CDCLAP: Efficient distillation framework for zero-shot audio event classification.
Abstract
This paper presents a novel Cross-Modal Distillation Contrastive Language-Audio Pretraining (CDCLAP) framework to address the challenges of zero-shot audio event classification (ZAEC) under limited hardware resources. The proposed method transfers visual knowledge from an ImageNet-pretrained MobileNetV3 to a lightweight student audio network, and employs a CLAP model-comprising HTSAT and natural language processing model GPT-2-as the teacher to enhance the representational capacity and semantic generalization of the audio encoder. Furthermore, a description generation strategy using GPT-2 is introduced to augment the audio-text pairs, thereby improving model performance. Extensive evaluations on three benchmark datasets-ESC50, US8K, and TUT17 demonstrate that CDCLAP achieves superior performance compared to existing state-of-the-art methods, while maintaining significantly lower model complexity. These results validate the effectiveness of the proposed cross-modal distillation strategy in improving zero-shot recognition accuracy and robustness, offering new insights into efficient multimodal audio understanding.