Speaker
Description
Marine soundscapes contain valuable information about local ecosystems, making comprehensive analysis of their constituent events essential for accurately assessing ecosystem health. However, they are subject to high levels of continuous noise, hindering our ability to disentangle these events and analyse the soundscape. One potential approach to mitigating noise is the use of a noise-reduction front end. However, because noise-reduction algorithms alter the input distribution seen by the pre-trained embedding networks used in event detection, they must be applied carefully. At the same time, appropriate tools for synthesising marine acoustic recordings at controlled signal-to-noise ratios (SNRs) are currently not available. In this work, we introduce an open toolbox for synthesising controlled recordings aimed at evaluating noise-reduction front ends. Users can select the number of events, the SNR, and the length of the recorded files to simulate underwater recordings containing acoustic events originating from biological, anthropological, or geological processes from the Marine Soundlib. We introduce binary time-frequency masks indicating where event energy is located. These masks are used in the mixing process and in evaluating noise reduction, allowing us to assess performance in the time-frequency bins where event energy is active. We then mix the extracted events with noise recordings from different locations at different SNRs to assess noise-reduction effectiveness. Using this toolbox, we show the relation between event detectability and SNR, and how different pre-trained embedding networks react to noise reduction using a simple Wiener filter; some models show improved event-detection performance, whereas others degrade. This illustrates that the effects of noise-reduction methods should be studied carefully within a detection pipeline.