---
license: apache-2.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: dev
    path: data/dev-*
  - split: test
    path: data/test-*
---

# 📀 Common Voice 22.0 — Parquet Repack (Community Version)

> ⚠️ **Unofficial community repackage. Not affiliated with Mozilla.**  
This dataset is a community-generated Parquet version of the original **Mozilla Common Voice 22.0** release.  
It was converted to a more training-friendly format to speed up loading, improve compatibility with modern ML frameworks, and support distributed training.

## 🚀 TL;DR

- **Original Source:** Mozilla Common Voice 22.0 (public domain voice dataset)
- **This Version:** Restructured into **Parquet** for faster I/O + easier ML training
- **Why:** To make fine-tuning speech models (Whisper, MMS, Wav2Vec2, etc.) less painful
- **License:** Apache-2.0 for *this repack only* — original audio remains under **CC-0** by Mozilla contributors

---

## 📚 Dataset Details

### 📝 What This Is

This is **not the official dataset** — it’s a **re-packaged mirror** for convenience & performance.

The raw `.tar.gz` archives from Mozilla were:
1. Downloaded from the official Common Voice hosting
2. Extracted + validated
3. Converted to `.parquet` with structured metadata + 16kHz audio
4. Splits preserved: `train`, `dev`, `test`

### 🌍 Languages

Currently includes: **French (fr)** from CV22  
(*Extendable to other languages if the community wants to contribute*)

### 🤝 Credits

| Role | Entity |
|------|--------|
| Original data creators | Mozilla + Common Voice community |
| Re-packaged by | Community for educational & research use |
| Affiliation | ❌ No affiliation with Mozilla |

---

## 📂 Dataset Structure

| Field | Type | Description |
|-------|--------|----------------|
| `client_id` | string | Anonymous speaker ID |
| `path` | string | Audio file path |
| `sentence_id` | string | Sample unique ID |
| `sentence` | string | Ground-truth transcription |
| `sentence_domain` | string | Domain / category of sentence |
| `up_votes` | string | Community upvotes |
| `down_votes` | string | Community downvotes |
| `age` | string | Optional user self-reported |
| `gender` | string | Optional user self-reported |
| `accents` | string | Accent info if provided |
| `variant` | null | Unused |
| `locale` | string | Locale code |
| `audio` | audio (16kHz) | Audio object |
| `segment` | null | Unused |

---

## 🧠 Intended Uses

### ✅ Good For

- Fine-tuning speech-to-text models (Whisper, Wav2Vec2, MMS, etc.)
- ASR evaluation and benchmarking
- Research on accents, pronunciation, and speech diversity
- Training small and large-scale ASR models efficiently (Parquet = faster)

### ❌ Not Recommended For

- Commercial use *without reviewing original Common Voice licensing*
- Speaker identification or deanonymization research
- Training models intended to profile demographic or identity attributes

---

## 🧪 Dataset Creation

### Why This Exists

Loading CV22 from raw MP3/TSV is **slow as hell** for modern GPU pipelines.  
This repack aims to:
- Reduce dataset loading overhead
- Improve compatibility with HF `datasets` / PyTorch / JAX / TPU
- Make community fine-tuning more accessible

### Source Data Collection

All audio was **originally donated by volunteers** to Mozilla under **CC-0**.  
This version applies **no extra filtering** beyond the original release.

### ⚖️ Personal & Sensitive Info

- Contains **voice data**, which is inherently biometric
- Age, gender, and accent are **self-reported** and optional
- All speaker identifiers are **anonymized IDs** from Mozilla

Users should avoid re-identification or any non-ethical use of voice data.

---

## ⚠️ Bias, Risks & Limitations

- Age/gender/accent labels may be inaccurate or incomplete
- Speech data may not represent all demographics equally
- Model trained on this may reflect bias from accents or speaker distribution
- Not ideal for extremely low-resource or domain-specific speech tasks

### Recommendations

- Combine with other datasets for balanced performance
- Evaluate ASR models across demographics + accents to detect bias

---

## 📎 Citation

If you use this dataset, cite **both**:

### Mozilla Common Voice

```bibtex
@misc{mozilla_common_voice_2023,
  title = {Mozilla Common Voice Dataset},
  howpublished = {https://commonvoice.mozilla.org/},
  year = {2023}
}
@misc{common_voice_22_parquet_community,
  title = {Common Voice 22.0 — Community Parquet Repack},
  year = {2025},
  note = {Unofficial preprocessing for research/education}
}