6 subdirectories (wavs_00 through wavs_05).

Data Instances

The dataset follows the standard LJSpeech format. The metadata.csv file contains pipe-separated values with no header:

wavs_00/kab_00000.wav|Text in Kabyle Latin script here.
wavs_00/kab_00001.wav|Another sentence in Taqbaylit.
...
wavs_05/kab_59461.wav|The final utterance.

Data Fields

• filename (string): Relative path to the WAV file (e.g., wavs_00/kab_00000.wav).
• text (string): The raw, normalized Kabyle text transcript corresponding to the audio.

Intended Use & Phonemization

This dataset is explicitly prepared for training neural TTS models.

Recommended Phonemizer

For the best results, it is highly recommended to use orthography2ipa (often exposed as the "o2i" backend in tools like phoonnx). This data-driven G2P engine handles Kabyle-specific context-aware rules (e.g., occlusivization, specific consonant mappings) far more accurately than generic phonemizers like eSpeak or Epitran.

Example Usage with phoonnx

# Configuration for phoonnx train_vits.ipynb
LANG = "kab"
PHONEMIZER = "o2i"          # Native orthography2ipa backend
ALPHABET = "ipa"
HF_DATASET = "boffire/kabyle-ljspeech-22khz"
SAMPLE_RATE = 22050

Dataset Creation

Curation Rationale

Kabyle is a low-resource language in the NLP and TTS domains. This dataset was curated to provide a clean, standardized, and sufficiently large corpus to train robust, natural-sounding TTS models from scratch, without relying on cross-lingual transfer from unrelated languages.

Source Data

• Mozilla Common Voice: Crowd-sourced, validated Kabyle voice clips.
• Tatoeba: High-quality, community-translated Kabyle sentences paired with synthetic or cleaned audio.

Annotation Process

The text transcripts are the original, validated strings from the source datasets. No manual phonetic annotation was applied; G2P is intended to be handled dynamically during the TTS preprocessing pipeline.

Considerations for Using the Data

Social Impact

Training TTS models for Kabyle helps preserve and promote the language in the digital age, improving accessibility for native speakers through screen readers, voice assistants, and educational tools.

Limitations and Bias

• Accent Diversity: As with many crowd-sourced datasets, the audio may contain a mix of regional Kabyle accents (e.g., Tizi Ouzou, Béjaïa, Bouira).
• Audio Quality: Some utterances from Common Voice may contain minor background noise or varying microphone qualities, though validation steps were taken by the original platforms to minimize this.
• Text Normalization: The text is provided in raw form. Numbers, dates, or URLs should be expanded into spoken Kabyle words during your preprocessing pipeline if your chosen G2P tool does not handle them automatically.

Additional Information

Dataset Curators

• Prepared and formatted by: boffire

Licensing Information

• The audio and text are derived from Mozilla Common Voice (typically licensed under CC0 1.0) and Tatoeba (typically CC0 1.0 or CC-BY 4.0). Users should verify the specific licenses of the underlying source sentences if using this dataset for commercial purposes.

Citation

If you use this dataset in your research or project, please consider citing the original sources:

@misc{commonvoice,
  author = {Ardila, R. and Branson, M. and Davis, K. and Henretty, M. and Kohler, M. and Meyer, J. and Morais, R. and Saunders, L. and Tyers, F. M. and Weber, G.},
  title = {Common Voice: A Massively-Multilingual Speech Corpus},
  year = {2020},
  publisher = {arXiv},
  url = {https://arxiv.org/abs/1912.06670}
}