ICPhS XVII Special Session Hong Kong, 17-21 August 2011 108 VOICELESS INTERVALS AND PERCEPTUAL COMPLETION IN F 0 CONTOURS: EVIDENCE FROM SCALING PERCEPTION IN AMERICAN ENGLISH Jonathan Barnes a , Alejna Brugos a , Nanette Veilleux b & Stefanie Shattuck-Hufnagel c a Boston University, USA; b Simmons College, USA; c MIT, USA jabarnes@bu.edu; abrugos@bu.edu; veilleux@simmons.edu; sshuf@mit.edu ABSTRACT Intonation models describing F0 alignment and scaling in terms of peak and valley localization can face challenges when F0 contours are interrupted (e.g., during voiceless segments). It is often assumed that some form of perceptual completion or “filling in” of such intervals occurs that resolves these issues. This study uses the perceived scaling of High pitch accents both with and without missing peaks due to F0 gaps to adjudicate between three possible accounts of how speakers treat missing F0 in intonation perception. Results provide strong evidence against both extrapolation and interpolation across the missing region, supporting instead the hypothesis that listeners simply ignore these regions. This suggests that a non-turning-point-based model, such as TCoG, should be considered as an alternative to standard target-and-interpolation models. Keywords: F0 alignment, intonation, interpolation, extrapolation, F0 plateau 1. INTRODUCTION It is often observed in the literature on intonation that, although F0 contours are routinely shot through with discontinuities (e.g., at voiceless intervals), our experience of the signal is one of a continuous melody carried over the length of the utterance (e.g., [10, 11, 17]). This might suggest that gaps in the F0 record are sufficiently unobtrusive, in duration, positioning, or both, to avoid disruption of perceived continuity of the F0 signal. A stronger interpretation, apparently informing mainstream assumptions in speech intonation research, is that the “missing” F0 in voiceless intervals is actually restored perceptually. Nooteboom [17], for example, uses a term from the perceptual completion literature, “filling in”, to characterize this process (p. 644). Furthermore, nearly all approaches to F0 contour modelling (with notable exceptions, e.g., [1, 13]) employ some form of F0 interpolation through gaps in the F0 record. In some cases (e.g., [8, 19]), this move may be purely pragmatic. In other cases, though, perceptual issues are clearly at stake. The MOMEL algorithm [12], for example, using quadratic spline fitting to produce a continuous F0 curve, which is then reduced to “a series of target points” that can serve as an “appropriate phonetic representation” of the contour in question. Crucially, these target points will often fall within the missing F0 intervals. If perceptual “filling in” of missing F0 is real, this is good news for target-and-interpolation approaches to tonal phonetics/phonology, such as the Autosegmental-Metrical model [16, 18]. In this model, F0 turning points (such as those provided by MOMEL, hereafter TPs) are typically seen as critical cues for phonological tone specifications. The absence of key TPs should thus cause serious problems for tonal perception. Perceptual completion, on the other hand, predicts such problems should not arise. To see how perceptual completion might be accomplished in tone perception, take, for example, a symmetrical F0 rise and fall separated by a voiceless interval. One solution would be for listeners to extrapolate the missing F0 peak based on observed trajectories to either side of the gap. (Fig. 1a.) Dannenbring [6], however, provides evidence against this option: when presented with tone glides of this shape, separated by noise, listeners failed to extrapolate a peak between rise and fall, reporting instead a peak F0 equal to or somewhat lower than the real F0 maximum. Figure 1: Schematic showing 3 predictions for the perceptual contribution of a no-F0 region to the scaling of a high accentual contour: a) extrapolation, b) interpolation, and c) “ignoring an absence”.