Voices-in-the-Wild-Bench

Voices-in-the-Wild-Bench is a bilingual benchmark for evaluating robust automatic speech recognition under real-world acoustic conditions. It contains 5,000 speech examples covering Chinese and English, synthetic and real-recorded speech, and eight acoustic scenarios.

Dataset Composition

Group	Count
Synthetic speech	3,500
Real-recorded speech	1,500
Chinese	2,500
English	2,500
Total	5,000

Splits

Splits are organized by recording source and acoustic category. Prefixes denote source type:

• real_: real-recorded speech
• syn_: synthetic speech

Available categories are noise, far_field, obstructed, distortion, recording, echo, dropout, and mixed.

Example split names:

real_noise
syn_noise
real_mixed
syn_mixed

Fields

• audio: embedded audio feature.
• file_name: release-relative audio file name.
• audio_path: release-relative audio path.
• question: transcription instruction.
• answer: reference transcription.
• text: same as answer, provided for ASR tooling compatibility.
• subset: original subset label.
• prediction: empty placeholder for model outputs.
• name: sample name.
• index: sample index.

Usage

from datasets import load_dataset

dataset = load_dataset(
    "zhifeixie/Voices-in-the-Wild-Bench",
    split="syn_noise",
)

sample = dataset[0]
audio = sample["audio"]
reference = sample["answer"]

Evaluation

We report CER for Chinese samples and WER for English samples. See the project repository and leaderboard for evaluation scripts, model wrappers, and submission instructions:

• Repository: https://github.com/xzf-thu/Voices-in-the-Wild-Bench
• Leaderboard: https://xzf-thu.github.io/Voices-in-the-Wild-Bench/

Citation

If you use this benchmark, please cite the associated paper. Citation metadata will be updated after release.