Exploring the role of temporal dynamics in acoustic scene classification

Identification of acoustic scenes often relies on finding the most informative features that best characterize the physical nature of sound sources in the scene. In this paper, we propose a framework that provides a detailed local analysis of spectro-temporal modulations augmented with generative modeling that map both the average modulation statistics of the scene using Gaussian Mixture Modeling (GMM) as well temporal trajectories of these modulations using Hidden Markov Modeling (HMM). Our analysis shows that a hybrid system of these two representations can capture the non-trivial commonalities within a sound class and differences between sound classes. The proposed hybrid system outperforms current systems in the literature by about 30 % and surpasses the performance of the individual GMM and HMM systems suggesting that these representations provide complimentary information about acoustic scenes.