Comparison of Vocal Tract Shape Modelling Methods: MRI vs AR

Abstract

Investigation was carried out on the vocal tract structural data obtained with the magnetic resonance imaging and the acoustic reflectometery techniques during the vocalization of vowels. This was carried out as a determination of the merits of the data acquisition techniques in question. This investigation included data for both techniques from 5 speakers. Vocal tract structural data was captured by the magnetic resonance method as a series of sagittal images which is converted into cross-sectional area functions through manipulation of the images in the open source modelling tool of CMGUI [http://www.cmiss.org/cmgui]. The data collected by the acoustic reflectometer is in the form of cross-sectional areas along the vocal tract. Voice recordings for each target vowel were also collected. The cross-sectional area functions obtained by the two methods are processed, and with the application of the lossless tube model and the linear predictive coding method, the spectrum and resonances of a give vocal tract structure was obtained. The length and resonance values of each vowel was recorded and analysed. Vocal tract shapes obtained from the two methods were compared and discussed, and it was found that both methods were successful in capturing the expected vocal tract geometry of the target vowels, though a few compromises to the shapes from the acoustic reflectometery method was observed due to the compromising nature of the measurement mouthpiece. The resonances deduced from the vocal tract shapes were compared to the formants extracted from recorded speech. It was found that the magnetic resonance method yielded a more accurate estimate of the formant values (9 out of 11 monophthong with reasonable estimations) while the acoustic reflectometery method was much less accurate (3 out of 9 monophthongs with reasonable estimation) It was concluded that while the acoustic reflectometery method was able to capture the general shape of the vocal tract, only a number of vowels were able to be modelled accurately enough for the calculated resonances to be comparable to the formants of real speech. However, future research may allow for methods with counter acts the compromising effects of the mouth piece, and yield more promising results.