SELD-Synth-Mini

A compact synthetic Sound Event Localization and Detection (SELD) set in 4-channel First-Order Ambisonics, DCASE-style polar metadata.

SELD is the most active structured spatial-audio task (DCASE Task 3). SELD-Synth-Mini is a small, fully-reproducible synthetic set for prototyping and sanity-checking SELD pipelines: 600 ten-second FOA clips (ACN channel order, SN3D normalization, 16 kHz), each containing 1–3 overlapping sound events of distinct classes, every event spatialized through a real room impulse response from the RIR-Bench-Hard bank with exact direction-of-arrival, onset/offset, and class labels in DCASE-SELD polar metadata format.

from datasets import load_dataset
ds = load_dataset("mandipgoswami/seld-synth-mini", split="dev")
print(ds[0])   # clip_id, wav_path, meta_path, n_events, classes, events_json, ...

Why this dataset

• Drop-in DCASE-SELD format. Per-clip CSV: frame,class,source,azimuth,elevation,distance at a 100 ms frame hop — the metadata layout SELD toolkits expect.
• Spatial ground truth from real rooms. Each event's DOA comes from an actual RIR's geometry; reverberation is physically real.
• Small and instant. 600 clips total — fast to download and iterate on; a sanity set, not a training mega-corpus.
• Reproducible & license-clean. Every event is seeded synthetic audio (see Honesty note).

Composition

clips	600 (10 s each, 16 kHz, 4-ch FOA ACN/SN3D)
total events	1,191
polyphony	1 event: 203 clips · 2 events: 203 · 3 events: 194
total audio	~768 MB
frame hop	100 ms

Splits: train 400 · dev 100 · test 100.

Classes (10): speech, alarm, knock, whistle, noise_burst, engine, bell, clap, footstep, beep — each 96–133 events, roughly balanced.

Format

• data/foa/clip_XXXX.wav — 4-channel FOA mixture.
• data/metadata_dev/clip_XXXX.csv — DCASE-SELD polar labels, one row per active (frame, source):

  frame,class,source,azimuth,elevation,distance
  12,3,0,45.0,10.0,3.21
  

  class is the index into the class list above; azimuth ∈ [−180,180], elevation ∈ [−90,90] degrees; distance in metres.
• metadata.parquet / .csv — top-level manifest with events_json (per-clip event list: class, source, onset_s, offset_s, az, el, distance).

Task & metric

Joint detection + localization: for each frame, predict active classes and their DOA. Standard DCASE-SELD metrics apply (location-dependent F-score / error rate, and DOA error). A dependency-light reference scorer and an oracle/first-event baseline are in eval/.

Honesty note (read before using)

Every sound event is seeded synthetic audio (procedurally generated per class — e.g. harmonic/formant "speech", filtered-noise bursts, decaying impacts), not recorded natural sound. This is deliberate for a reproducible evaluation/prototyping set: labels are exact and the whole set regenerates from seeds. It does not replace recorded SELD corpora (e.g. STARSS) for final evaluation. The room impulse responses are real (physics-simulated).

Related

• RIR-Bench-Hard — the RIR bank these events are spatialized through
• foa-acoustics-bench — companion FOA parameter/DOA benchmark
• audio-eval-suite · reverb-speech-mini

Limitations

• Synthetic (not recorded) sound events — see Honesty note.
• First-order ambisonics only; static sources (no moving trajectories).
• One RIR (direction) per event; no moving-source Doppler.

Citation

@misc{goswami2026seldsynthmini,
  title  = {SELD-Synth-Mini: A Compact Synthetic Sound Event Localization and Detection Set in First-Order Ambisonics},
  author = {Mandip Goswami},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/mandipgoswami/seld-synth-mini}}
}

License: CC-BY-4.0.