Speakers of American English use a number of pronunciations for /r/, which are often classified into """"""""bunched"""""""" and """"""""retroflex"""""""" categories. These /r/s have similar acoustic profiles, with a third format (F3) at 2000 or below, in spite of very different vocal tract configurations. Because these articulatory differences may lead to different coarticulatory constraints for the variants, understanding the coarticulation of /r/ with adjacent segments requires knowledge of which variant was uttered. Likewise, knowledge of which variant a speaker uses may aid therapy. At present, an acoustic method for identifying which variant was uttered does not exist, and only acoustic mechanisms for retroflex /r/ have been described. The goal of this research is to look for acoustic cues that distinguish between /r/ variants. As part of this goal, we will formulate and refine acoustic models of both /r/ articulations. We take the view that an in-depth look will reveal acoustic details reflecting variants' substantial articulatory differences. In this project, same-speaker data on tongue, jaw, and lip movement, and vocal tract shape, will be collected concurrently with acoustic recordings for different bunched and retroflex variants of /r/, using an Electro-Magnetic Midsagittal Articulometer (EMMA) system, and Magnetic Resonance Imaging (MRI). The type of /r/ produced will be determined by whether the tongue tip is higher than the tongue dorsum (suggesting a retroflex articulation), or whether the tongue dorsum is higher than the tongue tip (suggesting a bunched articulation) at the lowest point of F3 for /r/. Additional EMMA measures will include (I) tongue tip vs. tongue dorsum movement over time, and (b) lip movement for rounding. Acoustic measures will include spectral peaks, bandwidths, and magnitudes during /r/ for Fl through F5. MRI measures will be vocal tract cavity volumes and constriction locations during sustained production of both variants. A screening test will be used to select-subjects who use both variants of /r/. These data will be used to (1) characterize subjects' patterns of/r/ use in different phonetic conditions, (2) construct acoustic models corresponding to bunched and retroflex /r/ articulation, (3) test candidates for acoustic cues to distinguish the variants. Our work will be guided by the notion that F3 lowering in /r/ is best approximated by a model that includes an extra vocal tract cavity contributing a additional resonance/ antiresonance pair. From pilot work, we expect to find a splitting of F3 into two peaks for both /r/ types, with the lower normally identified as F3 and the upper weakened by an anti resonance produced by the extra cavity. We expect other candidate differences to emerge during analysis-by- synthesis between actual acoustic output for /r/ variants vs. prediction of the model. In this process, we expect to refine our models of the two /r/s, and identify cues to production of the variants. Since production and perception of /r/ is problematic for many disordered populations, acoustic methods for characterizing individual patterns of /r/ use and /r/ coarticulation may aid in designing better therapy and in the design and interpretation of studies of disordered speech.
