Downstepped Pitch Accent in American English is Categorical and Predictable Tae-Jin Yoon; Jennifer Cole University of Illinois at Urbana-Champaign, tyoon@uiuc.edu; jscole@uiuc.edu Résumé: A travers une analyse acoustique et une expérience didactique sur machine, cette présentation met en évidence une distinction de catégorie en anglais américain entre les tons hauts accentués H* dont le registre est abaissé (downstep: !H*) et ceux dont il ne l'est pas. Cette étude offre une explication quant aux découvertes contradictoires de recherches acoustiques precedents (Liberman et Pierrehumbert 1984; Dainora 2001a,b) sur le downstep qui donnent à réfléchir sur l'hypothèse que le statut de l'accent !H* en anglais américain constituerait une catégorie légitime. 1. Introduction The paper presents evidence from acoustic analysis and a machine learning experiment for a categorical distinction between downstepped and non-downstepped high-toned pitch accents (H* vs. !H*) in American English. The present study offers an explanation for the contradictory findings from prior acoustic studies of downstep (Liberman & Pierrehumbert 1984 vs. Dainora 2001a,b), which call into question the status of the downstepped accent in American English as a legitimate prosodic category. Dainora (2001a,b) suggests that there is a single phonological High tone that can be used in the specification of pitch accent melody, and “downstepped” pitch accents are illusory, being no more than a subset of variants taken from the normal distribution of H* peak values. 2. Is there !H* in American English? Pierrehumbert (1980) adopts the analysis of a downstepped pitch accent as a phonologically derived feature. Later work in intonation building on Pierrehumbert’s model allows the possibility that downstep is non-automatic and encoded as a contrastive tonal category (e.g., Ladd 1983; Beckman & Ayers 1997). Pierrehumbert (2000) notes, however, that despite the experimental findings of Liberman & Pierrehumbert (1984) in support of downstep, the categorical status of !H* is not substantiated by any large-scale study of naturally occurring speech. More recently, the categorical status of !H* is called into question by Dainora (2001a,b). Based on the analysis of Radio News speech (news stores read by 7 professional announcers), Dainora investigates the status of !H* by comparing the pitch drop in the tonal sequences (H* !H*) and (H* H*). She shows that the pitch drop measure defines a unimodal distribution, where H* and !H* belong to opposite ends of a single distribution in the F0 dimension, in contradiction of earlier claim that !H* forms a distinct tonal category. 3. Categorical Status of !H* We argue that Dainora’s study fails to consider the effects of peak height on the pitch measure; specifically, the F0 peak of the first H* in the sequence might condition the magnitude of the pitch drop to a following peak. In an alternative analysis developed here, we analyze the peak of the second pitch accent (both H* and !H*) in relation to the peak of the preceding H* in the target sequences. We apply this analysis to the same set of data from the Boston Radio News corpus used by Dainora, and on data from a Maptask corpus of spontaneous speech (Shattuck- Hufnagel et al. 2004) produced by 1 female speaker. Using regression analysis methods, we show that H* and !H* form two distinct distributions when the F0 peak is plotted against the peak height of a preceding H*. In regression analysis for peaks in the sequence H*H* in the Boston Radio News corpus, the slope and intercept is 1.0 and 15.93, respectively (Y=1.0X + 15.93). For peaks in the sequence H*!H*, the slope and intercept of the regression is 0.5 and 63.95, respectively (Y=0.5X+63.95) (see Figure 1). For peaks in sequences of H*H* in the Maptask corpus, the slope and intercept of linear regression is 0.82 and 36.60, respectively (Y=0.82X+36.60), and for peaks in the H*!H* sequence, the slope and intercept values are 0.66 and 50.56, respectively (Y=0.66X+50.56).