Prominence-based Evaluation of L2 Prosody

Abstract

Prosody in terms of word and sentence stress is one of the most difficult features for many second language (L2) speakers to learn and it can be hypothesized that assessing the learner's prosodic abilities could provide a good measure for assessing the learners' spoken language skills in general. Automatic assessment is, however, dependent on reliable automatic analyses of prosodic features for comparing the productions between native (L1) and L2 speech. Here we investigate, whether estimated prosodic prominence levels of syllables can be used to predict the prosodic competence of Finnish learners of Swedish. Syllable level prominence was estimated for 99 L2 and 25 native Swedish utterances using continuous wavelet transform analysis with combinations of f0, energy, and duration features. The L2 utterances were assessed by four expert raters using the revised CEFR scale for prosodic features. Correlations of prominence estimates for L2 utterances with estimates for L1 utterances and linguistic stress patterns were used as a measure of prosodic proficiency of the L2 speakers. The results show that these estimates correlate significantly with the assessments of expert raters. Overall, the results provide strong support for the use of the wavelet-based prominence estimation techniques in automatic assessment of L2 proficiency.

Fluency-related Temporal Features and Syllable Prominence as Prosodic Proficiency Predictors for Learners of English with Different Language Backgrounds

“…However, distinguishing between neighboring levels, especially at the higher end of the scale, proved more challenging. This was also the case in our previous studies using the same CEFR scale ( Kallio et al, 2017 , 2018 , 2020 ). We suspect one reason for this to be the cumulative nature of the rating scale: the differences between proficiency levels are greater at the lower part of the scale because the language learning process is seen as progressive or accumulative.…”

“…This allows us to take into account both word and sentence level production of prominence and compare the stress realizations between speakers in more detail. In Kallio et al (2018Kallio et al ( , 2020, we compared the estimated syllable prominence values between L2 and native speakers of Finland Swedish from read utterances and found that the higher the correlation in prominence values between native and L2 speakers, the better the prosodic proficiency grades of L2 speakers. However, the previous studies provided somewhat unexpected results regarding the feature combinations that best predicted the proficiency scores: energy and duration together provided the best results, which indicate that in the speech of L2 learners there might be a revealing pattern in the use of these features when producing prominence.…”

“…The individual signals were z-scored and the combinations were obtained as weighted sums of appropriate signals using the following feature weights 1.0, 1.0, and 0.5, for F0, en, and dur, as in Kallio et al (2018Kallio et al ( , 2020. The resulting combined signals were subjected to the CWT using a Mexican Hat mother wavelet with scales a quarter of an octave apart.…”

“…In this study, we estimate syllable prominence using a recent methodology based on the continuous wavelet transform (CWT) of prosodic features. The method has been shown to provide strong correlations with perceptual prominence ( Eriksson et al, 2018 ; Suni et al, 2017 ) as well as promising results in predicting prosodic proficiency of L2 learners ( Kallio et al, 2018 , 2020 ). Compared to previous methods used in detecting syllable stress, we use relative prominence values for all syllables in an utterance instead of categorizing syllables as stressed or unstressed.…”