Smart Ambient Sound Analysis via Structured Statistical Modeling

In this paper, we introduce a novel framework called SASA (Smart Ambient Sound Analyser) to support different ambient audio mining tasks (e.g., audio classification and location estimation). To gain comprehensive ambient sound modelling, SASA extracts a variety of acoustic features from different sound components (e.g., music, voice and background), and translates them into structured information. This significantly enhances quality of audio content representation. Further, distinguished from existing approaches, SASA’s multilayered architecture seamlessly integrates mixture models and aPEGASOS (adaptive PEGASOS) SVM algorithm into a unified classification framework. The approach can leverage complimentary strengths of both models. Experimental results based on three large test collections demonstrate the SASA’s advantages over existing methods on various analysis tasks.