Hugging Face's logo

Datasets:
mandipgoswami
/
rirmega

Tasks:
Audio-to-Audio
Audio Classification
Other
Modalities:
Audio
Formats:
soundfolder
Languages:
English
Size:
1K - 10K
ArXiv:
Tags:
audio
rir
room-impulse-response
room-acoustics
reverberation
dereverberation
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
Community
2
rirmega
File size: 4,650 Bytes
5438937
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
 
---
pretty_name: RIRmega v2
license: cc-by-nc-4.0

language:
  - en

task_categories:
  - audio-to-audio
  - audio-classification
  - other

tags:
  - audio
  - rir
  - room-impulse-response
  - room-acoustics
  - reverberation
  - dereverberation
  - acoustic-parameter-estimation
  - robust-asr
  - speech-enhancement
  - acoustic-scene-analysis
  - simulation
  - benchmark
  - regression
  - acoustics
  - speech
  - python
  - pytorch
---


# RIRmega v2 — Dataset card (Hugging Face)

This folder holds the v2 artifacts intended for the Hugging Face dataset [mandipgoswami/rirmega](https://huggingface.co/datasets/mandipgoswami/rirmega). When published, use revision **v2.0.0** for the v2 release.

## Dataset description

RIRmega v2 extends the existing RIRmega v1 dataset with:

- A **versioned metadata schema** (`metadata_v2.parquet`) with acoustic metrics (RT60, DRR, C50, C80, D50, EDT), quality-control grades, and provenance.
- A **QC report** (`qc_report.parquet`) with checks and outlier scores.
- **Official splits** in `splits/`: `random`, `unseen_room`, `unseen_distance` (train/validation/test).

All v2 fields are derived from v1 or computed from the RIR waveforms; no private data is assumed.

## v1 vs v2 changelog

| Aspect        | v1 | v2 |
|-------------|----|----|
| Metadata     | Compact CSV/JSON (id, family, split, fs, metrics) | Parquet with full schema (see below) |
| QC           | None | Grades A/B/C and qc_report.parquet |
| Splits       | train/valid/test only | + unseen_room, unseen_distance, random |
| Metrics      | In JSON when present | Computed from RIR when missing (Schroeder, DRR, C50/C80/D50) |

## Schema (metadata_v2.parquet)

| Column              | Type   | Description |
|--------------------|--------|-------------|
| sample_id           | string | Unique identifier |
| sample_rate         | int    | Sampling rate (Hz) |
| rir_length_samples  | int    | Length of RIR |
| onset_sample        | int    | Estimated onset |
| room_id             | string | Room or pseudo-room bucket |
| src_xyz, mic_xyz    | string/JSON | Source/receiver coordinates if available |
| distance_m          | float  | Source–receiver distance if available |
| RT60_T20_s, RT60_T30_s | float | Reverberation time |
| EDT_s               | float | Early decay time |
| DRR_dB              | float | Direct-to-reverberant ratio |
| C50_dB, C80_dB      | float | Clarity indices |
| D50                 | float | Early energy fraction (0–1) |
| band_rt60_octave    | string | JSON band-limited RT60 when available |
| qc_grade           | string | A / B / C |
| qc_flags            | string | JSON list of flags |
| generator_version   | string | e.g. 2.0.0 |
| provenance          | string | e.g. v1_rirmega_hf |

## Splits

- **random**: Random train/validation/test for baseline sanity.
- **unseen_room**: Test set = rooms held out by room_id.
- **unseen_distance**: Test set = distance buckets held out.

Each split is a JSON file with keys `train`, `validation`, `test` (lists of `sample_id`).

## Evaluation tasks

- **Parameter estimation**: Predict RT60 (T20), DRR, C50 from RIR (regression). Metrics: MAE, RMSE, R²; bootstrap 95% CIs.
- Baselines: classical (features + Ridge/RF), 1D CNN, 1D Transformer. See the GitHub repo for code and results.

## Quickstart (load v2 from Hugging Face)

```python
from datasets import load_dataset
# Load v1 (audio); v2 metadata is in parquet files in the repo
ds = load_dataset("mandipgoswami/rirmega", revision="v2.0.0", trust_remote_code=True)
# If v2 parquet are uploaded as data files:
# import pandas as pd
# meta = pd.read_parquet("metadata_v2.parquet")  # from repo files
```

For full v2 pipeline (build, QC, eval, paper), clone and run:

```bash
git clone https://github.com/mandip42/rirmega-v2-release.git
cd rirmega-v2-release && pip install -e . && python scripts/release_v2.py
```

## Citation

If you use RIRmega v2, please cite the dataset and the paper:

```bibtex
@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}
}

@misc{githubrirmegav2,
  author = {Goswami, Mandip},
  title  = {rirmega-v2-release},
  year   = {2025},
  url    = {https://github.com/mandip42/rirmega-v2-release}
}
```

## License

Metadata and QC artifacts: CC BY 4.0 where applicable. RIR audio content: same license as the upstream RIRmega v1 dataset (see v1 dataset card).