Performance of Sound Source Localisation For Bird Calls in Native New Zealand Bush

Abstract

This study directly compares the performance of existing source localisation algorithms using microphone arrays for bird call localisation in native New Zealand bush. Experiments were carried out to investigate the performance of generalised cross correlation with phase transform (GCC-PHAT) and multiple signal classification (MUSIC), and their variants using recordings of North Island brown kiwi calls played from a loudspeaker acquired in a native New Zealand bush environment. Four different microphone array configurations were tested at two locations in the bush. Experimental results showed that GCC-PHAT had the best accuracy for most array configurations. Conventional narrowband MUSIC consistently performed the worst, whereas the wideband MUSIC was comparable to GCCPHAT in some tests however the offset in the estimated angle was much higher. A variant of wideband MUSIC performed similarly to conventional wideband MUSIC, but it did not outperform GCC-PHAT. The small triangle array achieved the best performance by a large margin. The results also imply that the square array shape may perform better than a triangular array when their sizes are similar.