Nuclear Tunes lost and found: Modeling intonational tunes in American English with labeled vs. unlabeled data

We examine how intonational tunes in American English are represented by speakers, as assessed in an imitative speech paradigm, in which speakers reproduce tunes from model utterances. We test 8 distinct nuclear tunes defined in the inventory of American English intonational phonology. We present “bottom up” cluster analyses of unlabeled  data, and “top down” analyses of data which is labeled with the 8 tune shapes that are predicted to be distinct. We find that some predicted distinctions are lost in unlabeled clustering analysis, though they are detectable in small variations in f0 scaling and alignment in labeled data, and further reflected in neural net classification of labeled data. Lost distinctions in production are also reflected in poorer discrimination in an AX perceptual discrimination task. Results together suggest a hierarchy of distinctions among tune shapes, not directly predictable from the tonal inventory. We discuss implications for discrete intonational categories and continuous variation in intonational phonology.