• PADO introduces dynamic prompt tuning to optimize class distribution learning in semi-supervised sound event detection, effectively mitigating noisy interference from pseudo-labels and domain gaps.
• The method achieves significant performance improvements over state-of-the-art approaches on DCASE 2019, 2020, and 2021 challenge datasets.
• PADO maintains model generalization while improving the efficiency of class distribution learning, addressing the trade-off between pseudo-label quality and quantity.
• The framework is readily extendable to other benchmark models, demonstrating versatility beyond specific SSED architectures.