WildASR

Official dataset for Back to Basics: Revisiting ASR in the Age of Voice Agents.

Code: github.com/boson-ai/WildASR-public

Overview

WildASR is a multilingual diagnostic benchmark built from real human speech to stress-test ASR robustness under real-world out-of-distribution (OOD) conditions. We decompose robustness into three axes:

• Environmental Degradation (the where): reverberation, far-field, phone codec, noise gap, clipping
• Demographic Shift (the who): children, older adults, accented speech
• Linguistic Diversity (the what): short utterances, incomplete audio, code-switching

Dataset

Due to licensing constraints, we currently release 7 splits covering environment degradation (clean, clipping, far-field, noise gap, phone codec, reverberation) and demographic shift (accent). 10,058 samples, ~30 hours total. Each sample contains audio (16kHz WAV), transcript, and metadata (category, subset, language, etc.). More splits and languages will be added as licenses are cleared.

Usage

from datasets import load_dataset

# Load all splits
ds = load_dataset("bosonai/WildASR")

# Load a specific split
clean = load_dataset("bosonai/WildASR", split="environment_degradation__en__fleurs_clean_en")

# Play audio (in a notebook)
clean[0]["audio"]

Run evaluation with WildASR toolkit

pip install git+https://github.com/boson-ai/WildASR-public.git

# Save a split as parquet for the eval toolkit
clean.to_parquet("data/fleurs_clean.parquet")

from run_eval.eval import create_client, run_asr_evaluation, ASREvalConfig

client = create_client("whisper-large-v3", "en")
cfg = ASREvalConfig(
    model_name="whisper-large-v3",
    data_path="data/fleurs_clean.parquet",
    output_dir="results/whisper-large-v3",
    language="en",
    wer_method="qwen",
)
run_asr_evaluation(client=client, config=cfg)

Citation

@misc{wildasr2026,
  title   = {Back to Basics: Revisiting ASR in the Age of Voice Agents},
  author  = {Geeyang Tay and Wentao Ma and Jaewon Lee and Yuzhi Tang and Daniel Lee and Weisu Yin and Dongming Shen and Yi Zhu and Mu Li and Alex Smola},
  year    = {2026},
  note    = {arXiv:TODO}
}

License

Apache 2.0