Year: 2025
Type: preprint
Abstract: Marine mammal vocalization analysis depends on interpreting bioacoustic spectrograms. Vision Language Models (VLMs) are not trained on these domain-specific visualizations. We investigate whether VLMs can extract meaningful patterns from spectrograms visually. Our framework integrates VLM interpretation with LLM-based validation to build domain knowledge. This enables adaptation to acoustic data without manual annotation or model retraining.
Source: arXiv (Cornell University)
Authors: Nihal, Ragib Amin, Yen, Benjamin, Ashizawa, Takeshi, Nakadai, Kazuhiro
Language: English
Subfield: Ecology
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4414764109
Cites:
Cited by:
Related to: