FOA-Acoustics-Bench

Blind room-acoustic parameter estimation and source direction-of-arrival (DOA) from 4-channel First-Order Ambisonics (FOA).

FOA-Acoustics-Bench promotes single-channel blind acoustic analysis to the spatial / multi-channel setting that drives most current structured-audio research (DCASE SELD, FOA parameter estimation). Each item is a 3-second, 16 kHz, 4-channel FOA (ACN channel order, SN3D normalization) recording produced by convolving a seeded, license-clean synthetic excitation with a real room impulse response from the RIR-Bench-Hard bank and encoding it along the geometric direct-path direction. Every clip is annotated with ground-truth acoustic parameters and source azimuth/elevation.

from datasets import load_dataset
ds = load_dataset("mandipgoswami/foa-acoustics-bench", split="test_hard")
print(ds[0])   # foa_path, t60_measured, drr_db, c50_db, edt_s, azimuth_deg, elevation_deg, ...

Why this dataset

• Spatial, not mono. 4-channel FOA with exact DOA labels — the format active SELD/spatial pipelines consume.
• Dual task. (a) regress {T60, DRR, C50, EDT} from the ambisonic signal; (b) estimate source azimuth/elevation. Both from audio alone.
• Adversarial acoustics. Inherits RIR-Bench-Hard's five hard regimes (long T60, coupled rooms, near-field, low-DRR, non-shoebox), so reported numbers reflect deployment-hard conditions, not easy rooms.
• Reproducible & license-clean. Sources are seeded synthetic speech-like probe signals, not real speech (see Honesty note); the whole set regenerates from seeds.

Composition

	count
clips	2,595 (one per valid RIR)
channels	4 (FOA, ACN/SN3D: W, Y, Z, X)
sample rate	16 kHz
duration	3.0 s each
total audio	~997 MB

Splits (inherited from RIR-Bench-Hard, weighted toward hard conditions): train 1,946 · dev 260 · test_hard 389. Regimes: long_t60 520 · low_drr 520 · near_field 520 · non_shoebox 519 · coupled_rooms 516.

Measured label ranges (per regime)

regime	T60 (s)	DRR (dB)	C50 (dB)	EDT (s)
long_t60	0.55 – 3.25	−16.7 – 10.2	−3.9 – 13.5	0.12 – 2.26
low_drr	1.94 – 8.39	−16.0 – −5.7	−1.6 – 5.1	0.86 – 2.00
near_field	0.21 – 0.48	5.6 – 24.8	11.2 – 33.6	0.02 – 2.40*
non_shoebox	0.09 – 0.42	−14.0 – 5.5	1.2 – 34.0	0.15 – 0.93
coupled_rooms	0.14 – 0.77	−11.9 – 0.4	−2.3 – 56.7	0.34 – 1.61

*A small number of near-impulsive RIRs yield unstable EDT estimates; these are retained but EDT should be treated with caution on near_field (see Limitations).

Fields

id, category, split, foa_path (4-ch WAV), fs, duration_s, t60_measured, drr_db, c50_db, edt_s, azimuth_deg [−180,180], elevation_deg [−90,90], distance_m, room_dims, source_excitation, seed, valid.

Tasks & baselines

Task A — acoustic-parameter regression (primary): predict {T60, DRR, C50, EDT} from the FOA signal; metric = mean per-descriptor MAE on test_hard, lower is better.

Task B — source DOA: estimate (azimuth, elevation); metric = mean great-circle angular error (deg).

A reference active-intensity DOA baseline is included in eval/. Honest note on its number: because each clip encodes a single direct-path direction into an otherwise clean mix, the intensity estimator recovers azimuth/elevation near-exactly (mean angular error ≈ 0° on test_hard). This confirms the ambisonic encoding is physically correct, but it also means Task B is only meaningful under added interference/diffuseness — so Task B is positioned as a correctness check and a starting point for multi-source extensions, while Task A (parameter regression under adversarial reverberation) is the substantive benchmark.

Honesty note (read before using)

The source audio is synthetic speech-like probe signal (seeded harmonic+formant voiced segments with fricative-like unvoiced bursts), not real speech. This is a deliberate, defensible choice for an evaluation set: it is fully reproducible, license-clean, and isolates the acoustic channel. It does not substitute for natural-speech evaluation. The room impulse responses are real (physics-simulated) and are the acoustic ground truth.

Related

Part of a room-acoustics evaluation program:

• RIR-Bench-Hard — the single-channel adversarial RIR benchmark this is built on
• audio-eval-suite · reverb-speech-mini
• seld-synth-mini — the companion synthetic SELD set

Limitations

• Synthetic (not natural) sources — see Honesty note.
• First-order only (no higher-order ambisonics).
• Single direct-path source per clip; diffuse/multi-source conditions are future work.
• EDT on near-impulsive near_field RIRs can be numerically unstable.

Citation

@misc{goswami2026foaacousticsbench,
  title  = {FOA-Acoustics-Bench: Blind Room-Acoustic Parameter and Direction-of-Arrival Estimation from First-Order Ambisonics},
  author = {Mandip Goswami},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/mandipgoswami/foa-acoustics-bench}}
}

License: CC-BY-4.0.