基频归一化——如何处理声调的随机差异?

F0 Normalization:How to Deal with Between-speaker Tonal Variations?

一个语言信号的物理性质表现形式是无限多的。归一化的主要目的是消减人际随机差异,提取恒定参数,在语际变异中找到共性,从而使得人际比较和语际比较的研究成为可能。本文比较六种归一化方法,以及各种选点、定频域的方法,引入了判断方法好坏的标准指数和离散系数。结果表明对数 z-score(LZ)方法最好,与此相应的频域定义以均值和标准差为好,用以归一化的采样点以选用语音目标处的时点为佳。

This paper compares six mathematical transforms for F0 normalisation using eleven speakers of Shanghai,a dialect of Wu Chinese.Various strategies for defining range and selecting nor- malisation points are also exploited.The logarithmic z-score transform with normalisation points at F0 targets is found to be superior due to its higher normalisation indices,its advantages in statistics and, perhaps,its accordance with perception and production.