Exploiting hierarchy in environmental sound classification
Jakub Bajzík, Roman Jarina · 2022
Discovering machine learning potential for the computational analysis of audio signals has become a topical issue of recent years. Speech and music have been studied widely. In the field of environmental sounds analysis, there is still a place for developing the systems for several tasks such as classification and detection. This work focuses on environmental sounds classification using convolution neural networks. The novelty of this work is in embedding the hierarchical sound categorization into the classification process. For the training and evaluation of the proposed models, the ESC-50 dataset is used. The dataset comes with 50 fine classes, loosely arranged into 5 coarse categories.