Acoustic Scene Classification Using Spectrograms

For human beings, it's an easy task to hear an environment sound and identify from where it belongs. But for machines, it's different. Thinking on that, the Acoustic Scene Classification task was raised with the objective of automatically classify the place/environment where an audio sample was originally recorded. Thereat, this work investigates the performance of an automatic classification system, using the database of the DCASE 2016 challenge, that recognizes environments/places where an audio sample was originally recorded, considering 15 different categories. The sample's audio signal was converted to a spectrogram and after, features were extracted using texture descriptors. The complementary between the existing signals from the left and right channels of an digital audio recorded in stereo was investigated. Combinations using acoustic and visual features were tested as well, looking for an improvement in the classification rate. After all, an accuracy of 80.17% was reached, surpassing the original baseline provided by the DCASE 2016 challenge itself, that presented accuracy of 77.2%. Then, we have concluded that the features obtained in the visual domain can support the development of an efficient classification system in this application.