https://biblioteca.sbrt.org.br/articles/1929
https://repositorio.ufsc.br/handle/123456789/193287
This paper presents new contributions aiming at enhancing the performance of the baseline acoustic scene classifier proposed in the context of the Detection and Classification of Acoustic Scenes and Events 2018 (DCASE2018) Challenge. These contributions consist on modifications of the structure of the baseline convolutional neural network as well as on the use of data augmentation and model ensemble strategies. As a result, an accuracy of 72.04% is obtained for the development dataset, whereas 68.5% accuracy is obtained for the evaluation dataset. This corresponds to a significantly better performance in comparison with the baseline system, which is of 59.7% and 61.0%, respectively.