There are currently multiple exciting tasks undertaken by the Computer Science sub-team. These include:
ARTwarp
Once extracted and measured using Raven and PAMGuard, whistle contours are categorised into types. This is done using a MATLab-based programme called ARTwarp, which uses an ART2 neural network to categorize tonal animal sounds. These sounds are represented by frequency contours obtained by analysing their audio spectrograms in Raven and PAMGuard.
Figure 3: ARTwarp user interface during categorisation. Current whistle contour being analysed is shown in the top left box. The current categorisation, or neuron, it’s being compared to is highlighted in red.
ARTwarp’s Dynamic Time Warping algorithm allows whistles to be temporally stretched or squished. This allows broader similarities between whistle contours to be captured. Hence, ARTWarp allows us to determine how many types of whistle dolphins produce. It also counts how many whistle types occur. This allows us to compare differences in whistle repertoire between locations.
The interactive user interface displays various details of the real-time processing taking place (Figure 3). Each sound category or neuron is represented by a reference frequency contour, to which new whistles are compared to. If the new whistle matches the reference contour of a particular category according to a set vigilance (or similarity) level, then the new whistle is added to that category. If not, the whistle creates its own new category.
The VIP aims to improve ARTwarp’s user interface and tweak its network parameters. The improvements and analyses conducted will help inform dolphin conservation and management strategies.
Transfer Learning
Transfer Learning involves using machine learning models that have already been trained on previous data. These models are used to classify new dolphin whistle datasets.
On this VIP, a comparison is currently being drawn between using different cloud machine learning providers, e.g. Google Cloud, Microsoft Azure and AWS Rekognition, to distinguish between whistles produced by bottlenose, common, and melon-headed dolphins. This approach may be preferable to an in-house Machine Learning solution for species classification as it can be hosted on fast servers and should require little technical knowledge.
