Avian Language Models: Rethinking Bioacoustic Pipelines with Self-Supervised Learning. In Data Science & Engineering Master of Advanced Study (DSE MAS) Capstone Projects

For our capstone project in the MAS Data Science and Engineering program, we focused on automated bird species identification from ambient soundscapes, a task traditionally constrained by the laborious process of manual audio annotation. Tracking bird populations can serve as a leading indicator to problems in the environment, such as pollution or food chain imbalances. As many bird species are small and spend most of their time hidden, counting populations through sound rather than sight is much easier and more accurate. The data used in this project consists of audio files containing bird vocalizations collected using focal and omnidirectional soundscape recording devices. The data is available publicly through the BirdSet and Cornell Birdcall Identification (CBI) datasets. We conducted controlled ablation studies to isolate the effects of LLMs within the NatureLM model, and compared to the same architecture without the LLM component, as well as our own encoder-only model based on CLAP.