Capturing L1 Influence on L2 Pronunciation by Simulating Perceptual Space Using Acoustic Features

Abstract

Theories of second language (L2) acquisition of phonology/phonetics/pronunciation/accent often resort to the similarity/dissimilarity between the first language (L1) and L2 sound inventories. Measuring the similarity of two speech sounds could involve many acoustic dimensions, e.g., fundamental frequency (F0), formants, duration, etc.. The measurement of the sound inventories of two languages can be further complicated by the distribution of sounds within each inventory as well as the interaction of phonology and phonetics between the two inventories. This paper attempts to propose a tentative approach to quantify similarity/dissimilarity of sound pairs between two language inventories and to incorporate phonological influence in the acoustic measures used. The language pairs studied are English and Mandarin Chinese and only their vowel inventories are considered. Mel-Frequency Cepstral Coefficients (MFCCs) are used as features, and Principle Component Analysis (PCA) is used and slightly adjusted to simulate the perceptual space. Similarity/dissimilarity of sound pairs between the language inventories are examined and potential L2 error patterns are predicted based on the proposed approach. Results showed that predicted results using the proposed approach can be well related to those by Speech Learning Model (SLM), Perceptual Assimilation Model for L2 (PAM-L2) and Native Language Magnet Model (NLM).

Domain Generalization for Language-Independent Automatic Speech Recognition

“…A vowel classifier trained on English is able to recognize the duration differences of some Japanese vowel pairs, but not others (Nishi et al, 2008 ). A Mandarin vowel classifier, applied to English vowels, finds American English /u/ to be closer to the Mandarin central unrounded vowel /ɨ/ than to the Mandarin /u/ (Shi et al, 2019 ).…”