PSDS for Sound Event Detection Evaluation

Sound-event detection tries to say which sounds occurred and when. A model often outputs a confidence trace for each class, and a threshold turns that trace into discrete events. An F1 score at one threshold can change substantially if the threshold was chosen differently. The PSDS framework was developed to compare detectors over a range of operating points, using a polyphonic detection receiver-operating-characteristic view. Its original paper emphasizes robust evaluation when multiple event classes and overlaps occur. Later work introduced efficient threshold-independent computation from event scores. Evaluation still depends on choices. A detection must be matched with a reference event under temporal-intersection or related criteria. The allowed overlap, treatment of cross-class confusions and false-positive cost affect the score. DCASE tasks have used scenarios with different emphases, such as tighter timing in one and class-confusion penalties in another. A score without scenario and evaluation code is hard to interpret. It is not a universal percentage of correct events. PSDS reduces the opportunity to cherry-pick one threshold, but it does not solve data-quality or deployment questions. If reference onsets are wrong or an event class is missing, the result can mislead. A high area under an operating curve can still conceal a poor operating point at the false-alarm rate a real product can tolerate. Report class-level behavior and show performance near the intended threshold in addition to the summary. For an alarm, one false wake-up per day may matter more than a small benchmark-average gain. The metric is for detection, not source separation or listener preference. A model can score well on timed events while sounding poor if used to generate audio, and a clip tagger without event times cannot be judged as a temporal detector without further outputs. Use strong reference intervals, specify preprocessing and postprocessing, and evaluate under representative background conditions. PSDS supports comparison; a deployment decision still needs a user-centered error budget.