For every second of a recording, predict which of fifteen everyday sound events are happening — footsteps, running water, a microwave, a light switch. A sparse, imbalanced, multi-label problem where the entire challenge is honest evaluation. Built as a JKU machine-learning project, finished in a competitive challenge that reached 0.70 macro-F1.
The dataset is a crowd-collected library of everyday indoor sounds: 3,656 recordings, cut into roughly 168,000 one-second segments, each described by 960 precomputed audio features and labelled across fifteen event classes — footsteps, running water, keyboard typing, doors, cutlery and dishes, microwave, vacuum, light switch, and more. The task is multi-label and segment-level: any combination of events can overlap within a single second. Crucially, the labels are extremely sparse — 94% of all class slots are silent — and imbalanced, with the most common event roughly 29× more frequent than the rarest. A model that simply predicts "nothing, everywhere" scores 94% accuracy, so accuracy is meaningless. The real objective was a trustworthy evaluation and a model that genuinely beats trivial baselines.
Most of the effort went into evaluation discipline, because that is where audio benchmarks quietly leak. Recordings are split by collector, so no contributor — and none of their overlapping segments — appears in both training and test; measured cross-split overlap is exactly zero. Labels from one to five annotators per file are reconciled by majority vote. Every feature is standardized using statistics fit on the training split only; fitting on the full dataset would shift per-feature scale by up to 11%, a leak I quantified rather than ignored. Performance is reported as macro-F1 and macro-AP against three reference points: an always-silent baseline (0.00), a class-prior baseline (0.06), and a human annotator-agreement ceiling (0.78).
On that footing, logistic-regression and random-forest classifiers are swept across regularization strength and tree depth, producing clean bias-variance curves; logistic regression settles at 0.38 macro-F1, comfortably above the prior. A case study then runs the model on unseen recordings, pairing mel-spectrograms with predicted-versus-true timelines and a per-class error-overlap matrix. It makes the structure of the errors legible: continuous sounds like running water and vacuum are easy, while brief transients — light switch, window, wardrobe — are genuinely hard.
A follow-on challenge pushed for the highest achievable macro-F1, and the progression was itself the lesson. Tuning per-class decision thresholds, instead of leaving them at the default 0.5, lifted logistic regression from 0.41 to 0.50; lightly median-filtering the prediction timeline added a little more. A CRNN trained from scratch on log-mel spectrograms matched the classical system at 0.51 — and no further.
The decisive move was transfer learning. Frozen embeddings from an AudioSet-pretrained model with a small BiGRU head reached 0.70 macro-F1 — and, tellingly, fine-tuning that same backbone end-to-end did worse (0.60), a textbook small-data result. Across both phases the consistent finding held: calibration and leakage control moved the score more than raw model capacity.
The technical report, per-class evaluation, error-analysis figures, and the challenge pipeline can be shared for review.