Robust Acoustic Event Classification for Environmental Monitoring through CLAP-Based Transfer Learning

Environmental noise monitoring requires reliable tools for acoustic source attribution and automatic screening of complex sound environments. Sound event classification systems may operate as black-box decision modules, but their predictions must remain trustworthy in practical monitoring workflows. Contrastive Language–Audio Pretraining (CLAP) provides general-purpose audio representations with strong zero-shot capabilities, but its direct application to environmental monitoring may lead to unbalanced class-wise performance and limited reliability for rare or critical events. This paper presents an engineering-oriented CLAP-based transfer-learning framework for robust acoustic event classification in environmental monitoring. The contribution is not intended as a new deep-learning architecture, but as a reproducible validation of a transparent workflow for adapting general-purpose audio embeddings to a specific monitoring task. The CLAP audio encoder is used as a frozen feature extractor, while a supervised linear head is trained on DataSEC, comprising 22 environmental sound classes from public repositories and field measurements. The experimental protocol explores multiple segment lengths, optimizers and hyperparameters, adopting a four-way split into training, validation, early stopping and test sets. Model selection combines accuracy with the minimum and standard deviation of class-wise F 0.5 to obtain a conservative classifier that penalizes false positives. The selected configuration, trained on 7-second segments with Adam, increases test accuracy from 73.95% to 90.77%, while the macro-averaged F 0.5 rises from 74.84% to 90.78%. The minimum class-wise F 0.5 increases from 22.13% to 64.36%, and its standard deviation is approximately halved. These results demonstrate robustness across classes and support the proposed framework as a decision-support module for intelligent acoustic monitoring.