Weighing Phonetic Patterns in Non-native English Speech

Zhiyan Gao

Major Professor: steven h. weinberger, PhD, Department of English

Committee Members: Harim Kwon, Douglas Wulf, Dennis Perzanowski

Robinson Hall B, #434
November 26, 2019, 10:00 AM to 12:00 PM


Non-native (L2) speakers of English often speak English with a certain degree of foreign accent. While much research has investigated the accentedness of L2 English speech, very few studies have attempted to rank phonetic patterns in L2 speech according to their perceived foreign accentedness. This dissertation provides accentedness rankings of a large variety of phonetic patterns in L2 English speech. By investigating why some phonetic patterns in L2 English speech are perceptually more accented than others, this dissertation reveals the possible underlying mechanisms that govern foreign accent perception.

This dissertation focuses specifically on the segmental and syllable structural aspects of L2 speech. Two perception experiments were conducted to elicit native (L1) American English raters’ accentedness judgments on 100 L2 stimuli extracted from the Speech Accent Archive (Weinberger, 2019). In both experiments, raters heard L2 English speech samples and rated foreign accentedness for each speech sample. Accentedness rankings of various phonetic patterns in L2 speech were therefore obtained. Linear mixed-effects models were implemented to investigate the effect of different types of phonetic patterns on accentedness perception. Prosodic information of the stimuli was accounted for in the least intrusive manner by employing a Dynamic Time Warping approach. Results of the two experiments showed that (1) the consonant and syllable structural aspects of an L2 speech carry more weight in accentedness perception than vowels; (2) phonological context affects accentedness perception; (3) raters are aware of which phonetic patterns are allowed in L1 speech, showing that L1 phonetic and phonological knowledge (L1 knowledge) might affect accentedness perception.

A third experiment was further conducted to investigate how raters’ L1 knowledge affected their accentedness judgment. A Naïve Discriminative Learning Model (NDL) was employed to estimate raters’ L1 knowledge by examining the co-occurrences of pronunciations (e.g., [æsk]) and lexical outcomes (e.g., “ask”) in L1 speech of American English. The 100 L2 stimuli were subsequently evaluated against listeners’ L1 knowledge to estimate the degree of dissimilarity (NDL- distance) between L1 and L2 speech samples. The results showed that NDL-distance correlates significantly with accentedness ratings, suggesting that L1 knowledge, as esimated by the NDL model, could have affected accentedness judgments.

This dissertation contributes to the field of foreign accent by providing accentedness rankings of various phonetic patterns in L2 speech. In lieu of ad hoc explanations for why some phonetic patterns are more accented than others, this dissertation directly examines how raters’ L1 knowledge affected their accentedness judgments on L2 speech, providing insights into the nature of foreign accent perception