Datasets:
Download README.md from mandipgoswami/rir-mega-speech: direct link, hf CLI and curl.
- Browser
- Download file 7.58 kB
-
https://huggingface.co/datasets/mandipgoswami/rir-mega-speech/resolve/main/README.md
- Command line
-
hf download hf://datasets/mandipgoswami/rir-mega-speech/README.md
-
curl -L -o README.md https://huggingface.co/datasets/mandipgoswami/rir-mega-speech/resolve/main/README.md
language:
- en
license: cc-by-nc-4.0
tags:
- speech
- reverberation
- room-impulse-response
- acoustics
- asr
- whisper
- reproducibility
- synthetic
task_categories:
- automatic-speech-recognition
- audio-to-audio
- audio-classification
pretty_name: RIR-Mega-Speech
size_categories:
- 100K<n<1M
RIR-Mega-Speech
Dataset Summary
RIR-Mega-Speech is a large-scale reverberant speech corpus created by convolving LibriSpeech utterances with simulated room impulse responses (RIRs sampled from the RIR-Mega collection). Each reverberant utterance includes per-file acoustic metadata computed from the source RIR, enabling controlled analysis of reverberation effects on speech processing systems.
This dataset emphasizes transparency and reproducibility: acoustic metrics are explicitly defined, and the underlying RIR corpus, generation scripts, and evaluation methodology are publicly available.
Relationship to RIR-Mega
This dataset is derived from the RIR-Mega room impulse response corpus:
- Hugging Face: https://huggingface.co/datasets/mandipgoswami/rirmega
- Zenodo DOI: https://doi.org/10.5281/zenodo.17387402
- rirmega arXiv paper: https://arxiv.org/abs/2510.18917
- rirmega-speech arXiv paper: https://arxiv.org/abs/2601.19949v1
If you require access to the raw RIR waveforms or wish to generate alternative speech corpora, please refer to the RIR-Mega dataset directly.
Motivation
Despite extensive work on reverberant speech, comparing methods remains difficult because many corpora lack per-file acoustic annotations or provide limited documentation for reproduction. RIR-Mega-Speech is intended as a standardized resource where acoustic conditions are explicit and results can be independently verified.
Source Data
- Clean speech: LibriSpeech (public-domain audiobook speech)
- Room impulse responses: RIR-Mega (simulated, physics-based)
Data Generation
For each clean utterance (x[n]), a room impulse response (h[n]) sampled from RIR-Mega is applied via time-domain convolution: [ y[n] = (x * h)[n]. ] RIRs that produce clipping or fail metadata validation are excluded.
Acoustic Metadata
Metadata is computed from the source RIR prior to convolution:
- RT60: Schroeder backward integration with linear fit between −5 dB and −35 dB, extrapolated to −60 dB (ISO 3382-1 compliant)
- DRR: Direct-to-reverberant ratio using a 2.5 ms direct-only window centered on the first arrival
- C50: Clarity index with a 50 ms boundary
- LUFS: Integrated loudness (ITU-R BS.1770)
- duration_s: Utterance duration in seconds
Note: The DRR definition uses a narrow direct-only window and may produce very low values for weak direct arrivals.
Dataset Structure
rir-mega-speech/ ├── audio/ │ ├── train/ │ ├── dev/ │ └── test/ ├── metadata/ │ ├── metadata.csv │ ├── train.csv │ ├── dev.csv │ └── test.csv └── examples/
- Total reverberant files: 53,200
- Total duration: ~117.5 hours
- Underlying RIR pool: ~10,000 simulated RIRs(via RIR-Mega)
Splits
Train, development, and test splits are speaker-stratified to prevent speaker overlap across partitions. Acoustic parameters are not explicitly stratified, resulting in uneven coverage across the RT60–DRR plane.
Metadata Schema
| Column | Description |
|---|---|
audio |
Relative path to reverberant WAV file |
rt60 |
Reverberation time (seconds) |
drr |
Direct-to-reverberant ratio (dB) |
c50 |
Clarity index C50 (dB) |
lufs |
Integrated loudness (LUFS) |
duration_s |
Utterance duration (seconds) |
clean_id |
Source LibriSpeech utterance ID |
rir_id |
Identifier of the RIR-Mega impulse response |
split |
Dataset split (train / dev / test) |
Recommended Use
- Robust ASR benchmarking under controlled reverberation
- Dereverberation and speech enhancement evaluation
- Analysis of ASR error trends versus acoustic parameters
Limitations
- RIRs are simulated and may not capture all real-room effects (e.g., furniture, HVAC noise).
- Acoustic coverage across RT60 and DRR is uneven due to non-stratified sampling.
- The DRR definition differs from perceptual conventions that include early reflections.
Reproducibility
The underlying RIR corpus, dataset generation scripts, and evaluation pipeline are publicly available. Users can regenerate the dataset or construct alternative speech corpora using the same RIR pool.
License
- LibriSpeech: CC BY 4.0
- RIR-Mega RIRs: MIT
- Derived reverberant speech: CC BY 4.0
Related benchmark: RIR-Bench-Hard
For a dedicated adversarial evaluation benchmark built in the same research program, see RIR-Bench-Hard — 2,600 synthetic RIRs across five deliberately hard regimes (long-T60 halls, coupled rooms, near-field sources, low-DRR, non-shoebox geometries) with standardized est_hard leaderboard splits, a reproducible evaluation harness, and per-RIR acoustic metadata (T60, DRR, C50, EDT). RIR-Bench-Hard is the stress-test complement to this dataset: use it to measure how acoustic-parameter-estimation and dereverberation methods hold up under extreme conditions.
Related eval sets
Part of the same room-acoustics research program:
- mandipgoswami/audio-eval-suite — small audio-robustness eval set (clean / reverberant / noisy conditions) with reference RT60-regression and reverb-classification tasks.
- mandipgoswami/reverb-speech-mini — tiny dry/wet reverberant-speech pairs for fast dereverberation / robust-ASR sanity checks.
- mandipgoswami/RIR-Bench-Hard — adversarial room-impulse-response benchmark across five hard regimes.
Independent benchmark & evaluation portal
An independent, reproducible benchmark portal provides interactive data exploration, a five-minute reproducibility quickstart, and an open, pull-request-based result-submission workflow for this dataset.
- Portal: https://rirmega-benchmark.vercel.app
- Benchmark suite & submissions: https://github.com/mandip42/rirmega-benchmark
The suite runs three tracks — acoustic validity (recompute RT60/DRR/clarity from the waveforms and compare to the published metadata), synthetic-to-real generalization (a protocol against measured corpora under their own licenses), and downstream ASR robustness. Results are grouped by provenance (self-reported / artifact-checked / independently-reproduced); passing automated validation does not imply endorsement.
Citation
If you use this dataset, please cite both the speech corpus and the underlying RIR collection:
RIR-Mega-Speech
Goswami, M. (2026). RIR-Mega-Speech: A Reverberant Speech Corpus with Comprehensive Acoustic Metadata and Reproducible Evaluation. arXiv:2601.19949. https://arxiv.org/abs/2601.19949v1
RIR-Mega
Goswami, M. (2025). RIR-Mega: A Large-Scale Room Impulse Response Corpus with Benchmarks for Industrial and Building Acoustics. arXiv:2510.18917. https://arxiv.org/abs/2510.18917