Data from: Characterizing vocal repertoires - hard vs. soft classification approaches

To understand the proximate and ultimate causes that shape acoustic communication in animals, objective characterizations of the vocal repertoire of a given species are critical, as they provide the foundation for comparative analyses among individuals, populations and taxa. Progress in this field has been hampered by a lack of standard in methodology, however. One problem is that researchers may settle on different variables to characterize the calls, which may impact on the classification of calls. More important, there is no agreement how to best characterize the overall structure of the repertoire in terms of the amount of gradation within and between call types. Here, we address these challenges by examining 912 calls recorded from wild chacma baboons (Papio ursinus). We extracted 118 acoustic variables from spectrograms, from which we constructed different sets of acoustic features, containing 9, 38, and 118 variables; as well 19 factors derived from principal component analysis. We compared and validated the resulting classifications of k-means and hierarchical clustering. Datasets with a higher number of acoustic features lead to better clustering results than datasets with only a few features. The use of factors in the cluster analysis resulted in an extremely poor resolution of emerging call types. Another important finding is that none of the applied clustering methods gave strong support to a specific cluster solution. Instead, the cluster analysis revealed that within distinct call types, subtypes may exist. Because hard clustering methods are not well suited to capture such gradation within call types, we applied a fuzzy clustering algorithm. We found that this algorithm provides a detailed and quantitative description of the gradation within and between chacma baboon call types. In conclusion, we suggest that fuzzy clustering should be used in future studies to analyze the graded structure of vocal repertoires. Moreover, the use of factor analyses to reduce the number of acoustic variables should be discouraged.

为阐明驱动动物声学通讯的近因与终极因，对特定物种的鸣声库（vocal repertoire）进行客观表征至关重要，因为此类表征可为个体、种群及类群间的比较分析奠定基础。然而，该领域的研究进展长期受制于方法学缺乏统一标准的困境。其一，不同研究者可能会选取不同变量对鸣声进行表征，这会对鸣声分类结果产生影响；更关键的是，学界尚未就如何依据鸣声类型内部及类型间的渐变程度来最优地表征鸣声库的整体结构达成共识。
本研究通过分析912份野生chacma狒狒（Papio ursinus）的鸣声样本，对上述挑战展开研究。我们从声谱图（spectrogram）中提取了118项声学变量，并据此构建了包含9、38及118个变量的多组声学特征集；同时还通过主成分分析（principal component analysis）得到了19个因子。我们对K均值（k-means）聚类与层次聚类（hierarchical clustering）得到的分类结果进行了比较与验证。
声学特征数量更多的数据集，其聚类效果优于特征数量较少的数据集。在聚类分析中使用因子变量时，所得新鸣声类型的分辨率极差。另一项重要发现为：所有采用的聚类方法均未为某一特定聚类解提供强有力的支持。与之相反，聚类分析结果显示，在明确的鸣声类型内部可能存在亚型。由于硬聚类（hard clustering）方法难以捕捉鸣声类型内部的此类渐变特征，我们采用了模糊聚类（fuzzy clustering）算法。结果表明，该算法可对chacma狒狒鸣声类型内部及类型间的渐变特征进行细致且定量的描述。
综上，我们建议未来在分析鸣声库的渐变结构时应采用模糊聚类算法；同时应避免使用因子分析来缩减声学变量的数量。