Learning to Tag Environmental Sounds in Nightlong Audio

Kurzfassung

Environmental sound events are defined as sounds occurring naturally or produced due to human activity. Devices need to identify the environmental sounds for better perception of the environment. The emergence of environmental sound classification using machine learning has led to the development of more context-aware technologies like smart homes, multimedia search, etc. In this thesis, different sound events and their starting and ending times are determined using classification models. The audio data is provided by the German Aerospace Center and is composed of night-long audio clips that are recorded near an airport and contain different environmental sounds. Many challenges like noise in data, sounds of different lengths, and their effects on the classifier are also discussed. We investigated if different models prefer to identify sounds of different lengths in the same audio. An overlapping window approach is devised to improve the identification of starting and ending times of the predicted events. The results of the thesis are encouraging as the best model is able to identify various sound events with an accuracy of 0.94 on a balanced dataset of 4 different classes. Finally, a system is also conceptualized where the environmental sounds are identified, and their spans are visualized against time.