Datasets:
Formats:
parquet
Languages:
Kabyle
Size:
100K - 1M
Tags:
automatic-speech-recognition
kabyle
common-voice
mozilla-data-collective
taqbaylit
low-resource
License:
Dataset Viewer
The dataset viewer is not available for this dataset.
Unexpected token '<', "<html>
<h"... is not valid JSON
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Common Voice Scripted Speech 26.0 - Kabyle (Huge, Cleaned)
Full cleaned dataset of Mozilla Common Voice 26.0 for Kabyle (Taqbaylit) ASR. No speaker cap, no splits — all validated, cleaned, GlotLID-filtered clips.
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
- Quality filter: ≥2 upvotes, 0 downvotes
- Character cleaning: Fixed 29 false friend types (Greek/Cyrillic epsilon → ɛ/Ɛ, gamma → ɣ/Ɣ, etc.)
- Language filtering: GlotLID v3 (kab_Latn ≥ 0.95)
Statistics
| Metric | Value |
|---|---|
| Total clips | 567,727 |
| Hours | ~520.4 |
| Speakers | 1,330 |
| Avg clip duration | ~3.3s |
Usage
from datasets import load_dataset
ds = load_dataset("boffire/common-voice-scripted-speech-kab-26-huge", split="train")
Audio Format
- Sampling rate: 16,000 Hz
- Format: MP3 (original), decoded to float32 arrays
Citation
Mozilla Common Voice: https://commonvoice.mozilla.org/
- Downloads last month
- 15