Checklist APOAI 2026 Contest · Task 2
Audio Event Detection
Classify 3-second telephone-like audio clips into eight acoustic event classes from pre-computed log-mel spectrograms, with higher weight on human sounds.
The task
Each 3-second clip at 16 kHz simulates a phone-call scenario (speech, background noise and occasional acoustic events) and belongs to exactly one of eight classes: cough, sneeze, laughter and cry (weight 2.0), and dog_bark, siren, noise and none (weight 1.0).
There are 24,000 training, 4,000 validation and 4,000 test clips (per-class counts are given in the statement). Pre-computed log-mel spectrograms (64 mel bins, n_fft 1,024, hop 160, shape [64, 300], log10(max(S, 1e-10))) are provided as .npy files with labels in train.csv; the validation and test sets are only available inside the evaluation environment through environment variables.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
- Training log-mel spectrograms (.npy) and
train.csv(id, split, label,wav_path,mel_pathand synthesis metadata such as event/background SNR). - You submit
- A notebook named
submission.ipynb(training and prediction) that writessubmission.zipcontainingsubmission_val.jsonandsubmission_test.json, each a flat JSON object mapping sample ID to one of the 8 labels. - Scoring
- Weighted macro-F1: per-class F1 averaged with weights 2.0 (cough, sneeze, laughter, cry) and 1.0 (
dog_bark, siren, noise, none). Leaderboard A = validation, B = test. Baseline on B: 0.5886; scientific committee reference: 0.7030. - Rules
- Pretrained model weights are not allowed, including weights pretrained on ImageNet or any other external dataset.
- External large language model APIs (e.g. GPT, Claude) may not be used for prediction, feature generation, data labelling or model ensembling.
- Training + inference within 25 minutes on a Tesla L20 GPU.
- Use the pre-computed log-mel spectrograms; computing STFT from raw audio is not needed.
- Evaluated in the 'ioai3.6' image.
- Format
- APOAI 2026 contest (first edition): a six-hour examination on Saturday 27 June 2026, taken online on the Bohrium platform with local invigilated sites; the four tasks were republished on Bohrium for practice from 15 June to 31 July 2026 (platform time UTC+8).