To build a training dataset of calls for recogniser development, we first manually selected and extracted a minimum of 100 reference calls for each species using Adobe Audition CC and Raven Pro 1.5 software. To maximise representativeness, we (a) selected calls from as many surveys sites as possible, to capture geographical variation, and (b) selected calls of varying quality and amplitude, to capture soundscape variation. These are important steps to improve the similarity between call templates and ‘real-world’ sound data. Building a recogniser solely from calls that are loud and clear would perform poorly if the species’ calls are rarely loud and clear in field recordings. Given the complexity of frog choruses, variations in ambient noise and differences in amplitude among calls (e.g. from variations in the distance of the frog from the sound recorder), capturing diversity in call templates is a critical component of recogniser construction.