基于发音轨迹的可视语音合成  

VISUAL SPEECH SYNTHESIS BASED ON ARTICULATORY TRAJECTORY

作　　者：郑红娜[1,2] 白静[1] 王岚[2] 朱云[2] 

机构地区：[1]太原理工大学信息工程学院山西太原030024 [2]中国科学院深圳先进技术研究院广东深圳518055

出　　处：《计算机应用与软件》2013年第6期253-257,261,共6页Computer Applications and Software

摘　　要：着眼于语音可视化。为呈现真实说话人各发音器官的运动,提出一种改进的CM协同发音模型,并采用该方法合成汉字的发音轨迹,进而以此驱动和控制虚拟3D音视频说话人头模型,直观呈现通常可见及不可见发音器官的发音运动。实验证明:由改进方法获得的合成发音轨迹更逼近真实发音轨迹。同时,为了定量比较舌读和唇读在语音感知和识别中的作用,设计三组感知实验,实验结果显示:叠加了唇读信息的感知识别率较单纯的加噪语音的识别率提高了25.8%;而叠加了舌读信息的识别率较单纯听觉信息的识别率提高了26.7%。因此,当语音恶化时,舌读信息可提供大于唇读信息的补充作用,且舌读信息与唇读信息一样具有识别力。This paper focuses on speech visualisation. A modified CM co-articulation model is presented in order to display the movement of each articulatory organ of real speakers, it is employed to have synthesised the articulatory trajectory of Chinese characters, and is further used to drive and control a virtual 3D audiovisual talking head model, intuitively shows the articulatory motions of the articulatory organs usually visible and invisible. Experiment proves that the synthesis articulatory trajectory obtained from the modified method approaches more the actual articulatory trajectory. Meanwhile, in order to quantitatively compare the roles of tongue reading and lip reading in articulatory perception and recognition, three groups of perception experiments are designed. Experimental results demonstrate that the perception recognition rate with the lip reading information superposed improves 25.8% compared with the recognition rate of pure speech with noise superposed; while the recognition rate with the tongue reading information superposed is higher by 26.7% than that of the sole audio information. Therefore, when the speech degrades, the tongue reading information can play more supplementation role than the lip reading information, and the tongue reading information has the recognition capability the same as lip reading information.

关 键 词：语音可视化 CM协同发音模型 改进的CM协同发音模型 舌读 唇读 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]