Oulasong/Real_Audio_Bench

Real Audio Bench

Real Audio Bench is a compact human-recorded benchmark for streaming spoken reasoning. The benchmark contains 186 English audio recordings. Each item is a natural spoken question in which an early answer can be misleading until a later cue is heard.

Audio files are hosted at: https://huggingface.co/datasets/Oulasong/RealAudioBench

Files

• —audio/: 186 WAV files.
• —real_audio_bench_manifest.jsonl: one metadata row per audio file, including the transcript-style question, corrected gold answer, and item metadata.
• —real_audio_bench_manifest.csv: CSV version of the same manifest.
• —real_audio_bench_gold_answers.csv: compact table with audio path, question, gold answer, and basic tags for manual inspection.

Manifest schema

Each row contains:

• —audio_path: relative path of the WAV file in the Hugging Face dataset.
• —reader: speaker/reader id.
• —item: item id within the reader packet.
• —source_id: source item identifier used during benchmark construction.
• —question: spoken prompt text.
• —gold_answer: reference answer used for evaluation.
• —category, difficulty, scenario, mechanism, duration: construction metadata.
• —critical_cue, why_early_answer_is_wrong, speaker_note: explanatory metadata from the reader packet.

Construction summary

The benchmark was constructed from scripted spoken-reasoning prompts and recorded by five internal readers. The final public split keeps 186 recordings after removing ambiguous or unsuitable items from the 200-item reader packet. The manifest contains the reference answers used for evaluation.

Intended use

This benchmark is intended as a compact transfer check for models that answer spoken questions from real human recordings. It is not intended as a large-scale user study or as a general-purpose speech recognition benchmark.

License

This dataset is released under CC BY 4.0.

Release notes

This release contains 186 manifest rows and 186 unique audio paths. Each row has a non-empty reference answer.