Skip to main content
University of Bath

Deep-learning and detection of acoustic trends and transients in marine observatories

We aim to detect underwater acoustic events in a long duration recording to assess the impact of noise from commerce and shipping on animal population.

To analyse long recordings, high-performance computing environment is required to overcome computer memory and speed constraints.
To analyse long recordings, high-performance computing environment is required to overcome computer memory and speed constraints.

The project aims to detect and analyses automatically underwater acoustic events, which can be long-term events (shipping) or short-term events (vocalisations by animals, seismic prospection or small-scale weather observations) in a ten-years recording taken from a set of hydrophones.

Project outline

The detection of acoustics events at different time-scale in a very large dataset raises key research questions:

  • At which scale should we look at the data to capture all the features of different sounds?

Indeed, a short modulation like clicks is better to be studied in a very short time-window whereas "longer" sounds like vessels signature, whale calls and songs have some periodical features that can only be studied in larger time-windows.

  • What are the more relevant information that we can extract from the data?

The metrics used to analyse the signal are based on the software PAMGuide developped at the University of Bath that provides habitat characterization techniques. Detection algorithms have been created to extract “true” acoustic events from the background noise.

  • How can we speed up the analysis of the data?

The duration of the recording gives rise to problems with the computer memory constraints and to solve this problem with greater speed and more memory space, the high-performance computing environment at the University of Bath (Balena) is required to use parallel programming and parallel computers.


The final purpose is to offer new perspectives on optimising acoustic processing of very large datasets and machine learning techniques to detect and classify different sound sources. Processing techniques have been presented at the international conference of the Acoustical Society of America that took place in Victoria (British Columbia) in November 2018.


The detection of sound sources in a long duration recording is important to assess the impact of increased ambient noise due to increased commerce and international shipping on animal population. Do the animals adapt to noisier environments? Do they migrate? Long-duration recording are necessary to answer these questions and ensure the protection of animals since they enhance a trend in animal populations or shipping activities.

This is the PhD project of Amelie Klein, supervised by Dr Philippe Blondel and Dr Kari Heine.