A Re-trained Model Based On Multi-kernel Convolutional Neural Network for Acoustic Scene Classification

Abstract

This paper proposes a deep learning framework applied for Acoustic Scene Classification (ASC), which identifies recording location. In general, we apply three types of spectrograms: Gammatone (GAM), log-Mel and Constant Q Transform (CQT) for front-end feature extraction. For back-end classification, we present a re-trained model with a multi-kernel CDNN-based architecture for the pre-trained process and a DNN-based network for the post-trained process. Our obtained results over DCASE 2016 dataset show a significant improvement, increasing by nearly 8% compared to DCASE baseline of 77.2%.