Multi-level attention model with deep scattering spectrum for acoustic scene classification

Acoustic scene classification (ASC) refers to the classification of audio into one of predefined classes that characterize the environment. People are used to combine log-mel filterbank features with convolutional neural network (CNN) to build ASC system. In this paper, we explore the use of deep scattering spectrum (DSS) features combined with a multi-level attention model based on CNN for ASC tasks. First, the time scatter and frequency scatter coefficients of DSS with different resolutions are explored as ASC features. Second, we incorporate a multi-level attention model into CNN to build the classification system. We then evaluate the proposed approach on the IEEE challenge of detection and classification of acoustic scenes and events 2018 (DCASE 2018) dataset. Results show that the DSS features provide between a 11%-14% relative improvement in accuracy over log-mel features, within a state-of-the-art framework. The application of multilevel attention model on CNN can improve the accuracy by nearly 5%. The highest accuracy of our proposed system is 78.3% on the development set.