Context. Biodiversity is declining worldwide, with many species decreasing in both number and range. Acoustic monitoring is rapidly becoming a common survey method in the ecologist’s toolkit that may aid in the conservation of endangered species, but effective analysis of long-duration audio recordings is still challenging.

Aims. The aims of this study were to: (1) develop and test call recognisers for the endangered southern black-throated finch (Poephila cincta cincta) as well as the similar sounding, but non-endangered, double-barred finch (Taeniopygia bichenovii); and (2) compare the ability of these recognisers to detect these species with that of on-ground bird surveys at under-surveyed locations in the Desert Uplands bioregion which is at the edge of the known range of the black-throated finch.

Methods. A range of convolutional neural network call recognition models were built and tested for both target species, before being deployed over new audio recordings collected at 25 sites during 2020, 2021 and 2022, and compared with the results of on-ground bird surveys at those same sites.

Key results. Call recognisers for both species performed well on test datasets from locations in the same area as the training data with an average area under the precision-recall curve (PRAUC) of 0.82 for black-throated finch and 0.87 for double-barred finch. On-ground bird surveys in the Desert Uplands bioregion detected black-throated finches at two locations in different years, and our call recognisers confirmed this with minimal post-validation of detections. Similar agreement between methods were obtained for the double-barred finch, with site occupancy in the Desert Uplands bioregion confirmed with audio recognition in all nine surveys with on-ground detections, as well as during four additional surveys that had no on-ground detections.

Conclusions. Using call recognisers to survey new locations for black-throated finch presence was equally successful as on-ground surveys, and with further refinements, such as retraining models with examples of commonly misclassified vocalisations added to the training data, minimal validation should be required to detect site presence.

Implications. Acoustic monitoring should be considered as a valuable tool to be used alongside manual surveys to allow effective monitoring and conservation of this endangered species.