Deep audio embeddings for vocalisation clustering

This dataset includes embeddings of animal vocalisations from different feature extraction methods (auto-encoder embeddings, predefined acoustic features, whole spectrograms). Each vocalisation has been assigned a label by an expert which describe its unit type (vocalisations of these animals come in discrete classes from a fixed repertoire). With this data, we study the relevance of vocalisation representations to cluster them by type, proposing methods to help in repertoire discovery and animal communication studies in general.

本数据集收录了通过多种特征提取方法生成的动物发声嵌入（embeddings），包含自编码器嵌入（auto-encoder embeddings）、预定义声学特征以及完整语谱图（spectrograms）三种形式。每一段发声样本均经专家标注其单元类型，该类动物的发声源自固定的发声库，且可划分为若干离散类别。依托该数据集，本研究旨在探究发声表征按类型聚类的相关性，并提出可助力发声库发现及通用动物通信研究的相关方法。