Evaluation of segmentation approaches and constriction degree correlates for spirant approximant consonants

The segmentation of Spanish spirant approximant consonants poses several methodological challenges given the gradual transition from these consonants to neighbouring segments. Most studies to date employ manual segmentation approaches, although these methods are highly subjective and time-consuming. Alternative automated approaches have used the minimum intensity in the consonant and maxima from the surrounding segments as landmarks for relative intensity measurements. Multinomial Logistic Regression (MLR) analyses were used to assess whether acoustic measurements obtained from manual and automated segmentation approaches are able to predict the level of a known categorical outcome variable (phoneme category). A third analysis was included to explore the predictive capabilities of data from several constriction degree correlates, based on relative intensity or spectral energy measurements. Results showed a slight advantage for the automated segmentation method. As for constriction degree correlates, some measurements – Intensity Ratio and Intensity Difference A – were unable to fully predict the outcome categorical variable.