Bioacoustic detection with wavelet-conditioned convolutional neural networks

• DOI: 10.1007/s00521-018-3626-7
• Publication Date: 2018-08-01T05:51:30Z
• Authors: Ivan Kiskin; Davide Zilli; Yunpeng Li; Marianne Sinka; Kathy Willis; Stephen Roberts
• Keywords: 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology

Many real-world time series analysis problems are characterized by low signal-to-noise ratios and compounded scarce data. Solutions to these types of often rely on handcrafted features extracted in the or frequency domain. Recent high-profile advances deep learning have improved performance across many application domains; however, they typically large data sets that may not always be available. This paper presents an for acoustic event detection a challenging, data-scarce, problem. We show convolutional neural networks (CNNs), operating wavelet transformations audio recordings, demonstrate superior over conventional classifiers utilize features. Our key result is offer clear benefit more commonly used short-time Fourier transform. Furthermore, we features, particular dataset, do generalize well other datasets. Conversely, CNNs trained generic able achieve comparable results multiple datasets, along with outperforming human labellers. present our both detecting presence mosquitoes classification bird species.