Here are proposed two automatic detectors of Barau's petrel (Pterodroma baraui) and tropical shearwater (Puffinus bailloni) vocalisations in noisy audio recordings (1) trained with a low number of positive training instances, and (2) whose performances would be the highest possible. To do so, acoustic recordings were performed in one Barau's petrel colony between February and May 2014 (85 h) and in two tropical shearwater colonies in March and April (21 h). Manual and automatic methods of segmentation were combined. Manual segmentation allowed (1) to miss a very few number of positive segments and (2) to avoid introducing false positive instances. Automatic segmentation provided quickly a diversified set of negative instances. Manual labelling must be regarded as an investment, for current and future works. A random forest classifier and classical methods of acoustic signal characterisation (cepstral coefficients, spectral moments, etc.) were tested. Best models were able to discriminate each target species calls from other sounds of its colony with F1 scores of 88% (Barau's petrel, 1015 samples) and 85% (tropical shearwater, 1217 samples). The acoustic monitoring of nocturnal burrow-nesting seabirds based on (1) data collected by autonomous recording units in harsh, windy and wet environments and (2) automatic analysis tools is feasible. The size of our database was limited. Consequently further works will be necessary to study robustness of models on long time-series data.