RIR-Bench-Hard / eval /README.md
mandipgoswami's picture
Add RIR-Bench-Hard: 2,600 adversarial synthetic RIRs, metadata, splits, eval harness, dataset card (part 8)
9f76cac verified
|
Raw History Blame Contribute Delete
3.36 kB
# RIR-Bench-Hard — Evaluation & Leaderboard
This directory is the **adoption magnet**: a self-contained scorer plus the
exact submission format, so anyone can evaluate a dereverberation / room-
acoustics system against RIR-Bench-Hard and report a comparable number.
`evaluate.py` depends on **numpy, scipy, soundfile only** — no deep-learning
stack required.
```bash
pip install numpy scipy soundfile pandas # pandas optional (parquet reader)
python evaluate.py --metadata ../metadata.parquet --pred my_system.csv
```
## What is scored
For every benchmark item the reference metadata carries four acoustic
descriptors. A submission is scored by the **Mean Absolute Error (MAE)** of its
recovered descriptors against the reference, reported **per category**, for the
**`test_hard`** split, and **overall**. Only rows with `valid == True` are
scored (5 of 2600 rows are flagged non-physical and excluded; see the dataset
card's Quality-control section).
| Metric | Definition |
|---|---|
| `t60_measured` | Reverberation time (s). Schroeder backward-integration EDC, line fit on the −5 dB → −35 dB region (T30), extrapolated to 60 dB. |
| `drr_db` | Direct-to-reverberant ratio (dB). Energy in a ±2.5 ms window around the peak vs. the remainder. |
| `c50_db` | Clarity (dB). `10·log10( E[0–50 ms] / E[50 ms–end] )`, referenced to the direct arrival, clamped to ±60 dB. |
| `edt_s` | Early decay time (s). Slope over 0 → −10 dB, extrapolated to a 60 dB decay. |
The estimator source is duplicated verbatim inside `evaluate.py` (and lives in
`scripts/generate_rirs.py`) so your scores use the identical definitions that
built the benchmark.
## Two submission modes
### 1. Predictions CSV (recommended — framework-agnostic)
Produce one row per benchmark `id` with your system's estimate of whichever
metrics it recovers. Unknown metrics may be omitted (they are skipped, not
penalised).
```csv
id,t60_measured,drr_db,c50_db,edt_s
long_t60_0000,1.19,-11.1,1.9,1.20
long_t60_0001,1.28,-5.0,3.0,0.95
...
```
```bash
python evaluate.py --metadata ../metadata.parquet --pred my_system.csv --out my_results.csv
```
### 2. Enhanced-audio directory (signal-level)
Drop your enhanced / dereverberated outputs as `<id>.wav` in a directory. The
harness recomputes the four metrics on **your** audio with the benchmark's own
estimators and reports the same MAE table.
```bash
python evaluate.py --metadata ../metadata.parquet --enhanced-dir my_outputs/
```
### Identity baseline (default)
With neither flag, the harness runs the **identity (do-nothing) baseline**: it
recomputes the metrics on the reference WAVs. Because the reference metrics were
derived from those WAVs, this defines the measurement floor (near-zero MAE,
limited only by int16 storage quantisation). The seeded result ships as
[`baseline_results.csv`](baseline_results.csv) so the leaderboard has a value on
day one.
## Submitting to the leaderboard
Open a discussion / PR on the dataset repo with:
1. your `--out` results CSV (the per-bucket MAE table),
2. a one-line system description and a link to code or a paper,
3. which mode you used (predictions CSV or enhanced-audio).
Report the **`test_hard` overall MAE** as your headline number — that split is
weighted toward the longest-T60 and lowest-DRR samples where dereverberation and
far-field ASR break.