README.md

---
license: cc-by-sa-4.0
task_categories:
- audio-classification
size_categories:
- 1M<n<10M
configs:
- config_name: default
  data_files: "*.tar"
  default: true
---

# VoxCeleb2 - dev set

This is a copy of VoxCeleb2 dev set in WebDataset format. The audio data is the original AAC-encoded files without any transcoding. Refer to https://arxiv.org/abs/1806.05622 for more details about the dataset.

There are 1,092,009 samples covering 5,994 unique speakers. The dataset is split into 779 shards of ~100MB.

## Usage

```python
import torchaudio
import webdataset as wds
from datasets import load_dataset

def decode_audio(sample):
    audio, fs = torchaudio.load(sample.pop("m4a"))  # requires FFMPEG to decode AAC. refer to torchaudio doc
    # optionally resample audio and other pre-processing
    sample["audio"] = audio
    return sample

# using webdataset library
ds = wds.WebDataset("https://huggingface.co/datasets/gaunernst/voxceleb2-dev-wds/resolve/main/voxceleb2-dev-{0000..0778}.tar")
ds = ds.map(decode_audio)
next(iter(ds))

# using HF datasets library
ds = load_dataset("gaunernst/voxceleb2-dev-wds", split="train", streaming=True)
ds = ds.map(decode_audio)
next(iter(ds))
```

The original filename is kept. In other words, if you download all shards and untar them, it will be exactly the same as the original folder (with extra `.cls` files containing pre-defined speaker_id-to-integer mapping). You can also retrieve the original speaker ID and YouTube video ID from the `__key__` field.

## Citation

```
@InProceedings{Chung18b,
  author       = "Chung, J.~S. and Nagrani, A. and Zisserman, A.",
  title        = "VoxCeleb2: Deep Speaker Recognition",
  booktitle    = "INTERSPEECH",
  year         = "2018",
}
```