Production of Cantonese Lexical Tones by Native Speakers of Brazilian Portuguese: A Comparative Analysis

Abstract

This study presents a quantitative analysis of the production of lexical tone by speakers of a non-tonal language. More specifically, we look at how well native speakers of Brazilian Portuguese are able to produce the six tones of Cantonese. An experiment was conducted where six subjects listened to a number of tone words previously produced by a native speaker of Cantonese. For each word they heard, the subjects were asked to try to reproduce it as best as they could. No further instructions were provided. A total of 150 tokens per subject were produced (5 phonetic strings × 6 tones × 5 repetitions). The quality of the non-native productions was assessed by comparing them to the productions of the native speaker. This was done by extracting the F0 contours from all recorded tokens (both native and non-native), approximating them by polynomials (first, second, and third order) and then using the polynomial parameters to group the tokens into six classes corresponding to the six lexical tones of Cantonese. The analysis was performed in two ways: (1) by using an F0 onset × offset graph already proposed in the literature and (2) through a classification task using a Linear Discriminant Analysis (LDA). In both cases, each subject was analyzed separately. The F0 onset × offset graphs presented differences between native and non-native data; the non-native productions showed great variability while the native productions were better defined, presenting less variability. Classifiers trained on the Cantonese native speaker’s data resulted in accuracies above 90%, whereas those trained on the Brazilian Portuguese speakers’ data resulted in accuracies between 40 and 60%.