README.md

---
language:
- en
task_categories:
- audio-to-audio
- automatic-speech-recognition
- audio-classification
license: cc-by-nc-4.0
tags:
- audio
- rir
- acoustics
- dereverberation
- robust-asr
- simulation
- room-acoustics
---
## 🔖 How to Cite
If RIR-Mega helps your research, please cite both the paper and the dataset:

Paper
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

Dataset
Goswami, M. (2025). RIR-Mega Dataset (v1.0.0). Zenodo. https://doi.org/10.5281/zenodo.17387402
```
@misc{goswami2025rirmega,
  title        = {RIR-Mega: A Large-Scale Room Impulse Response Corpus with Benchmarks for Industrial and Building Acoustics},
  author       = {Goswami, Mandip},
  year         = {2025},
  eprint       = {2510.18917},
  archivePrefix= {arXiv},
  primaryClass = {cs.SD},
  url          = {https://arxiv.org/abs/2510.18917}
}

@dataset{goswami_2025_rirmega_zenodo,
  author       = {Goswami, Mandip},
  title        = {RIR-Mega Dataset},
  year         = {2025},
  publisher    = {Zenodo},
  version      = {v1.0.0},
  doi          = {10.5281/zenodo.17387402},
  url          = {https://doi.org/10.5281/zenodo.17387402}
}
```
```bash
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

Goswami, M. (2025). RIR-Mega Dataset (v1.0.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17387402
```
# RIR-Mega

[Paper](https://huggingface.co/papers/2510.18917) | [Code](https://github.com/mandip42/rirmega) | [Project page (Zenodo)](https://doi.org/10.5281/zenodo.17387402)

RIR-Mega provides thousands of simulated room impulse responses for research in dereverberation, robust speech recognition, and acoustic scene analysis.
This Hugging Face release hosts a lightweight, representative subset — 1 000 linear-array and 3 000 circular-array RIRs — for quick exploration, tutorials, and reproducible baselines.

The complete 50 000-RIR archive is permanently preserved on Zenodo and described in the accompanying paper:

🤗 Subset for streaming: (https://huggingface.co/datasets/mandipgoswami/rirmega)
📦 Technical Paper: ([arxiv.org/abs/2510.18917](https://doi.org/10.48550/arXiv.2510.18917))
📦 RIR Mega-Speech Paper: ([https://arxiv.org/abs/2601.19949v1])

## ✨ What’s inside
- `data/` — RIR audio and `metadata/metadata.csv` (compact schema)
- `rirmega/dataset.py` — Hugging Face Datasets loader
- `benchmarks/rt60_regression/` — a lightweight RT60 regression baseline
- `scripts/` — utilities (validation, checksums, mini subset)
- *(optional)* `data-mini/` — tiny subset for quick demos and Spaces

## Contents
| Folder | Description |
| ----------------------------------------- | ------------------------------------------------------------------------------ |
| `data/audio/linear` | 1 000 RIRs simulated for linear microphone arrays |
| `data/audio/circular` | 3 000 RIRs simulated for circular arrays |
| `data/metadata/metadata.csv` / `.parquet` | Compact schema linking each file to acoustic metrics and simulation parameters |
| `rirmega/dataset.py` | Hugging Face Datasets loader (supports streaming) |
| `benchmarks/rt60_regression/` | Baseline RT60 regression example |
| `scripts/` | Validation + checksum utilities |
| `figs/` | Overview and validation plots for reference |


## 📦 Schema (compact)
| Column | Meaning |
| -------------------------------------- | ----------------------------------------------------------------------------- |
| `id` | unique identifier |
| `family` | “linear” or “circular” |
| `split` | train / valid / test |
| `fs` | sampling rate (Hz) |
| `wav` | relative path to audio file |
| `room_size`, `absorption`, `max_order` | simulation parameters |
| `metrics` | JSON string with `rt60`, `drr_db`, `c50_db`, `c80_db`, and band-limited RT60s |
| `rng_seed` | random seed for reproducibility |

## 🚀 Getting started

```bash
from datasets import load_dataset
ds = load_dataset("mandipgoswami/rirmega", trust_remote_code=True)
print(ds["train"][0]["audio"])       # lazy-loads waveform
print(ds["train"][0]["rt60"])        # scalar metadata
```

## For streaming or partial download:
```bash
ds = load_dataset("mandipgoswami/rirmega", streaming=True)
```
## 🧪 Baseline: RT60 regression
Lightweight features + RandomForest to predict RT60-like targets from RIR signals.

```bash
python benchmarks/rt60_regression/train_rt60.py
```
```bash
pip install soundfile numpy pandas scikit-learn
python benchmarks/rt60_regression/train_rt60.py
# or choose a specific target key present in `metrics`
python benchmarks/rt60_regression/train_rt60.py --target rt60
```
**Default target search order:**  

## Technical Validation 
A random subset of 1 000 samples was analyzed for internal consistency.
The RT60 values derived from Schroeder energy decay curves correlated strongly with the metadata values:
| Metric | Correlation | MAE (s) | RMSE (s) |
| ---------------------- | ----------- | ------- | -------- |
| RT60 (metadata vs EDC) | 0.96 | 0.013 | 0.022 |


### Reference numbers (example)
- Train/Valid used: 36,000 / 4,000 (auto 10% valid)
- Metric: MAE = **0.013 s**, RMSE = **0.022 s** (auto target)

## 🏅 Leaderboard (RT60 regression)
| Date | Team / Author | Method | Target | Train/Valid | MAE (s) | RMSE (s) | Seed | Code |
|---|---|---|---|---|---:|---:|---:|---|
| 2025-10-19 | Baseline (RIR-Mega) | RF on light feats | auto | 36k / 4k | 0.013 | 0.022 | 0 | `benchmarks/rt60_regression` |

> 📫 **Submit a result:** Open a PR adding a row (see **Submitting**).

## 📤 Submitting
See **SUBMITTING.md** for rules and a PR template. Minimum info:
- Command (incl. `--target` if used), seed, dataset tag (e.g., `v1.0.0`)
- Train/Valid sizes used
- MAE (s) and RMSE (s)
- Link to code (repo, gist, or HF Space)