DFHOAFH/whisper-rirmega-bench

Whisper-RIR-Mega 是一个用于评估自动语音识别（ASR）系统对房间声学鲁棒性的基准数据集，包含成对的干净和混响语音样本。每个样本包括：干净音频（来自LibriSpeech test-clean，16 kHz）、混响音频（同一语音与RIR-Mega中的脉冲响应卷积）、参考文本、RIR元数据（如rir_id、RT60、DRR、C50等）以及技术论文链接。数据集的分割基于RT60或DRR进行分层，以确保声学条件的平衡。该数据集可用于：比较Whisper（或任何ASR）在干净和混响语音上的表现并报告混响惩罚（Δ WER）、评估跨RT60/DRR区间的鲁棒性、复现官方Whisper-RIR-Mega排行榜。

Whisper-RIR-Mega is a benchmark dataset of paired clean and reverberant speech for evaluating ASR robustness to room acoustics. Each sample consists of: clean speech (LibriSpeech test-clean, 16 kHz), reverberant speech (same utterance convolved with one RIR from RIR-Mega), reference transcript, RIR metadata (rir_id, RT60, DRR, C50, etc.), and a link to the technical paper. Splits are stratified by RT60 or DRR to ensure balanced benchmarking across acoustic conditions. Use this dataset to: benchmark Whisper (or any ASR) on clean vs. reverberant speech and report reverb penalty (Δ WER), evaluate robustness across RT60/DRR bins, and reproduce the official Whisper-RIR-Mega leaderboard.