http://swrc.ontoware.org/ontology#TechnicalReport
Estimation of Three-Dimensional Tongue Shape from Midsagittal Tongue Contour using Regression Models
en
ポスターセッション
Faculty of Intelligence and Informatics, Konan University
Fourth Department of Forensic Science, National Research Institute of Police Science
Department of Electrical and Electronic Engineering, Tohoku Institute of Technology
Tatsuya Kitamura
Hisanori Makinae
Masashi Ito
In this study, we investigated methods to estimate the tongue contours of the outer sagittal planes from a midsagittal tongue contour using linear and machine-learning-based regression models. To validate the method, tongue contours were extracted from each frame of a 3D MRI movie of the three sagittal planes, including the midsagittal plane, recorded while a male adult speaker produced the Japanese vowel sequence / aiueo /. The extracted tongue contoius were then converted into harmonic amplitude profiles (HAPs) using a 2D Fourier transformation to reduce the number of dimensions. Finally, we calculated the coefficients of a multiple regression model and trained a random forest regression model to map the HAP of the midsagittal plane to those of the outer sagittal planes. In comparison with the multiple regression model, the random forest regression model demonstrated a higher estimation accuracy and exhibited an average distance error of less than 2.0 mm in a 10-fold cross-vafidation test.
In this study, we investigated methods to estimate the tongue contours of the outer sagittal planes from a midsagittal tongue contour using linear and machine-learning-based regression models. To validate the method, tongue contours were extracted from each frame of a 3D MRI movie of the three sagittal planes, including the midsagittal plane, recorded while a male adult speaker produced the Japanese vowel sequence / aiueo /. The extracted tongue contoius were then converted into harmonic amplitude profiles (HAPs) using a 2D Fourier transformation to reduce the number of dimensions. Finally, we calculated the coefficients of a multiple regression model and trained a random forest regression model to map the HAP of the midsagittal plane to those of the outer sagittal planes. In comparison with the multiple regression model, the random forest regression model demonstrated a higher estimation accuracy and exhibited an average distance error of less than 2.0 mm in a 10-fold cross-vafidation test.
AN10442647
研究報告音声言語情報処理（SLP）
2019-SLP-130
14
1-6
2019-11-29
2188-8663