--- language: - zgh license: cc0-1.0 task_categories: - automatic-speech-recognition pretty_name: Common Voice 19.0 Tamazight (Combined) dataset_info: features: - name: text dtype: string - name: subset dtype: string - name: audio dtype: audio - name: duration dtype: float64 splits: - name: train num_bytes: 34567216 num_examples: 1728 download_size: 34538106 dataset_size: 34567216 configs: - config_name: default data_files: - split: train path: data/train-* --- # Common Voice 19.0 — Tamazight This dataset is a reformatted version of the Tamazight (zgh) portion of Mozilla Common Voice Corpus 19.0. ## Original source https://huggingface.co/datasets/fsicoli/common_voice_19_0 Original project: https://commonvoice.mozilla.org/ ## Processing The original train, dev, test, and other splits were concatenated and shuffled using seed 42. The invalidated split was excluded. The original MP3 recordings are preserved without resampling or silence trimming. The `duration` column contains the original MP3 duration in seconds. ## Dataset columns - text: transcription in Tifinagh - subset: original Common Voice split (train/dev/test/other) - audio: embedded MP3 recording - duration: audio duration in seconds ## Important notes This is a combined corpus, not an official Common Voice train/test partition. The original dev and test recordings are included. Do not use those recordings as independent evaluation data after training on the combined dataset. The `other` subset may contain unvalidated recordings. Transcriptions should be reviewed before use in high-quality supervised training. ## License CC0 1.0 — as specified by the source dataset. This repository is an unofficial reformatted copy of Mozilla Common Voice 19.0.