boffire/common-voice-scripted-speech-kab-26-tiny

Common Voice Scripted Speech 26.0 - Kabyle (Cleaned)

This is a cleaned, speaker-disjoint subset of Mozilla Common Voice 26.0 for Kabyle (Taqbaylit) ASR.

Source

• Original: Mozilla Common Voice 26.0 (cv-corpus-26.0-2026-06-12)
• Dataset ID: cmqim4fux00tynq07ljtyhzfh (Mozilla Data Collective)
• License: CC0-1.0
• Generated: 2026-07-12

Cleaning Pipeline

Step	Input	Output	Filter
Quality filter	609,940	573,073	≥2 upvotes, 0 downvotes
Character cleaning	573,073	573,073	29 false friend types fixed
Language filter (GlotLID v3)	573,073	567,727	kab_Latn ≥ 0.95
Speaker cap	567,727	567,727	Max 50 clips per speaker
Speaker-disjoint splits	567,727	31,128	No speaker overlap

Character Cleaning Details

• Greek epsilon (ε) → Latin epsilon (ɛ)
• Greek gamma (γ) → Latin gamma (ɣ)
• Cyrillic lookalikes → Canonical Kabyle forms
• French/Spanish/Turkish diacritics → Base Latin or Kabyle equivalents
• Modified sentences: 12,923 (2.26%)

Language Filtering (GlotLID v3)

• Model: cis-lmu/glotlid (model_v3.bin)
• Kabyle label: __label__kab_Latn
• Confidence threshold: ≥ 0.95
• Non-Kabyle rejected: 5,346 clips

Statistics

Split	Clips	Hours	Speakers	Avg Clip Duration
Train	24,944	~22.9	1,064	~3.3s
Dev	3,001	~2.8	133	~3.3s
Test	3,183	~2.9	133	~3.3s
Total	31,128	~28.5	1,330

Dataset Features

• audio: Audio(sampling_rate=16000, decode=True)
• transcription: string (Kabyle text)
• client_id: string (hashed speaker ID)

Usage

from datasets import load_dataset

# Load full dataset
ds = load_dataset("boffire/common-voice-scripted-speech-kab-26")

# Or load specific split
train = load_dataset("boffire/common-voice-scripted-speech-kab-26", split="train")
dev = load_dataset("boffire/common-voice-scripted-speech-kab-26", split="dev")
test = load_dataset("boffire/common-voice-scripted-speech-kab-26", split="test")

# Access sample
sample = train[0]
print(sample["transcription"])  # Kabyle text
print(sample["audio"]["array"])  # Audio waveform (16kHz)
print(sample["client_id"])  # Speaker ID (hashed)

Audio Format

• Sampling rate: 16,000 Hz (Whisper-compatible)
• Original format: MP3
• Decoded: float32 arrays via HuggingFace Datasets
• Average duration: ~3.3 seconds

Speaker Information

• Speaker IDs: Hashed UUIDs (privacy-preserving)
• Speaker-disjoint: No speaker appears in more than one split
• Gender distribution: See original Common Voice metadata (not included in this subset)
• Age distribution: See original Common Voice metadata (not included in this subset)

Limitations

• Read speech only (not spontaneous/conversational)
• Average clip length ~3.3s (short utterances)
• No demographic metadata in this release (speaker IDs only)

Citation

@dataset{common_voice_kabyle_26,
  title = {Common Voice Scripted Speech 26.0 - Kabyle (Cleaned)},
  author = {Mozilla Foundation / ButterflyOfFire},
  year = {2026},
  url = {https://huggingface.co/datasets/boffire/common-voice-scripted-speech-kab-26}
}

Acknowledgments

• Mozilla Common Voice contributors and validators
• Kabyle language community on Common Voice
• GlotLID v3 (cis-lmu/glotlid) for language identification

Contact

• Dataset maintainer: boffire
• Original data: Mozilla Data Collective