Datasets:
File size: 5,685 Bytes
0280d8f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 | ---
tags:
- automatic-speech-recognition
- robustness
- speech
- benchmark
license: cc-by-4.0
---
# Whisper-RIR-Mega: Paired Clean↔Reverberant Speech Robustness Benchmark
## Dataset Summary
**Whisper-RIR-Mega** is a benchmark dataset of *paired* clean and reverberant speech for evaluating ASR robustness to room acoustics. Each sample consists of:
- **audio_clean**: Clean speech (LibriSpeech test-clean, 16 kHz)
- **audio_reverb**: Same utterance convolved with one RIR from [RIR-Mega](https://huggingface.co/datasets/mandipgoswami/rirmega) (v2)
- **text_ref**: Ground-truth transcript
- **RIR metadata**: `rir_id`, RT60, DRR, C50, etc. when available
- **Technical paper**:([Whisper-RIR-Mega](https://arxiv.org/abs/2603.02252))
Splits are stratified by RT60 (or DRR) when metadata exists, so the benchmark is balanced across acoustic conditions.
**Use this dataset to:**
- Benchmark Whisper (or any ASR) on clean vs. reverberant speech and report **reverb penalty** (Δ WER)
- Evaluate robustness across RT60/DRR bins
- Reproduce the official Whisper-RIR-Mega leaderboard
---
## 30-Second Quickstart
```python
from huggingface_hub import snapshot_download
from datasets import load_from_disk
import whisper
import jiwer
# Download and load (use load_dataset if your HF datasets supports it)
path = snapshot_download("mandipgoswami/whisper-rirmega-bench", repo_type="dataset")
ds = load_from_disk(path + "/hf_dataset")["test"]
model = whisper.load_model("base")
# One sample
row = ds[0]
clean_wer = jiwer.wer(row["text_ref"], model.transcribe(row["audio_clean"]["path"], language="en")["text"])
reverb_wer = jiwer.wer(row["text_ref"], model.transcribe(row["audio_reverb"]["path"], language="en")["text"])
print(f"Clean WER: {clean_wer:.4f} Reverb WER: {reverb_wer:.4f}")
```
---
## Dataset Structure
| Column | Type | Description |
|---------------|--------|--------------------------------------|
| sample_id | string | Unique ID (from LibriSpeech + RIR) |
| audio_clean | Audio | Clean 16 kHz audio |
| audio_reverb | Audio | Reverberant 16 kHz audio |
| text_ref | string | Reference transcript |
| rir_id | string | RIR-Mega sample ID |
| split | string | train / validation / test |
| rir_* | mixed | RIR metadata (RT60_T30_s, DRR_dB, …)|
Splits: **validation** and **test** for benchmarking; **train** optional (default config uses test + validation only).
---
## How It’s Built
1. **Speech**: LibriSpeech test-clean (CC BY 4.0), streamed from Hugging Face.
2. **RIRs**: [mandipgoswami/rirmega](https://huggingface.co/datasets/mandipgoswami/rirmega) (v2.0.0), with metadata (RT60, DRR, C50, etc.).
3. **Pipeline**: For each utterance we sample one RIR (stratified by RT60), convolve at 16 kHz, normalize RIR energy and peak-normalize output. No added noise by default.
4. **Splits**: Deterministic assignment to validation/test (e.g. 20% / 80%) with optional stratification by acoustic bins.
Full reproducibility: see the [GitHub repo](https://github.com/mandipgoswami/Whisper_RIRMega) and run:
```bash
python -m bench.build_and_publish --config configs/default.yaml
```
---
## Leaderboard
The leaderboard is generated by the same pipeline and updated on each release. Example (your run may vary):
| model_id | clean | reverb | Δ WER |
|--------------------|---------|---------|---------|
| openai/whisper-tiny | … | … | … |
| openai/whisper-base | … | … | … |
| openai/whisper-small| … | … | … |
| openai/whisper-medium| … | … | … |
| openai/whisper-large-v3| … | … | … |
See the [Space](https://huggingface.co/spaces/mandipgoswami/whisper-rirmega-benchmark) for interactive charts (WER vs RT60/DRR) and the latest leaderboard.
---
## Limitations
- English only (LibriSpeech).
- Single RIR per utterance in the default setup; multi-RIR variants can be built by changing `k_rirs_per_utt` in the config.
- RIR metadata (RT60, DRR) may be missing for some RIR-Mega samples; the pipeline stores whatever is available.
---
## License & Citation
- **Speech**: LibriSpeech (CC BY 4.0).
- **RIRs**: RIR-Mega license (see [mandipgoswami/rirmega](https://huggingface.co/datasets/mandipgoswami/rirmega)).
- **Benchmark curation**: MIT (this repo).
**Citation (BibTeX):**
```bibtex
@misc{whisper-rirmega-bench,
title = {Whisper-RIR-Mega: Paired Clean-Reverberant Speech Robustness Benchmark},
author = {Goswami, Mandip},
year = {2025},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/mandipgoswami/whisper-rirmega-bench},
note = {Dataset built with LibriSpeech and RIR-Mega.}
}
```
**RIR-Mega citation:**
```bibtex
@misc{goswami2025rirmega,
title = {RIR-Mega: A Large-Scale Room Impulse Response Corpus with Benchmarks},
author = {Goswami, Mandip},
year = {2025},
eprint = {2510.18917},
archivePrefix= {arXiv},
primaryClass = {cs.SD},
url = {https://arxiv.org/abs/2510.18917}
}
```
---
## How to Reproduce
1. Clone the repo and install: `pip install -e .`
2. Set `HF_TOKEN` (and optionally reduce `n_utterances` in `configs/default.yaml` for a quick run).
3. Run: `python -m bench.build_and_publish --config configs/default.yaml`
4. This builds the dataset, runs Whisper baselines, generates reports, and can push the dataset and Space to the Hub (if `HF_TOKEN` is set).
For a <5 minute smoke test: `python scripts/sanity_check.py`
|