zhifeixie/Voices-in-the-Wild-test-v2

Voices in the Wild

Voices in the Wild is an automatic speech recognition dataset for robustness evaluation and training under diverse acoustic conditions. Each example contains an augmented speech waveform, a transcription, and a normalized acoustic subset label.

Repository Structure

Voices-in-the-Wild/
├── README.md
├── metadata.jsonl
├── meta.jsonl
└── audio/
    ├── noise/
    ├── far_field/
    ├── far_field_noise/
    └── ...

For Hugging Face loading, metadata.jsonl contains a repository-relative file_name field. The audio files use sequential public names such as sample_000000.wav.

Data Fields

• file_name: relative path to the audio file.
• audio_path: audio path retained for compatibility with local ASR tooling.
• audio: decoded by Hugging Face when loading from AudioFolder or Parquet.
• text: transcription alias copied from answer.
• answer: reference transcription.
• question: transcription instruction.
• subset: normalized acoustic condition category.
• prediction: empty placeholder for model output.
• name: public sample identifier.
• index: integer sample index.

Dataset Size

• Total examples: 590499
• Subset categories: 37
• Audio format: WAV

Subset Distribution

subset	count
noise	89418
far_field_noise	76044
barrier	42826
echo	42620
far_field	42610
distortion	42575
dropout	42478
recording	42353
far_field_dropout	30300
noise_dropout	29886
far_field_recording	29854
barrier_dropout	3197
echo_dropout	3150
echo_noise	3130
far_field_recording_distortion	3113
echo_distortion_dropout_noise	3102
echo_recording_dropout	3100
echo_recording_distortion_noise	3098
barrier_distortion	3095
echo_distortion_dropout	3090
echo_recording_dropout_noise	3088
echo_recording_distortion_dropout	3079
echo_recording_distortion_dropout_noise	3073
recording_distortion_dropout_noise	3072
far_field_distortion_dropout_noise	3063
echo_distortion	3061
echo_recording_distortion	3056
echo_recording	3040
barrier_recording	3032
far_field_recording_dropout	3032
far_field_recording_distortion_dropout	3004
far_field_distortion	3003
barrier_noise	2995
far_field_recording_distortion_noise	2990
far_field_distortion_dropout	2963
far_field_recording_dropout_noise	2957
far_field_recording_distortion_dropout_noise	2952

Loading

from datasets import load_dataset, Audio

ds = load_dataset("Prummn/Voices-in-the-Wild")
ds = ds.cast_column("audio", Audio())
print(ds["train"][0])

For local AudioFolder loading:

from datasets import load_dataset, Audio

ds = load_dataset("audiofolder", data_dir="/data/haobin/Voices-in-the-Wild")
ds = ds.cast_column("audio", Audio())

Parquet Format

The recommended upload format stores audio bytes directly inside Parquet shards. The audio column is a Hugging Face-compatible struct with bytes and path fields. Metadata fields such as text, answer, subset, name, and index are stored in the same rows.