Signal Denoising

Vocalisations of marine mammals, fish and other organisms are of interest to many scientists who study animal behavior. However, it can be difficult to make high quality recordings of underwater sound due to unwanted environmental noise. This can be a problem not only for the scientists but also for the animals that have to live in increasingly noisy environments and whose health and well-being might be compromised by not being able to hear useful sound signals.

Contrary to popular opinion, the sea in general is a surprisingly noisy place and coastal zones even more so. Natural noise sources include wave and wind action, bubbles, rainfall and even snowflakes landing on water; noise produced by marine mammals, fish, and particularly Snapping shrimp (Alpheus heterochaelis); and man-made noise such as shipping and small boat noise, oil drilling, mining and general off-shore geological explorations, and sonar systems.

While some of these sources produce sound in limited frequency bands (specific pitches) and are therefore relatively easy to filter out of scientific recordings of marine mammal vocalisations, some sources cover a very wide range of frequencies and are much more difficult to remove. This is a similar problem to trying to listen to someone talking at a very noisy traffic junction or during a rock concert: you can shout to a certain extent but sometimes the problem is impossible to overcome. Snapping shrimp are a particularly important source of this kind of broadband environmental noise. They occur in large colonies in shallow tropical waters, particularly around rocks and man-made structures and the noise made by these small crustaceans each snapping a single claw many times dominates the sound field in tropical coastal regions. The result sounds like an egg frying in oil. Click here to a short audio clip of snapping shrimps.

Unwanted noise is not limited to the study of marine mammal vocalisations and is a common problem in more general acoustic signal processing. MMRL has developed several techniques for removing acoustic noise and separating out different signal components from a composite signal comprising of ‘transient sounds’ such as dolphin echolocation clicks, and ‘tonals’ such as dolphin whistles and Humpback whale song.

Despite a lot of research over many years, there is still no easy way to remove broadband transient noise from broadband transient acoustic signals because they both share superficially common signal features. MMRL is working with the NUS Department of Electrical and Computer Engineering and TMSI’s Acoustic Research Laboratory to address this problem which is fundamentally mathematically based. Some of the methods we are exploring are as follows:-

Imaged-Based Acoustic Signal Denoising
This new method based on a combination of signal and image processing techniques, aims to remove acoustic noise from FM tones, enhance the spectrogrammes and use spectrogramme segmentation to perform reliable tracing of the fundamental frequency variation of a dolphin or dugong whistle (or other FM tones). This package of tools is necessary for accurate classification of whistles produced by marine mammals.

Dolphin whistle corrupted in noise

Traced whistle after de-noising

Single Spectrum Analysis and Higher Order Statistics
This technique is jointly developed by MMRL and the Acoustic Research Laboratory (ARL). SSA is a Singular Value Decomposition (SVD) based procedure which decomposes a time series into a number of times series components. The sum of all these time series components is equal to the original time series. However, the time series components can be grouped together prior to summation. Currently grouping is performed by using statistical properties such as kurtosis and is summed separately to separate tonals, transients and spectrally smooth noise. However, the number of components from the decomposition has to be sufficiently large for adequate grouping, and, as the length of the time series and number of components increase, this translates to higher requirements for computation time and memory. The technique has proved useful when extracting dolphin echolocation clicks from tonal noise such as radio interference, and boat noise.

Time series of a signal corrupted with noise

Time series after applying SSA