Catching Toad Calls in the Cloud: Commodity Edge Computing for Flexible Analysis of Big Sound Data

Abstract

Passive acoustic recording has great potential for monitoring both endangered and pest species. However, the automatic analysis of natural sound recordings is challenging due to geographic variation in background sounds in habitats and species calls. We have designed and deployed an acoustic sensor network constituting an early warning system for a vocal invasive species, in particular cane toads. The challenging nature of recognising toad calls and the big data arising from sound recording gave rise to a novel edge computing system which permits both effective monitoring and flexible experimentation. This is achieved through a multi-stage analysis system in which calls are detected and progressively filtered, to both reduce data communication needs and to improve detection accuracy. The filtering occurs across different stages of the cloud system. This permits flexible experimentation, for example when a new call or false positive is received. Furthermore, to balance the loss of data from aggressive filtering (call recognition), novel overview techniques are employed to provide data summaries. In this way an end user can receive alerts that a toad call is present, the system can be tuned on the fly, and the user can view summary data to have confidence that the system is functioning correctly. The system has been deployed and is in day-to-day use. The novel approaches taken are applicable to other edge computing systems, which analyse large data streams looking for infrequent events and the system has application for monitoring other vocal species.