Parallel Algorithm for 3-D Vocal Tract Simulator

在三维声道形状的声学分析中，我们利用MRI数据，开发了有限元分析的并行算法。该方法能有效地评价口腔或咽部形状与包含个人信息的声学特征之间的关系，并可进行数值模拟。在常规有限元分析中，形状数据的线性方程组采用系数矩阵的Cholesky分解算法来求解。传统的声道声学分析方法计算量大，不可能得到声道形状随时间变化的声学分析结果。相反，我们的方法引入了一种适合于并行计算的系数矩阵预条件迭代算法，并研究了从50 Hz的低频到8 kHz的高频的声道传递函数的高速计算方法。新算法的开发成功，速度是传统方法的5~10倍。

In the acoustic analysis of 3-D vocal tract shapes, we employed the MRI data, and developed a parallel algorithm of Finite Element Analysis (FEM). The method has availability of evaluation the relation between the shape of the oral cavity or pharynx and acoustic characteristics including personal information, and the numerical simulation could be performed.In the conventional FEM analysis, the linear equations form the shape data is solved using the algorithm of the Cholesky decomposition for the coefficient matrix. The conventional method requires huge computational cost, ant it is impossible to obtain the acoustic analysis for time-varying shape of the vocal tract. Contrary our method introduced an iterative algorithm with pre-conditioning for the coefficient matrix, which is suitable for parallel computation method, and we investigated the high-speed computation method to obtain the vocal tract transfer function from low frequency as 50Hz to high frequency as 8kHz. The development of the new algorithm was succeeded, which is 5 to 10 times faster than the conventional method.