“…On the basis of previous successes of ML methods applied to audio in the field of music information retrieval, for example, the studies of Downie (10) and Li et al (11), the analysis input in our study is the event spectrogram, computed as short-time Fourier transforms. Previous work on human perception of sonified seismic data demonstrated that humans are surprisingly sensitive to subtle spectral aspects of the signals that reflect physical aspects of the earthquake source and wave propagation (12)(13)(14)(15)(16)(17)(18). The use of the spectrogram (instead of the waveform) as input enables certain pattern matching techniques that lead to unprecedented levels of pattern identification in and among signals [for example, the study of Cotton and Ellis (19)].…”