Using robust features with multi-class SVMs to classify noisy sounds

In a sounds recognition system, the most encountered problem is the background noise that can be captured with the sounds to be identified. This paper describes work that has been performed to address this problem. First, the robustness to the environmental noise is investigated for specific kinds of acoustic representation. The representations considered are RASTA-PLP, J-RASTA and wavelets-based processing. Then, we propose to apply multi-class support vector machines (SVMs) as a discriminative framework in order to address audio classification. The experiments conducted on a multi- class problem show that this classifier clearly overperforms the conventional HMM-based system, and hence, we can efficiently address a sounds classification problem characterized by complex real-world datasets, even under important noise degradation conditions.