Comparison of Several Acoustic Features for the Vowel Sequence Reproduction of a Talking Robot
Comparison of Several Acoustic Features for the Vowel Sequence Reproduction of a Talking Robot
复制标题
说话机器人元音序列再现的几种声学特征的比较
DOI：
10.1109/icma.2016.7558722
复制
发表时间：
2016
期刊：
影响因子：
--
通讯作者：
Vo Nhu Thanh and Hideyuki Sawada
中科院分区：
文献类型：
--
作者：
Yoko MIYOSHI;Junko YOSHIMURA;Miwa YAMAMOTO;福山慧，池内秀隆;Fang Jin and Chikamune Wada;Vo Nhu Thanh and Hideyuki Sawada
本研究比较了几个声学特性，为一个会说话的机器人开发了一个自动元音序列再现系统，这是一个模仿人类发音系统的机械发声系统。基于Matlab的控制系统用于分析记录的声音并驱动说话机器人的发音电机。提出了一种基于短时能量分析的语音提取方法，并将其转换为一个音元序列，用于元音序列的再现。然后，几个音素检测方法，包括直接互相关分析，线性预测编码（LPC）的关联，部分相关（PARCOR）系数分析，共振峰频率比较被应用到每个声音元素，给出正确的命令，说话机器人重复的声音顺序。最后，进行了实验，以比较这些技术和验证的机器人的工作行为。测试结果表明，机器人能够重复一系列的元音由一个人说的成功率超过70%的PARCOR分析技术和共振峰频率比较技术。重复序列的最大精度由共振峰比较方法给出，而直接互相关方法提供的精度最低。
This study compares several acoustic features for developing an automatic vowel sequence reproduction system for a talking robot, which is a mechanical vocalization system modeling the human articulatory system. Matlab-based control system is used to analyze a recorded sound and drives the articulatory motors of the talking robot. A novel method based on short-time energy analysis is used to extract a human speech and translate into a sequence of sound elements for the sequence of vowels reproduction. Then, several phonemes detection methods including the direct cross-correlation analysis, the linear predictive coding (LPC) association, the partial correlation (PARCOR) coefficients analysis, and the formant frequencies comparison are applied to each sound element to give the corrected command for the talking robot to repeat the sound sequentially. Finally, experiments to compare these techniques and verify the working behavior of the robot are performed. The result of the tests indicates that the robot is able to repeat a sequence of vowels spoken by a human with a successful rate of more than 70% for the PARCOR analysis technique and the formant frequencies comparison technique. The greatest accuracy for repeating the sequence is given by the formant comparison method, while the direct cross-correlation method delivers the least accuracy.