Automatic environmental soundscape classification of continuous field recordings around Lake George, NY

Abstract

The amount of audio data required for long-term bioacoustics monitoring is often too large to be manually sorted. Automatic environmental sound recognition techniques are therefore applied to extract relevant acoustic stimuli and classify these stimuli after a training period. In this work, a continuous 24-h, remote audiovisual recording system was developed and deployed near a tributary stream of Lake George, NY for the automatic collection of environmental and animal sounds. Besides monitoring natural environments, this system can also be used to establish an automatic protocol for collaborative business meetings. In conjunction with the efforts of the RPI/IBM Jefferson Project to create a network of sensors continuously monitoring the lake, the goal of this research is to automatically transcribe the soundscape of the lake using a network of directional microphone arrays positioned throughout the watershed. After establishing a training database of relevant acoustic stimuli, a convolutional neural network is used to classify unprocessed audio data. Pilot studies show good results when learning takes place directly from spectrograms, since the variability of non-speech acoustic events often requires more rich detail than can be provided by MFCCs or other compressed feature sets. [Work supported by NSF #1631674, CISL, and the Jefferson Project.]