视听言语合成技术综述  

Audio visual speech synthesis: an overview

作　　者：张金光 

机构地区：[1]北京大学中国语言文学系语音乐律实验室,北京100871

出　　处：《电声技术》2017年第7期103-107,共5页Audio Engineering

摘　　要：听觉言语合成经历了共振峰合成、基音同步叠加拼接合成、单元选择拼接合成,正在走向基于统计模型的合成技术。技术原理从预设规则的参数合成,发展到基于原始信号波形的拼接合成,再到基于统计规则的参数合成,经历了一个循环。视觉言语合成从生理解剖模型,到几何特征模型,再到基于大语料库的数据驱动的拼接合成,也在走向基于统计规则的参数合成。各种技术的本质是什么?优缺点有哪些?本文尝试进行视听言语合成技术发展综述。Auditory speech synthesis has gone through three critical stages： formant synthesis, pitch-synchronous overlap and add synthesis, unit-selection synthesis. Now it is walking to statistical model based synthesis. The technological basis has undergone reincarnation, changing from preset parameters, to original waveform concatenation, and back to parameters again, but statistical ones. Visual speech synthesis includes anatomy-based models and geometric property based models. The dominant practice in visual speech synthesis is to use corpus to do data-driven concatenation, but there is a tendency that the statistical method will be widely used in the near future. What are the essences of those technologies？ What are their pros and cons？ This article tries to give a brief overview on the development of audio-visual speech synthesis.

关 键 词：言语合成 规则合成 拼接合成 数据驱动 

分 类 号：TN912.33[电子电信—通信与信息系统]