拓扑独立分量分析在方言特征提取中的应用研究

Research of Extracting Dialect Features with TICA

将基于Mel频率域能量(简称Mel能量)的拓扑独立分量分析算法(ME-TICA)用于方言特征提取。拓扑独立分量分析(TICA)算法能够实现同组信号之间存在高阶相关性时的盲分离,ME-TICA算法在保留这种性能的基础上,引入Mel能量相关的概念来体现人类听觉特征,实现了从无标注的方言语句集中提取包含基音频率和调长两种指标的方言声调特征。文中设计了完整的特征提取过程。英语和汉语两种方言特征提取和识别的仿真实验验证了ME-TICA算法所提取的方言特征具有精度高和鲁棒性强的优势。

An algorithm of TICA, ME-TICA (TICA based on Mel energy), was proposed to extract dialect features. Topographic independent component analysis (TICA) is a blind separation algorithm, in which there is higher-order dependency among the components that are neighbors to each other and the non-neighbors are independent. ME-TICA introduces Mel energy in it to get the acoustic feature of the speech. The tone feature of pitch frequencies and time length was learned from un-labeled sentences database using ME-TICA,which improve the TICA performance. An integrated procedure was designed to extract the features. The experiment of English and Chinese dialect recognition shows that the features extracted with ME-TICA algorithm have higher accuracy and better robustness than others.