Analysis of American English corner vowels produced by Mandarin, Hindi, and American accented speakers and a baseline accent recognition system

Individuals' speaking styles such as their accent present challenges for designing speaker independent systems and affect the performance of the systems. Accent recognition systems can be used to improve voice activated systems and Automatic Speech Recognition (ASR) systems. In this paper, we focus on analyzing American English corner vowels produced by Mandarin, Hindi and American accented speakers by using Mel-frequency cepstral coefficients (MFCCs), Linear Predictive coding (LPC), and the first two F1-F2 formant frequencies. A baseline system is designed and tested. The vowel spaces of Mandarin and Hindi accented male speakers of American English are found to be smaller compared to the female L2 speakers' vowel spaces. A significant difference among the L1 and L2 speaker groups is observed for the back corner vowel /a/. In the baseline system, MFCCs feature set achieved higher classification accuracies than the LPCs set.