The study of the static and dynamic aspects of speech production can profit from technologies such as electromagnetic midsagittal articulography (EMA) and real-time magnetic resonance (RTMRI). These can improve our knowledge on which articulators and gestures are involved in producing specific sounds and foster improved speech production models, paramount to advance, e.g., articulatory speech synthesis. Previous work, by the authors, has shown that critical articulator identification could be performed from RTMRI data of the vocal tract, with encouraging results, by extending the applicability of an unsupervised statistical identification method previously proposed for EMA data. Nevertheless, the slower time resolution of the considered RT-MRI corpus (14 Hz), when compared to EMA, potentially influencing the ability to select the most suitable representative configuration for each phone — paramount for strongly dynamic phones, e.g., nasal vowels —, and the lack of a richer set of contexts — relevant for observing coarticulation effects —, were identified as limitations. This article addresses these limitations by exploring critical articulator identification from a faster RTMRI corpus (50 Hz), for European Portuguese, providing a richer set of contexts, and testing how fusing the articulatory data of two speakers might influence critical articulator determination.
Cite as: Silva, S., Teixeira, A., Cunha, C., Almeida, N., Joseph, A.A., Frahm, J. (2019) Exploring Critical Articulator Identification from 50Hz RT-MRI Data of the Vocal Tract. Proc. Interspeech 2019, 874-878, doi: 10.21437/Interspeech.2019-2897
@inproceedings{silva19_interspeech,
author={Samuel Silva and António Teixeira and Conceição Cunha and Nuno Almeida and Arun A. Joseph and Jens Frahm},
title={{Exploring Critical Articulator Identification from 50Hz RT-MRI Data of the Vocal Tract}},
year=2019,
booktitle={Proc. Interspeech 2019},
pages={874--878},
doi={10.21437/Interspeech.2019-2897}
}