A Simple fusion of deep and shallow learning for acoustic scene classification

Comunicació presentada a: 15th Sound and Music Computing Conference (SMC2018). Sonic crossing, celebrat a Limassol, Xipre, del 4 al 7 de juliol de 2018.In the past, Acoustic Scene Classification systems havebeen based on hand crafting audio features that are input toa classifier. Nowadays, the common trend is to adopt datadriven techniques, e.g., deep learning, where audio repre-sentations are learned from data. In this paper, we proposea system that consists of a simple fusion of two methods ofthe aforementioned types: a deep learning approach wherelog-scaled mel-spectrograms are input to a convolutionalneural network, and a feature engineering approach, wherea collection of hand-crafted features is input to a gradientboosting machine. We first show that both methods pro-vide complementary information to some extent. Then, weuse a simple late fusion strategy to combine both meth-ods. We report classification accuracy of each method in-dividually and the combined system on the TUT AcousticScenes 2017 dataset. The proposed fused system outper-forms each of the individual methods and attains a classifi-cation accuracy of 72.8% on the evaluation set, improvingthe baseline system by 11.8%.This work is partially supported by the European Union’sHorizon 2020 research and innovation programme undergrant agreement No 688382 “AudioCommons”, and theEuropean Research Council under the European Union’sSeventh Framework Program, as part of the CompMusicproject (ERC grant agreement 267583), and a Google Fac-ulty Research Award 2017. We are grateful for the GPUsdonated by NVidia