Acoustic index-based models for determining time of day in long duration environmental audio recordings.

• Machine learning models using acoustic indices can distinguish time of day in audio. • Such models may misclassify samples from dawn and day. • Models may need to be trained (calibrated) for each observation site. • Signal-to-noise ratio, spectral entropy and spectral cover appear most informative. Environmental sounds, such as bird calls, insects, animal and human activities, support monitoring the health of an area being listened to. Historically these observations needed to be made in the field, but in recent decades non-intrusive acoustic recorders can be deployed for long periods instead, reducing time spent in the field and increasing the volume of raw data collected. The volume of data that can be collected makes human-based processing impractical, so automated analysis approaches are required. Based on the observation that different times of the day exhibit characteristically different soundscapes, this paper investigates predictive (i.e., machine learning) models that use acoustic indices (a calculated representation of some aspect of the recording) to learn and later identify the gross time of day (dawn, day, evening or night). The analysis was based on recordings from north-west Tasmania, Australia, captured in Spring 2017, with 1-min segments of audio sampled at regular intervals across each day. No attempt was made to eliminate unwanted noise (such as wind and rain) before the audio was processed. While audio recordings will typically have accompanying time stamps, this study can be used as the basis for future work in environmental acoustics on: the preferred machine learning classifier; portability of models across sites; the quantity of data required for training; and feature selection. [ABSTRACT FROM AUTHOR]