Modeling Sound Changes in Mandarin Spontaneous Speech Using Deleted Interpolation of Mixture Component Weights
会议名称：《第九届全国人机语音通讯学术会议》
会议日期：2007年
学科分类：1305[艺术学-设计学（可授艺术学、工学学位）] 13[艺术学] 081104[工学-模式识别与智能系统] 08[工学] 0804[工学-仪器科学与技术] 081101[工学-控制理论与控制工程] 0811[工学-控制科学与工程]
基 金：The work is partially supported by Toshiba and Tsinghua University Joint Project
关 键 词：Sound Changes Deleted Interpolation Hand-labeled
摘 要：The high error rate of recognition accuracy in spontaneous speech is due in part to the poor modeling of pronun-ciations variations. An analysis of the acoustic data reveals that the variations include both phone changes and sound changes. Sound changes are the variations within the phoneme, such as nasalization, centralization, voiceless, voiced, etc. Sound changes are flexible and include diacritics that have to be explicitly hand-labeled by linguists. Annotating such corpus is time consuming and the available hand-labeled samples of sound changes are very limited. In this paper, based on standard phonetic unit inventory, we use dynamic programming alignment together with data-driven method to extend the phone set automatically for sound change description. We propose using deleted interpolation to interpolate baseline models and the more refined, but less well-trained sound change models, with the goal of improving the robust ability of sound change models to cover the diversity of sound variations in spontaneous speech. The effectiveness of this approach is evaluated on the 1997 Hub4NE Mandarin Broadcast News Corpus （1997 MBN） with different styles of speech. It gives a significant 1.98% absolute syllable error rate reduction in spontaneous speech. Additional 1.04% absolute syllable error rate reduction is achieved compared to Gaussian mixture sharing method.