This paper studies open‑vocabulary sound event detection, the problem of finding when sounds happen in an audio clip using arbitrary text queries (for example, “a man is speaking” or “jackhammer”). Progress was hard to judge because prior papers tested on different datasets and used different rules. The authors build a single benchmark of six temporally annotated tasks — four with fixed class lists across domestic, urban and mixed indoor/outdoor scenes, and two that require free‑text grounding — and they evaluate several recent methods under the same zero‑shot protocol.

On this unified benchmark the authors introduce COSED (Closed‑ and Open‑world SED). COSED is a dual‑encoder system: one network encodes audio into a sequence of frame embeddings, and a separate text encoder turns queries into vectors. A small temporal refinement stage (a two‑layer bidirectional gated recurrent unit, BiGRU) refines audio frames. The system scores every text query against each audio frame with a calibrated cosine similarity; clip‑level scores are kept only as an auxiliary training signal. The audio encoder they use (ATST‑F) was pretrained on AudioSet weak labels and the text encoder comes from prior audio‑text models.

The unified evaluation compares five recent methods plus COSED on all six tasks with the same metrics and a label‑space zero‑shot rule (models cannot train on the test task labels). COSED outperforms prior work on five of the six tasks and matches the best on the sixth. For three tasks the improvement margins reported are large (12–33%). The authors also reproduced published results for earlier systems where possible, to make the comparison fair.

To understand what matters, the paper reports a leave‑one‑out ablation study. It shows three main sources of COSED’s gains. First, scoping negative examples to the corpus they came from (that is, choosing negative training pairs carefully) produced the largest single effect (about 25.8% of the aggregate gain). Second, training with both closed‑world supervision (fixed class queries) and open‑world supervision (free‑text captions) gave a substantial boost (about 16.8%). Third, improving temporal processing with the BiGRU and related choices contributed roughly 16.4%. Other architecture choices mattered too: using an audio‑text pretrained text encoder and a higher frame‑rate audio backbone helped.