Predicting Midsagittal Pharynx Shape From Tongue Position During Vowel Production

Abstract

The shape of the pharynx has a large effect on the acoustics of vowels, but direct measurement of this part of the vocal tract is difficult. The present study examines the efficacy of inferring midsagittal pharynx shape from the position of the tongue, which is much more amenable to measurement. Midsagittal magnetic resonance (MR) images were obtained for multiple repetitions of 11 static English vowels spoken by two subjects (one male and one female). From these, midsagittal widths were measured at approximately 3-mm intervals along the entire vocal tract. A regression analysis was then used to assess whether the pharyngeal widths could be predicted from the locations and width measurements for four positions on the tongue, namely, those likely to be the locations of a receiver coil for an electromagnetometer system. Predictability was quite high throughout the vocal tract (multiple r> 0.9), except for the extreme ends (i.e., larynx and lips) and small decreases for the male subject in the uvula region. The residuals from this analysis showed that the accuracy of predictions was generally quite high, with 89.2% of errors being less than 2 mm. The extremes of the vocal tract, where the resolution of the MRI was poorer, accounted for much of the error. For languages like English, which do not use advanced tongue root (ATR) distinctively, the midsagittal pharynx shape of static vowels can be predicted with high accuracy.

Predicting midsagittal pharyngeal dimensions from measures of anterior tongue position in Swedish vowels: Statistical considerations

“…There can be a number of reasons for this, including a small number of observations with respect to number of independent variables (e.g., Whalen et al, 1999), high variability in the data, as is often the case with children (e.g., McGowan et al, 2004), the complexity of algorithms, such as PARAFAC (parallel factors, e.g., Jackson, 1988), or because the data are not normally distributed. In these cases numerical experiments and/or nonparametric statistics provide a means for enabling us to make statistically meaningful statements.…”

“…Each measurement grid was based on four fixed reference points, P1 through P4, similar to the ones used in Whalen et al (1999). However, Whalen et al (1999) analyzed MR images, whereas this study analyzes tracings of x-ray images, leading to differences in choosing anatomical landmarks.…”

“…A point corresponding to P3 was identified on each image for each speaker, and the mean position (within speaker) was used as the fixed location of P3 in subsequent analysis. Following Tiede (1996), who described an earlier version of the image-measurement methods used in Whalen et al (1999), the point on the rear pharyngeal wall at the same height as the point at the bottom of the vallecular sinus was adopted to be P4. Table II reports the standard deviations of the x and y positions of P3 in rows seven and eight; P4 in rows nine and ten.…”

“…For instance, Nguyen et al (1994) showed that coordinates of three points on the tongue predict the acoustic output of the vocal tract in [s] and [∫] in French, and Lindau-Webb and Ladefoged (1989) showed that as few as two fleshpoint positions suffice to determine the entire tongue shape in static American English vowels. Whalen et al, (1999), using the coordinates of four manually estimated fleshpoints from magnetic resonance (MR) images of 11 American English vowels, showed high correlations between these eight variables (x and y coordinates of each fleshpoint) and midsagittal vocal tract cross distances measured along gridlines.…”

“…Thus, the significance of correlations reported in some speech production studies is open to debate due to the small number of observations and the non-normality of the data. For instance, studies such as Jackson (1988) with 16 observations (vowels) per speaker, and Whalen et al (1999) with 11 observations, typically report high correlations, over 0.95. However, studies using more data, such as Beaudoin and McGowan (2000) with over 3000 observations per speaker, report correlations of around 0.9.…”