The study of phonetic variation and change has tended to concentrate on particular variables in isolation, and it has proven challenging to move beyond an analysis of individual variables or small groups of variables, towards a better theoretical and empirical understanding of entire vowel systems. We develop a methodology that facilitates the study of co-variation, and introduce a large scale analysis of how elements of full sound systems co-vary across hundreds of speakers, demonstrating how constellations of vocalic variables operate together. Our data-set comprises F1 and F2 for 10 monophthongs of New Zealand English. We first obtain estimates of how advanced each speaker is with respect to changes in each of the vowels, irrespective of known predictors of sound change (i.e. year of birth, gender, speech rate). This is done by extracting by-speaker intercepts from Generalised Additive Models. We then use Principal Component Analysis on these intercepts to investigate the underlying structural co-variation that exists across the vocalic variables. Within a large subset of vowels, we see ‘leaders’ and ‘laggers’ of sound change; however, there are also groups of vowels which stand in opposition to each other, such that if a speaker is innovative in one, they tend to be conservative in the other. Some sets of covarying vowels could be linked by structural relationships (such as chain-shifting), but there are also covarying sets of vowels with no clear structural relationship, and which may be linked by shared social meaning. Our analysis provides novel insights into the structure of sound systems, demonstrating the existence of structured patterns in the realisations of specific vocalic variables across a large group of speakers. This approach offer a means to overcome long-standing methodological challenges in the study of phonetic co-variation, paving the way for research to move beyond the analysis of individual variables, towards an understanding of variation and co-variation in sound systems.

语音变化和变化的研究倾向于孤立地集中在特定变量上，并且已经证明超越对单个变量或小组变量的分析，对整个元音系统进行更好的理论和经验理解是具有挑战性的. 我们开发了一种促进协变研究的方法，并介绍了对完整声音系统的元素如何在数百个扬声器之间协变的大规模分析，展示了声乐变量的星座如何共同运作。我们的数据集包括新西兰英语的 10 个单元音的 F1 和 F2。我们首先获得每个说话者在每个元音变化方面的先进程度的估计，而不管声音变化的已知预测因素（即出生年份、性别、语速）。这是通过从广义加性模型中提取旁白截取来完成的。然后，我们对这些截距使用主成分分析来研究存在于声音变量中的潜在结构协变。在很大一部分元音中，我们看到声音变化的“领先者”和“落后者”；然而，也有成组的元音相互对立，如果说话者在一个方面有创新，他们在另一个方面往往比较保守。一些共变元音组可以通过结构关系（如链移）连接，但也有没有明确结构关系的共变元音组，它们可能通过共享的社会意义联系起来。我们的分析提供了对声音系统结构的新颖见解，证明了在一大群扬声器中实现特定声音变量时存在结构化模式。这种方法提供了一种方法来克服语音协变研究中长期存在的方法论挑战，为超越分析单个变量的研究铺平了道路，