Click reduction in fluent speech: a semi-automated analysis of Mangetti Dune !Xung

We compare click production in fluent speech to previously analyzed clear productions in the Namibian Kx’a language Mangetti Dune !Xung. Using a rule-based software system, we extract clicks from recorded folktales, with click detection accuracy about 65% f-score for one storyteller, reducing manual annotation time by two thirds; we believe similar methods will be effective for other loud, short consonants like ejectives. We use linear discriminant analysis to show that the four click types of !Xung are harder to differentiate in the folktales than in clear productions, and conduct a feature analysis which suggests that rapid production obscures some acoustic cues to click identity. An analysis of a second storyteller suggests that clicks can also be phonetically reduced due to language attrition. We argue that analysis of fluent speech, especially where it can be semi-automated, is an important addition to analysis of clear productions in understanding the phonology of endangered languages.