Common Voice Dataset

Common Voice Dataset Overview

Dataset Description

• Source: Speech contributions from the Common Voice community via the web platform.
• Release Frequency: Approximately every six months.
• Data Processing: All speech contributions, irrespective of validation status, are included. The dataset is packaged with the Common Voice Bundler tool and uploaded to S3.

Dataset Structure

• File Format: Each downloaded .tar.gz contains the following layout:

  [lang].tar.gz/ ├── clips/ │ ├── *.mp3 files |__ dev.tsv |__ invalidated.tsv |__ other.tsv |__ test.tsv |__ train.tsv |__ validated.tsv |__ reported.tsv (as of Corpus 5.0)

• TSV File Contents: Each .tsv lists audio files, original source sentences, hash client_id, validation data, and demographic information.

Dataset Fields

• Each row (audio clip) includes:
  • client_id
  • path
  • text
  • up_votes
  • down_votes
  • age
  • gender
  • accent
  • segment

Dataset Usage

• ML Applications: Processed with the Mozilla Corpora Creator to generate test, train, and development splits.
• Data Splits: Generation is nondeterministic to avoid duplication and demographic bias.

Access

• Download Recommendation: For large files, use curl with resume support.

Citation

• Academic Citation: When used in research, cite Common Voice: A Massively‑Multilingual Speech Corpus.