Novel Class Identification: A Human‐in‐the‐loop Approach for Efficient Species Recognition in Bioacoustic Data

Abstract Passive acoustic monitoring (PAM) is widely used to collect large‐scale wildlife data encompassing many species, but efficiently identifying all species—particularly rare and potentially endangered ones—without exhaustive manual examination of the entire dataset remains a major challenge. We formally introduce the machine learning task of novel class identification (NCI), which aims to identify as many classes (e.g. species) as possible while examining as few samples as necessary, and we propose a comprehensive evaluation framework for both overall and class‐presence‐specific performance. NCI uses human‐in‐the‐loop querying strategies based on the active learning paradigm, while its objective of efficiently identifying classes aligns with open‐world learning assumptions and is closely related to the goals of novel class discovery. Drawing on insights from the literature, we identify four promising query method families: embedding‐based methods, self‐supervised learning, similarity learning and audio‐language models. To explore the potential of these directions, initial experiments evaluate 44 query method configurations from the identified query method families across five multi‐label PAM datasets. The results demonstrate that the proposed evaluation framework effectively captures both overall and class‐presence‐specific performance, enabling robust comparison of different methods. These experiments further indicate that audio‐language models achieve the highest performance, providing a foundation for future research. Our findings highlight the potential of deep learning methods for rapid species identification, supporting faster and more informed conservation decisions from large PAM datasets.