Estimation of vocal tract configurations from magnetic resonance images and synthesizing speech sounds

本项目的主要研究是调查通过（1）磁共振图像（MRI）和（2）真实的语音信号估计的声道配置。特别是在这一阶段，对从真实的语音信号中估计声道结构进行了实验研究，在模拟人类语音产生机制的计算机系统上，采用“综合分析”算法估计声道结构。该模拟器由声源、声道和唇部辐射三部分组成。为了简化损失项插入道模型的过程，将每一部分与一个由频域和时域表示的混合系统相结合。声道模型由20个等长、不同截面积的圆柱形管道组成，利用声道模拟器合成了5个日语元音，其中声道的结构由MRI估计，共振峰模式（频率）由MRI估计。元音/a/和/i/的第一共振峰频率被估计为分别比真实的语音的第一共振峰频率低约120 Hz和70 Hz。除第四共振峰频率/i/外，其他元音的相对误差均在5%以内。对由真实的语音估计的形状合成的声音，实验结果与真实的声音的谱图有较好的近似，其相对误差均在3%以内。然而，/a/、/u/和/e/的第一共振峰频率的误差相对较大，其值在8 - 9%之间。

A main research of this project is investigating vocal tract configurations estimated by (1) magnetic resonance images (MRIs) and by (2) real speech signals. Especially in this term, experiments were conducted to the estimation of the vocal tract configuration from real speech signals.Developing a simulator on a computer system which is analogous to a mechanism of speech production process of human beings, the configuration of the tract was estimated by using "Analysis by Synthesis" algorithm. The simulator consists of three parts ; vocal source, vocal tract and lip radiation. Each part was combined with a hybrid system represented by frequency domain and time domain for simplicity of insertion of loss-term into the tract model. The model of vocal tract consists of 20 cylindrical tubes which are equal in length and different in cross sectional area.Five Japanese vowels were synthesized by using the vocal tract simulator where the configuration of the vocal tract was estimated by the MRIs, and their formant patterns (frequencies) were estimated. First formant frequencies of vowel /a/ and /i/ were estimated lower than those of real speech sound by about 120 Hz and 70 Hz, respectively. Relative errors were shown within 5 % in other vowels except fourth formant frequency of /i/.On sounds synthesized by the shape estimated from real speech sounds, experimental results were shown rather good approximation to spectral patterns of real sounds, and their relative errors were shown within 3 %. However, errors in first formant frequencies of /a/, /u/ and /e/ were relatively large and their values ranged in 8 - 9 %.