B0449
Title: Bayesian machine learning for bird call identification in soundscape analysis: An innovative approach
Authors: Hossein Masoumi Karakani - University of Pretoria (South Africa) [presenting]
Abstract: Bird species worldwide exceed 10,000, and their identification within an area yields valuable insights into habitat characteristics. Given their position in the food chain, birds serve as exceptional indicators of environmental degradation and pollution. Leveraging machine learning (ML) techniques, particularly sound detection and classification, researchers can enhance their capacity to monitor biodiversity trends and status in critical ecosystems, enabling them to better support global conservation efforts. Recent advances in machine listening have improved acoustic data collection. While frequentist statistical inference dominates common ML algorithms, the Bayesian perspective offers significant utility for real-world events, such as bird call identification, as it facilitates the integration of prior assumptions with empirical evidence to update beliefs. Extensive bird call data is leveraged and an innovative Bayesian ML model is presented that incorporates audio preprocessing and attention mechanisms. Various applications of Bayesian approaches are described for soundscape analysis. The model exhibits the potential for further development, allowing for the inclusion of additional confounding factors such as the influence of climate on bird species. Using the Streamlit open-source Python library, a web application will be developed to deploy the model in a production environment, enabling users to access it and make informed decisions.
