Abstract
Events occur simultaneously at some scenes. The ability that people are able to detect these events and analyse these scenes by listening is called auditory scene analysis, and the study to let computers have this ability is called computational auditory scene analysis. Sound event detection is a topic that relates to computational auditory scene analysis, which focuses on converting acoustic signal to concrete descriptions of its corresponding sound events. This technology can be used in many applications, such as house security, healthcare, and so on. Through methods developed in pattern recognition, acoustic signal can be turned into feature vectors first, then learning methods can be applied to train models with these feature vectors and their corresponding event labels. Since the data used here were recorded in environments with multiple sound sources, polyphonic sound event detection is required such that the system can detect multiple events at the same time. Compared to monophonic sound event detection, polyphonic sound event detection is more complicated and can be viewed as multi-label classification. This research used TUT Sound Events 2016 database, which was published by Detection and Classification of Acoustic Scenes and Events 2016 (DCASE 2016). The baseline system of the database used mel frequency cepstral coefficients for feature extraction and Gaussian mixture models for multi-label classification. This paper tries to improve the performance by introducing another feature extraction method based on a human auditory model and multi-label classification methods based on deep neural networks. After trying different combination of feature extraction method and multi-label classification methods, the proposed method reduces the error rate by 0.04 and increases the F-score by 6.6% compared to baseline.