Segmentation of bio-signals in field recordings using fundamental frequency detection

Monitoring animal species by means of the automatic sound recognition is nowadays a research field of high interest. One of the challenges of this area lies in the segmentation of the species vocalizations. Recordings acquired in natural habitats are contaminated with the sounds emitted by other species and different kinds of background noise. If the data is “clean” a robust segmentation is feasible, otherwise a pre-processing stage must be carefully performed in order to recover relevant information over the noise effect. In this manuscript, we propose the use of the Karnuhen-Loeve Transform noise reduction algorithm as an additional pre-processing sub-stage. We also propose a segmentation method based on the Fundamental Frequency detection and Voiced/Unvoiced segmentation, implemented with the Voice Analysis software Praat. Preliminary results show that the proposed pre-processing scheme improves the segmentation process, with results comparable to the commercial software Song Scope.