On the error distributions of speech production data

In previous works [J. Acoust. Soc. Am. 102, 3164–3165(A) (1997) and 105, 2(A), 1354(A) (1999)] the following problem was presented. Assuming that two‐dimensional data (such as that obtained in most kinematic speech studies) are approximately linear with error terms about the mean coming from a quadratic—rather than a normal—distribution, then the regression analysis is best performed using a maximum likelihood stochastic linear regression methodology, instead of a least squares methodology. It was shown that one could expect, among other results, sharper confidence intervals about the mean. It remains to be determined if the distribution of error terms about the mean (for speech production data) is better described by a quadratic than by a normal distribution. Indeed, using the classical Kolmogorov–Smirnov test, it is shown that in many cases one can assume that the error terms follow a quadratic distribution and that at any reasonable confidence level one can reject the hypothesis that error terms follow a normal distribution. The result of this test substantiates the theoretical premise assumed in the earlier work.