In the emerging field of acoustic novelty detection, most research efforts are devoted to probabilistic approaches such as mixture\nmodels or state-space models. Only recent studies introduced (pseudo-)generative models for acoustic novelty detection with\nrecurrent neural networks in the form of an autoencoder. In these approaches, auditory spectral features of the next short term\nframe are predicted from the previous frames by means of Long-Short Term Memory recurrent denoising autoencoders. The\nreconstruction error between the input and the output of the autoencoder is used as activation signal to detect novel events. There is\nno evidence of studies focused on comparing previous efforts to automatically recognize novel events fromaudio signals and giving\na broad and in depth evaluation of recurrent neural network-based autoencoders. The present contribution aims to consistently\nevaluate our recent novel approaches to fill this white spot in the literature and provide insight by extensive evaluations carried out\non three databases:A3Novelty, PASCALCHiME, and PROMETHEUS.Besides providing an extensive analysis of novel and state-ofthe-\nart methods, the article shows how RNN-based autoencoders outperformstatistical approaches up to an absolute improvement\nof 16.4% average
Loading....