Abstract: Aiming at the problems of insufficient acoustic  feature extraction under sample scarcity conditions, we proposed a small sample sound event detection method based on channel  and spatial compression.  The method constructed a dual compression attention mechanism to screen features  in channel dimension and achieved  feature focusing in the spatial dimension, effectively improving the feature discrimination ability of the prototype network in small sample scenarios. The experimental results show that F₁-score  of the method on the dataset DCASE (detection and classification of acoustic scenes and events) reaches 66.84%, an improvement of 4.11 percentage points compared to the prototypical network, providing more reliable technical  support for practical applications such as wildlife monitoring and ecological environment assessment.

Key words: sound event detection, prototype network, channel attention, spatial attention