A Speech Accent Detection Metrics (L2AM) for Processing Accented Speech Signal

Abstract

The article proposes a multi-dimensional metrics (L2AM) for measuring second language accent detection by extracting and validating attitude-carrying adjectives from real-time perception tests of leveled accented speech. In Experiment 1, 20 homogeneous Chinese (2 Cantonese and 18 Mandarin) advanced learners of English were asked to provide 10 adjectives to 8 excerpts in three levels of accentedness. Then, the most prominent adjectives of 4 types were elicited as standard rating terms for perceivers. In Experiment 2, 55 participants (7 Cantonese and 48 Mandarin) were told to use these adj ectives to evaluate the same material with 1–5 increments while providing intelligibility, comprehensibility and accentedness scores. Factor analysis results show that sociocultural bias on heavy accent lies most primarily in education and intelligence, and then status and professionalism, and finally centers on linguistic ontology. Interestingly, Mandarin speakers viewed Cantonese speech far more negatively than Mandarin speech in all three levels despite similar intelligibility results, revealing hidden dialectical identities and values. Finally, by correlating the evaluation terms to actual comprehensibility and accentedness through neural network fitting, we have acknowledged the validity of the new metrics.