摘要
This study aims to address the limitations in few-shot bioacoustic event detection, such as the scarcity of labeled data and the selection of appropriate frequency ranges. Data collection: We collected 17 bioacoustic recordings of varying lengths from a university campus and a popular hiking trail in Hsinchu, Taiwan, with 9 bird species subjectively identified. The total duration is 1 h and 12 min, with a total of 1389 positive events. Methods: Our approach was built on the 2023 Detection and Classification of Acoustic Scenes and Events (DCASE) Task 5 baseline, the Prototypical Network (PN), enhanced with an Automatic Frequency Range Identification (AFRI) method which computes the low- and high-frequency bounds by comparing the Mel-spectral power density difference between positive events and negative events. Results: Using PN alone, we achieved an F-score of 29.1%. Integrating AFRI with PN improved the F-score to 36.9%, demonstrating the effectiveness of the AFRI approach.