Abstract

Large-scale manual labeling process is error-prone and time-consuming, whereby it is not a work to be done by one person. The present paper describes the development of a collaborative labeling protocol with the objective of generating manual annotations of polyphonic audios for the task of sound event detection. The audio recordings correspond to 1039 files each of 5 minutes of recording, which were collected over a period of one year (2019-2020) in three urban wetlands in the city of Valdivia, Chile. In this context of collaborative work, it is necessary to create a set of definitions that allow understanding and sharing concepts. Moreover, we propose a taxonomy that fits the audio data. Specifically, to carry out the labeling work, the open-source web tool called Label-Studio will be used. To understand the dataset based on the taxonomy proposed, it is expected to study the label distributions according to their: a) temporal information, such as hour, day, season; b) spatial, for example, geographic location and urbanization gradient; and c) origin of the sound source, in particular, biophone, anthropophone and geophone. The experience described here provides a contribution to collaborative research in the task of sound event detection. Also, it is expected to contribute to highlighting the importance of the urban wetland soundscapes for collective well-being.