This work proposes bag-of-features deep learning models for acoustic scene classi?cation (ASC) – identifying recording locations by analyzing background sound. We explore the effect on classi?cation accuracy of various front-end feature extraction techniques, ensembles of audio channels, and patch sizes from three kinds of spectrogram. The back-end process presents a two-stage learning model with a pre-trained CNN (preCNN) and a post-trained DNN (postDNN). Additionally, data augmentation using the mixup technique is investigated for both the pre-trained and post-trained processes, to improve classi?cation accuracy through increasing class boundary training conditions. Our experiments on the 2018 Challenge on Detection and Classi?cation of Acoustic Scenes and Events - Acoustic Scene Classi?cation (DCASE2018-ASC) subtask 1A and 1B signi?cantly outperform the DCASE2018 reference implementation and approach state-of-the-art performance for each task. Results reveal that the ensemble of multi-spectrogram features and data augmentation is bene?cial to performance.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=20465
(381KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
