Browsing by Subject "L2 speech"

Speaking fluently is an important goal for second language (L2) learners. In L2 research, fluency is often studied by measuring temporal features in speech. These features include speed (rate of speech), breakdown (use of silent and filled pauses), and repair (self-corrections and repetitions) phenomena. Fluent speakers generally have a higher rate of speech and fewer hesitations and interruptions than beginner language learners. In this thesis, phonetic fluency of high school students’ L2 Finnish speech is studied in relation to human ratings of fluency and overall proficiency. The topic is essential for the development of automated assessment of L2 speech, as phonetic fluency measures can be used for predicting a speaker’s fluency and proficiency level automatically. Although the effect of different fluency measures on perceived fluency level has been widely studied during the last decades, research on phonetic fluency in Finnish as L2 is still limited. Phonetic fluency in high school students’ speech in L2 Finnish has not been studied before. The speech samples and ratings used in this thesis are a part of a larger dataset collected in the DigiTala research project. The analyzed data contained spontaneous speech samples in L2 Finnish from 53 high school students of different language backgrounds. All samples were assessed by expert raters for fluency and overall proficiency. The speech samples were annotated by marking intervals containing silent pauses, filled pauses, corrections and repetitions, and individual words. Several phonetic fluency measures were calculated for each sample from the durations of the annotated intervals. The contribution of phonetic fluency measures to human ratings of fluency and proficiency was studied using simple and multiple linear regression models. Speech rate was found to be the strongest predictor for both fluency and proficiency ratings in simple linear regression. Articulation rate, portion of long silent pauses, mean duration of long silent pauses, mean duration of breaks between utterances, and rate of short silent pauses per minute were also statistically significant predictors of both fluency and proficiency ratings. Multiple linear regression models improved the simple models for both fluency and proficiency: for fluency, a model with a combination of articulation rate and the portion of long silent pauses performed the best, and for proficiency, a model with a combination of speech rate and mean duration of short silent pauses. Perceived fluency level is often affected by a combination of different phonetic fluency measures, and it seems that human raters ground their assessments on this combination, although some phonetic fluency measures might be more important on their own than others. The findings of this thesis expand previous knowledge on phonetic fluency in L2 Finnish and can benefit both language learners and teachers, as well as developers of automatic assessment of L2 speech.