Acoustic Scene Classification using Binaural Representation and Classifier Combination

Detection and Classification of Acoustic scene is a subtask of DCASE 2017 challenge which is trying to classify noisy structured sounds to predefinedclasses. This is a challenging task due to the content of audio signals and the lack of enough data. Thus most of the recent works used different classifier ensemble methods for acoustic scene classification. In this paper, we use Harmonic-Percussive Source Separation (HPSS) to decompose audio spectrogram to its constructing components and then use its harmonic component. After that, we propose to use different audio forms based on a binaural representation of sound recordings. We also use multilayer perceptron (MLP) neural networks as our classifier and propose two weighing techniques for classifier combination: inverse of entropy at softmax layer output and binary weights for the classifiers. The proposed methods outperform the baseline system of DCASE 2017. The entropy based weighing and binary weighing methods achieved 70.55% and 72.09% accuracy on evaluation dataset of DCASE 2017 challenge in comparison to 61% accuracy of DCASE 2017 baseline system.