Context-based sound event recognition

The research domain of automatic sound event recognition aims to describe an audio signal in terms of the sound events that compose a sonic environment. The ability to recognize events in a real‐world environment requires a listener or system to separate the sound events from each other and the background. Furthermore, these separated events need to be recognized. To recognize a sound event implies that some representation of the event is already known to the receiver, and can be identified when it is encountered again. This representation is limited to an acoustic description in many automatic systems, which complicates the recognition of ambiguous sound events or sound events in background noise. For example, a sound of a purring cat may seem unique, but without any other information than represented by the audio signal, it can sound like an engine as well. In contrast to these types of systems, people do not only use acoustic information when they recognize sound events, but also the context in which the event occured. In this thesis we test how poeple benefit from knowledge about the context when they recognize sound events. Subsequently, we show that the task of recognizing sounds event can be alleviated with a similar approach, through a model of human memory that represents the context in which the event occurs.