We extracted around two million vowel tokens from a sample of sixty-four speakers (b. 1886–1965; 35M/29F; 16 African Americans/48 non-African Americans) across eight states in the American South in an NSF-funded project. We have validated automatic measurements with manual inspection of alignment samples and find that 87 percent of alignments are successful and another 6 percent are partially successful. This large body of tokens (big data) complements existing sociophonetic research by providing a more thorough, detailed picture of the phonetics of American English. We find that (1) there is a much wider range of realization for vowels than is typically represented, and (2) there is no central tendency for any vowel. Using spatial methods drawn from technical geography, we find that all distributions of tokens in vowel space are nonlinear. This suggests that traditional reliance on finding average acoustic properties of a vowel may be unrepresentative of what most speakers actually do. (3) Distributional patterns for vowels are fractal. When we break up the overall dataset into subgroups (e.g., male/female), the same nonlinear distributional pattern appears but with varying locations of highest density of tokens. These findings complement existing sociophonetic research and demonstrate methods by which variation can both be represented and analyzed.