RIR-Mega-Speech: A Reverberant Speech Corpus with Comprehensive Acoustic Metadata and Reproducible Evaluation

TL;DR

RIR-Mega-Speech addresses reproducibility gaps in reverberant ASR research by providing a large, fully annotated corpus created by convolving LibriSpeech with ~5,000 simulated RIRs from RIR-Mega. Each reverberant file includes ground-truth acoustics ($RT_{60}$, $DRR$, $C_{50}$), and the authors supply complete code to rebuild audio, compute metrics, and reproduce results, enabling independent verification. Using Whisper small, they quantify a 48% relative degradation in WER from clean to reverberant speech and show clear trends: $WER$ rises with $RT_{60}$ and falls with $DRR$. The work emphasizes transparency and reproducibility, offering a framework and data to evaluate dereverberation and robust ASR methods, while acknowledging limitations of simulated acoustics and uneven coverage that motivate future extensions.

Abstract

Despite decades of research on reverberant speech, comparing methods remains difficult because most corpora lack per-file acoustic annotations or provide limited documentation for reproduction. We present RIR-Mega-Speech, a corpus of approximately 117.5 hours created by convolving LibriSpeech utterances with roughly 5,000 simulated room impulse responses from the RIR-Mega collection. Every file includes RT60, direct-to-reverberant ratio (DRR), and clarity index ($C_{50}$) computed from the source RIR using clearly defined, reproducible procedures. We also provide scripts to rebuild the dataset and reproduce all evaluation results. Using Whisper small on 1,500 paired utterances, we measure 5.20% WER (95% CI: 4.69--5.78) on clean speech and 7.70% (7.04--8.35) on reverberant versions, corresponding to a paired increase of 2.50 percentage points (2.06--2.98). This represents a 48% relative degradation. WER increases monotonically with RT60 and decreases with DRR, consistent with prior perceptual studies. While the core finding that reverberation harms recognition is well established, we aim to provide the community with a standardized resource where acoustic conditions are transparent and results can be verified independently. The repository includes one-command rebuild instructions for both Windows and Linux environments.

Figures (10)

• Figure 1: RT60 distribution across all reverberant files. Most files fall between 0.2 and 0.8 seconds.
• Figure 2: DRR distribution using a 2.5 ms direct-only window. The long tail toward negative values reflects weak direct arrivals in some simulated RIRs.
• Figure 3: Coverage heatmap (RT60 vs DRR). Darker cells indicate more files. Coverage is uneven due to non-stratified sampling.
• Figure 4: Duration vs RT60. Long and short utterances are distributed across RT60 bins, reducing confounding in WER analysis.
• Figure 5: Spectrogram comparison for three utterances. Left column: clean. Right column: reverberant. Smearing of formant structure is visible.