Tonal coarticulation revisited: functional covariance analysis to investigate the planning of co-articulated tones by Standard Chinese speakers

TL;DR

This paper introduces a novel functional data analysis method using covariance analysis and generalized additive models to study tonal coarticulation in Standard Chinese, revealing detailed covariance structures.

Contribution

It presents a new two-step approach combining covariance functions and GAM residual analysis to better understand tonal coarticulation effects.

Findings

Covariance functions are key to capturing tonal coarticulation effects.

The method uncovers residual structures beyond GAM explanations.

Application to Chinese tone data demonstrates the approach's effectiveness.

Abstract

We aim to explain whether a stress memory task has a significant impact on tonal coarticulation. We contribute a novel approach to analyse tonal coarticulation in phonetics, where several f0 contours are compared with respect to their vibrations at higher resolution, something that in statistical terms is called variation of the second order. We identify speech recording frequency curves as functional observations and harness inspiration from the mathematical fields of functional data analysis and optimal transport. By leveraging results from these two disciplines, we make one key observation:we identify the time and frequency covariance functions as crucial features for capturing the finer effects of tonal coarticulation. This observation leads us to propose a 2 steps approach where the mean functions are modelled via Generalized Additive Models, and the residuals of such models are investigated for any structure nested at covariance level. If such structure exist, we describe the variation manifested by the covariances through covariance principal component analysis. The 2-steps approach allows to uncover any variation not explained by generalized additive modelling, as well as fill a known shortcoming of these models into incorporating complex correlation structures in the data. The proposed method is illustrated on an articulatory dataset contrasting the pronunciation non-sensical bi-syllabic combinations in the presence of a short-memory challenge