README.md

metadata

license: mit

VoxCeleb2 Language-Detected Subset

This dataset is a language-labeled version of the VoxCeleb2 speaker identification dataset. It was created using the ProgramComputer/voxceleb Hugging Face dataset and the speechbrain/lang-id-voxlingua107-ecapa language identification model.

Dataset Contents

The dataset consists of two CSV files:

• audio_clips_meta_data.csv
  Contains metadata for each audio clip, including:

  • clip_id: Unique identifier for the audio clip.
  • speaker_id: ID of the speaker (matches VoxCeleb2 format).
  • language: Language predicted using SpeechBrain's language ID model.
  • audio_file: The audio clip file name
  • 'path': The local file path of the audio clip.

• vox2_meta.csv
  Maps speaker_id values to the speaker's real-world identity:

  • speaker_id: ID of the speaker.
  • name: Full name of the speaker.

How the Dataset Was Created

1. The original VoxCeleb2 audio clips were accessed via the ProgramComputer/voxceleb. The combined zip archives were used.
2. For each audio clip, the speechbrain/lang-id-voxlingua107-ecapa model was used to infer the spoken language.
3. The resulting metadata, including language labels, was stored in a CSV file (audio_clips_meta_data.csv).