Human Vocal Tract Analysis by in Vivo 3D MRI during Phonation: A Complete System for Imaging, Quantitative Modeling, and Speech Synthesis

  • Authors:

  • Axel Wismueller;Johannes Behrends;Phil Hoole;Gerda L. Leinsinger;Maximilian F. Reiser;Per-Lennart Westesson

  • Affiliations:

  • Department of Imaging Sciences and Department of Biomedical Engineering, University of Rochester, New York, U.S.A. NY 14642-8648;Department of Imaging Sciences and Department of Biomedical Engineering, University of Rochester, New York, U.S.A. NY 14642-8648;Department of Phonetics, University of Munich, Munich, Germany 80799;Department of Radiology, University of Munich, Munich, Germany 80336;Department of Radiology, University of Munich, Munich, Germany 80336;Department of Imaging Sciences and Department of Biomedical Engineering, University of Rochester, New York, U.S.A. NY 14642-8648

  • Venue:
  • MICCAI '08 Proceedings of the 11th International Conference on Medical Image Computing and Computer-Assisted Intervention, Part II
  • Year:
  • 2008

Quantified Score

Hi-index 0.00

Visualization

Abstract

We present a complete system for image-based 3D vocal tract analysis ranging from MR image acquisition during phonation, semi-automatic image processing, quantitative modeling including model-based speech synthesis, to quantitative model evaluation by comparison between recorded and synthesized phoneme sounds. For this purpose, six professionally trained speakers, age 22-34y, were examined using a standardized MRI protocol (1.5 T, T1w FLASH, ST 4mm, 23 slices, acq. time 21s). The volunteers performed a prolonged (茂戮驴21s) emission of sounds of the German phonemic inventory. Simultaneous audio tape recording was obtained to control correct utterance. Scans were made in axial, coronal, and sagittal planes each. Computer-aided quantitative 3D evaluation included (i) automated registration of the phoneme-specific data acquired in different slice orientations, (ii) semi-automated segmentation of oropharyngeal structures, (iii) computation of a curvilinear vocal tract midline in 3D by nonlinear PCA, (iv) computation of cross-sectional areas of the vocal tract perpendicular to this midline. For the vowels /a/,/e/,/i/,/o/,/ø/,/u/,/y/, the extracted area functions were used to synthesize phoneme sounds based on an articulatory-acoustic model. For quantitative analysis, recorded and synthesized phonemes were compared, where area functions extracted from 2D midsagittal slices were used as a reference. All vowels could be identified correctly based on the synthesized phoneme sounds. The comparison between synthesized and recorded vowel phonemes revealed that the quality of phoneme sound synthesis was improved for phonemes /a/, /o/, and /y/, if 3D instead of 2D data were used, as measured by the average relative frequency shift between recorded and synthesized vowel formants (p