Classifying birdsong in environments with everchanging sounds
Classifying birdsong in environments with everchanging sounds lead image
Collecting environmental acoustic data over long time periods captures important changes in a soundscape. Such data can provide insights into how factors like climate, habitat destruction, and human noise impact animal communication.
Deep learning is commonly used to automatically classify these large datasets, detecting and identifying classes of animal sounds like bird calls. However, when new or rarely occurring sounds appear in the data set, traditional closed-set machine learning strategies run into problems. If those sounds are not part of the training set, it is not possible for the machine learning architecture to classify them correctly.
Morgan and Braasch explored open set classification strategies that can account for this difficulty. They examined a data set of bird calls and other environmental sounds with distinct and unique time-frequency signatures that make them well-suited for machine learning.
The researchers investigated two types of unknown data. In the first, the unknown data consisted of several unambiguous sound classes that the network had never encountered. In the second, the unknown data hosted many sounds that occurred too infrequently or were too faint to otherwise classify.
While the team began with a closed-set machine learning strategy, they changed the final step of the classification procedure to allow for the possibility of these unknowns. A certainty probability was assigned to each given classification prediction, and if the prediction fell below a certain threshold, the associated sound clip was deemed unknown and flagged for manual examination.
“We found that methods that were successful with one type of unknown were much weaker with the other, and vice versa,” said author Mallory Morgan. “There is no one-size-fits-all approach to open set classification.”
Source: “Open set classification strategies for long-term environmental field recordings for bird species recognition,” by Mallory M. Morgan and Jonas Braasch, The Journal of the Acoustical Society of America (2022). The article can be accessed at https://doi.org/10.1121/10.0011466 .
