reverb-speech-mini

A deliberately tiny, instant-loading reverberant speech-like test set — the quick sanity check for dereverberation and robust-ASR front-ends. Every item is a (dry, wet) pair: a clean synthetic speech-like source and its reverberant version through a real-synthetic room impulse response (RIR), with the room's ground-truth acoustics attached.

Small is the feature: it loads in seconds, so a download actually completes and a sanity run is cheap.

Dataset at a glance

Stratification (balanced across T60 × DRR bins):

• t60_bin: T60_low (<0.3 s) 155 · T60_mid (0.3–0.8 s) 143 · T60_high (0.8–1.5 s) 93 · T60_vhigh (≥1.5 s) 109
• drr_bin: DRR_vlow (<−8 dB) 163 · DRR_low (−8…−2) 157 · DRR_mid (−2…6) 91 · DRR_high (≥6) 89

Measured ground-truth ranges:

• T60: 0.092 – 7.363 s
• DRR: −16.73 – 23.75 dB
• C50: −4.21 – 55.67 dB
• RIR categories drawn: long_t60 160 · non_shoebox 107 · coupled_rooms 97 · near_field 86 · low_drr 50

How it was built

Dry sources are synthetic speech-LIKE probe signals, not real speech: seeded harmonic (voiced) segments with time-varying F0 (80–250 Hz) and a 3–4 resonance formant filter, 3–7 Hz syllabic amplitude modulation, interleaved with fricative-like noise bursts and short silences — a speech-like ~4–8 Hz envelope that is fully reproducible and license-clean. The wet signal is fftconvolve(dry, rir) trimmed to the dry length. RIRs are reused from RIR-Bench-Hard (2,600-RIR adversarial bank), selected to balance the T60 × DRR grid.

Loading

from datasets import load_dataset

ds = load_dataset("mandipgoswami/reverb-speech-mini", split="test")
row = ds[0]
print(row["id"], row["t60_bin"], row["drr_bin"])
print(row["dry_path"], "->", row["wet_path"])   # the pairing

Decode a dry/wet pair:

import soundfile as sf
from huggingface_hub import snapshot_download
root = snapshot_download("mandipgoswami/reverb-speech-mini", repo_type="dataset")
dry, fs = sf.read(f"{root}/{row['dry_path']}")   # fs == 16000
wet, fs = sf.read(f"{root}/{row['wet_path']}")
# dereverb model: estimate dry from wet; score e.g. SI-SDR against dry.

Data fields (data.parquet)

Intended uses

• A fast dereverberation sanity set: feed wet, recover dry, score with SI-SDR / PESQ-style metrics against the paired dry.
• A quick robustness probe for ASR/enhancement front-ends across reverberation and DRR regimes, with the room's ground truth attached.

Limitations & biases

• Sources are synthetic, not real speech — speech-like structure, no phonetic/linguistic content, no speaker identity.
• Clips are short (2–3 s) and the set is small by design — a sanity check, not a training or formal-benchmark corpus.
• RIRs are synthetic (simulated), not measured in physical rooms.
• No additive noise in this set (clean dereverberation only); see audio-eval-suite for the noisy conditions.

Related

• audio-eval-suite — the companion robustness eval (adds noisy conditions and classification/RT60 tasks).
• RIR-Bench-Hard — the 2,600-RIR adversarial bank these RIRs come from.
• Part of the rirmega line of reproducible room-acoustics resources.

Citation

@misc{goswami2026reverbspeechmini,
  title  = {Reverb Speech Mini: A Tiny Reproducible Dereverberation Sanity Set},
  author = {Goswami, Mandip},
  year   = {2026},
  howpublished = {Hugging Face Datasets},
  note   = {\url{https://huggingface.co/datasets/mandipgoswami/reverb-speech-mini}}
}

License

Released under CC-BY-4.0. Sources and RIRs are synthetic, so the set carries no third-party speech-corpus restrictions.