Datasets:
Two tiny, one-line-loadable audio eval sets (reverb robustness + dereverb sanity)
I just released two small, instant-loading evaluation datasets for room-acoustics / reverberant-speech models:
audio-eval-suite — 300 source items x 5 conditions (clean, reverberant, noisy @ 10/5/0 dB) = 1,500 clips, 144 MB. Ships two reference tasks with baselines: reverb-level classification (accuracy) and RT60 regression (MAE), plus an lm-eval-style task template and a quickstart notebook.
reverb-speech-mini — 500 dry/wet pairs, 81 MB, 2-3 s each. A fast sanity set for dereverberation / robust-ASR, with a reproducible reverb proxy and an identity-dereverb baseline.
Both load in one line (load_dataset(...)), are CC-BY-4.0, and use fully reproducible synthetic speech-like probe signals convolved through a physics-simulated RIR bank. Companion to RIR-Bench-Hard. Feedback and submissions welcome.