A Comparison of Tone Normalization Methods for Language Variation Research

A Comparison of Tone Normalization Methods for Language Variation Research Jingwei Zhang Department of Chinese Language and Literature, University of Macau Avenida da Universidade, Taipa, Macau, China jwzhang@umac.mo Abstract One methodological issue in tonal acoustic analyses is revisited and resolved in this study. Previous tone normalization methods mainly served for categorizing tones but did not aim to preserve sociolinguistic variation. This study, from the perspective of variationist studies, re- evaluates the effectiveness of sixteen tone normalization methods and finds that the best tone normalization method is a semitone transformation relative to each speaker’s average pitch in hertz. 1 Introduction In comparison with the large amount of work done on vowel normalization (see van der Harst, 2011 p. 90-107 for an overview), tone normalization, or more specifically, the normalization of the fundamental frequency associated with linguistic tone, has received relatively little attention (Rose, 1987). The goal of tone normalization is to eliminate anatomical variation between speakers and allow between-speaker comparison. The anatomical variation resides in the frequency variation mainly caused by different sizes of the vocal folds. Although the variation of vocal fold sizes is the source of the variation of fundamental frequencies, listeners are nevertheless able to neutralize such anatomy-based acoustic differences while retaining phonemic differences (ibid.). Hence, to accurately model the perception of tones by the human auditory system, these anatomy-based differences need to be removed by a proper normalization method. In variationist studies, an effective technique for the normalization of vowels is one that (1) preserves phonemic variation, (2) minimizes anatomical variation, and (3) preserves sociolinguistic variation (Adank, 2003; Adank et al., 2004; van der Harst, 2011). These criteria can be applied to tone normalization as well. Previous vowel normalization studies suggested that different normalization methods serve different goals (Disner, 1980; Thomas, 2002; Thomas, 2011). Section 2 reviews the aforementioned criteria for evaluating normalization methods and introduces a comparison method used in vowel normalization that has been adapted for tone. In Section 3, 16 normalization methods, ten existing ones and six variants, are briefly described and compared. Section 4 presents the comparison results among 16 methods. Section 5 concludes with a discussion of which normalization method is the best for tonal variationist research. 2 Literature Review 2.1 Previous evaluations of tone normalization methods Rose (1987) provided the first quantitative comparison of tone normalization methods by evaluating both Z-score and Fraction of Range (FOR) by the method Dispersion Coefficient (DC, equation 1). The DC is the ratio of mean between- speaker variance to overall sample variance and is a measure of the degree to which speakers’ values cluster. Rose (1987) found that Z-score is preferable since it has a smaller DC, which indicates a better convergence result. (1) �� = ���� ������������ ��� �������������������������� �������� ������ �������� PACLIC 32 823 32nd Pacific Asia Conference on Language, Information and Computation Hong Kong, 1-3 December 2018 Copyright 2018 by the author Zhu (1999) followed Rose’s study and further developed four normalization methods by Z-score and FOR: proportion of range (POR), ratio …

Content truncated to 500 words