Using AI to recognise dolphins by sound
Dolphins are arguably some of the most intelligent animals on Earth, and now scientists are using artificial intelligence to learn more about dolphin language...
Dolphins are experts in communication, and produce a variety of sounds for different purposes. They talk to each other using a language of “whistles” and, in some dolphin species, individuals even have their own signature whistles – a bit like having their own names. Dolphins also produce high-frequency echolocation “clicks”, which help them to locate and track their prey – just like bats do.
Because these clicks travel such long distances, they can be extremely useful for studying dolphins. Rather than having to spend long periods of time looking for the animals, scientists can leave acoustic sensors on the sea floor. These record clicks from a large area over several months. However, sorting through all of these sounds and working out which species each click belongs to is difficult. Now, writing in PLOS Computational Biology, Dr Kaitlin Frasier, from the Scripps Institution of Oceanography, has designed a new way to analyse these sounds.
“I spent a lot of time staring at computer screens thinking ‘I think this type of click is different from that one’, and then I realised that it would be better to use a computer to do this consistently,” explains Frasier. “So now we’re trying to use ‘unsupervised learning’ to help us understand our data better.”
Unsupervised learning is a kind of artificial intelligence (AI) in which the computer learns to recognise patterns and sort them into groups. Other pattern-recognition methods usually require the computer to be told in advance which patterns belong to which group, but unsupervised learning means that the computer works it out for itself.
By applying unsupervised learning to the recordings of dolphin clicks, the computer can separate clicks into different types, based on things like how high or low the pitch of the click is, and the rate of clicking. These types match up with what we know – the computer recognises all the clicks from one species of dolphin as being from one species. However, for most species of dolphin, we don’t know which clicks belong to them.
Now that the different click types have been identified, Frasier and her team hope to assign each click type to each species. They will then be able to identify species in their recordings, and this will help them understand how dolphin populations change over time - and how they are affected by problems like oil spills and over-fishing.
“By recording these clicks, you can do calculations to estimate how many animals are swimming through the area over time and look at how populations are changing. What this research is trying to do is take it to the next level: not just how many are there, but who? What animal, what species, what genus? We’re trying to answer those sorts of questions so that we can start to get a more detailed picture.”