RIR-Mega: reproduce, evaluate, challenge, extend

What RIR-Mega is

RIR-Mega is an open corpus of room impulse responses with acoustic metadata (reverberation time, clarity, direct-to-reverberant ratio, and more), plus baselines and tooling for estimating these quantities. It is designed for research on reverberation, dereverberation and reverberation-robust speech.

Scientific questions it addresses

• Can acoustic descriptors be estimated reliably from RIRs?
• How do simulated RIRs compare to measured ones?
• How does reverberation degrade downstream ASR?

Known limitations

• The corpus is largely synthetic; synthetic-to-real gaps exist.
• Some published metadata fields are not populated for every sample (see the Explorer for exactly which fields carry data).
• Licensed CC-BY-NC-4.0 (non-commercial).

Track A — Acoustic validity

Recompute RT60 (T20/T30), EDT, DRR, C50, C80 and D50 directly from the RIR waveforms with a documented Schroeder-backward-integration estimator, and measure agreement with the dataset's published metadata (bias, MAE, RMSE, bootstrap 95% CI).

Track B — Synthetic-to-real generalization

A reproducible protocol (not a data redistribution) to compare RIR-Mega against independently measured corpora such as ACE and REVERB under their own licenses. Enforces matched sample rate, channel convention and held-out separation, and reports domain shift and failure cases.

Track C — Downstream audio ML (ASR)

Wraps the published whisper-rirmega-bench pipeline: WER/CER on paired clean↔reverberant speech. Documents model versions, splits and compute. A lightweight dereverberation (WPE) baseline is optional.