Google's Advanced AI Model for Identifying Whale Sounds and Its Applications in Marine Research

TapTechNews September 27th news, according to Google's press release, Google is continuously investing resources to develop an AI model that can be used to identify the sounds of whales. Its latest multi-species whale sound model can identify the calls of 8 types of whales and further subdivide the different vocalization types of two of these whales.

Google said that this technology is mainly used for marine academic research and can help research teams obtain the behavioral patterns of whale groups such as the Bryde's whale that are'more difficult to observe'. Currently, the model has been publicly available for academic personnel to download and use. TapTechNews attaches the address (click here to visit).

It is known that Google has been collaborating with the Pacific Islands Fisheries Science Center of the National Oceanic and Atmospheric Administration (NOAA) in the United States since 2018 to conduct research on the classification of whale calls. Initially, they developed a detection model for humpback whales and successfully identified the chirping sounds of humpback whales, thereby successfully obtaining the chirping patterns of the relevant whale groups at specific times and places, and also discovered new habitats for the whale groups.

Google has also collaborated with the Department of Fisheries and Oceans (DFO) of Canada and RainforestConnection to develop a series of detection models for the endangered southern resident killer whales and applied them to the underwater microphone monitoring network, which can report the position of the killer whales to the DFO and related units in real time.

Google mentioned that now the company's whale sound recognition model can identify humpback whales, killer whales, blue whales, fin whales, minke whales, Bryde's whales, North Atlantic right whales and North Pacific right whales. Considering that the frequency range covered by whale sounds is extremely wide (such as the 10 Hz low frequency of blue whale vocalization and the 120 kHz high frequency of toothed whale vocalization), so this model is a pioneering work at the scientific research level.

TapTechNews learned that when developing this whale sound recognition model, the researchers first converted the original audio into a spectrogram as the basis for the model to recognize sounds, where each time window contains a 5-second sound clip. Then, the model uses the Mel scale method to adjust the sound frequencies of these spectrograms and highlights the sound features through compression and regularization, and then classifies them into the chirping categories of different whales.

Due to various sounds in the sea, the researchers also added a large number of background sounds to train the model to ensure accurate recognition of whale sounds and exclusion of noise. The experimental results show that this whale sound recognition model can well distinguish different types of whales, especially the sounds of minke whales and Bryde's whales.