Mix-Clap: Adaptive Fusion of Knowledge-Distilled Audio Embeddings for Noise-Aware Audio-Language Models
Real-world deployment requires sound event and acoustic scene classification systems to remain reliable in noisy, diverse environments on resource-constrained devices. Although contrastive language-audio pretraining (CLAP) models with Transformer-based audio encoders achieve strong zero-shot performance, their computational cost hinders deployment. In this paper, we propose Mix-CLAP, a computationally efficient, noise-aware CLAP model with knowledge-distilled audio encoders. Our method includes: (1) a two-stage knowledge distillation from teacher embeddings to two lightweight student encoders?one on clean audio, the other on noisy audio, and (2) adaptive inference that combines their embeddings together with a fusion parameter and minimizes the parameterized entropy at test time. Experiments show that Mix-CLAP with MobileNetV3-based audio encoders greatly improves computational efficiency, while achieving a comparable average accuracy of 52.58% to the Transformer-based CLAP model at 52.83% on the recorded ESC50 datasets with different devices including microphones and fiber-optic distributed acoustic sensors under diverse conditions, making it suitable for real-world, resource-constrained applications.
