Accurate measurement of vocal tract shapes using magnetic resonance imaging
Accurate measurement of vocal tract shapes using magnetic resonance imaging
复制标题
DOI：
10.1002/ecjc.4430780807
复制
发表时间：
1995-08
期刊：
影响因子：
--
通讯作者：
Chang-Sheng Yang;H. Kasuya;S. Kano;Toshihiko Sato
中科院分区：
文献类型：
--
作者：
Chang-Sheng Yang;H. Kasuya;S. Kano;Toshihiko Sato
本文提出了一种利用磁共振成像(MRI)准确测量声道三维形状的方法，旨在分析儿童、男性和女性发声时的声道形状特征。为了得到声道形状的精确数据，从嘴唇到图谱的冠状MR图像获得了口腔声道的数据。声道从声门到腭部的数据来自两种水平磁共振图像。采用高速摄影，每片成像时间为2.9S。从正中矢状面MR图像中估计声道中心线，并计算声道长度。基于正中矢状面的MR图像，从三种图像中的声道轮廓重建声道的三维形状。为了处理海量数据，提出了一种从MR图像中自动提取声道轮廓的方法。针对MRI无法成像的牙齿形状，提出了一种基于牙印MR图像的牙齿形状校正方法。使用声道模型对测量方法的误差进行了定量评估。用所提出的测量方法估计了日语/a/的传递函数，并将结果与现场记录的语音的LPC分析结果进行了比较。结果是F1、F2和F3的误差分别为8、17和141赫兹。
In this paper, a method is developed for an accurate measurement of the three-dimensional shape of the vocal tract, using magnetic resonance imaging (MRI), aiming at the analysis of the features of the vocal tract shape in the utterance by children, males and females. to derive the precise data for the shape of the vocal tract, the data for the vocal tract of the oral cavity is derived from the coronal MR image from the lips to the atlas. the data for the vocal tract from the glottis to the palate are derived from two kinds of horizontal MR images. High-speed photography is used, and the imaging time is 2.9 s for each slice. the central line of the vocal tract is estimated from the mid-sagittal MR image, and the vocal tract length is calculated. Based on the mid-sagittal MR image, the three-dimensional shape of the vocal tract is reconstructed from the vocal tract contour in the three kinds of images. To process the large amount of data, a method is developed that automatically extracts the vocal tract contour from the MR image. A correction method is proposed for the teeth shape which cannot be imaged by MRI, based on the MR image of the dental impression. the error of the measuring method is evaluated quantitatively using a model for the vocal tract. the transfer function for Japanese /a/ are estimated by the proposed measuring method, and the result is compared to the result of LPC analysis of the speech recorded at the site. It is seen as a result that the errors for F1, F2, and F3 are 8, 17, and 141 Hz, respectively.