The Discriminability of Nearly Merged Sounds.

In a near merger, speakers produce two contrasting words differently without being able to reliably discern the contrast in their own speech or in the speech of others. Acoustic measurements typically reveal small differences between the elements of near merged minimal pairs, along several acoustic dimensions. This paper argues that statistical evaluation of the potential distinctiveness of these near merged elements must take simultaneous account of all these dimensions. For that reason, discriminant analysis was used to assess the differences between near merged lil-dl, leI-eli, and IUI-ul! for five Utah speakers. In contrast with independent univariate Analyses of Variance of Fl, F2, fo, and spectral slope, the multivariate discriminant analyses suggest that all three contrasts are preserved by all five speakers. However, homophones like heel and heal were not distinguished by the discriminant analyses. Discriminant analysis is thus a powerful technique for assessing whether a reliable basis exists for the claim that two potentially contrastive items are in fact distinctive.