All day every day, a computer on each auto-detection buoy analyzes sounds as they arrive. Whales have deep voices, so the software focuses on sounds between 50 and 350 Hertz--roughly spanning the notes on the left-hand half of a piano keyboard. The first step is to separate discrete sounds from background noise, a computing task that’s similar to recognizing an object in a cluttered photograph.
Next, the software estimates each sound's similarity to a right whale up-call. It assesses a dozen characteristics such as starting, minimum, and maximum frequencies, and the sound’s duration. Right whale up-calls are typically about 1 second long and rarely longer than 2 seconds. The software assigns each 2-second segment of sound a score between 1 and 10, with higher numbers indicating the clip sounds more like a right whale. Clips that score less than 6 almost never come from right whales.
The unit keeps a running tally of the 10 highest-scoring sound clips. Every 20 minutes the buoy makes a cell or satellite phone call to a server at the Cornell Lab of Ornithology, where it uploads its most recent top-ten list, then clears its memory and starts building a new list.
Step 3 > Analysts at Cornell verify the detections