Comprehensibility diagnosis of spontaneous L2 English: Automated analysis of pausing and lexical stress patterns
Résumé
There are various automatic scoring systems that can estimate the proficiency level of a language learner based on an analysis of certain features of speech, such as segmental information, speech rate, intonation or the frequency of pauses. Although these tools are very good at predicting proficiency, they do not provide insight into the specific causes of a learner's difficulties, making them less useful in a training context (Zechner et al., 2019). To effectively identify areas for improvement, it is necessary to first define what constitutes "good pronunciation" in a foreign language. A common approach is to compare a learner's speech with that of a native speaker, and base the evaluation on the difference between the two utterances. However, there are so many possible variations between native speakers that it would be ambitious to think that all of them could be represented in a single model; and it would be unfair to consider only one theoretical standard, since native speakers themselves deviate widely from it (Walker et al., 2021). Another method of evaluation consists in measuring the effort required for the listener to understand the speaker (Isaacs et al., 2017). This would not involve measuring a factual difference with a model, but rather identifying the elements that cause comprehension difficulties or, conversely, that allow for fluidity and optimization of communication. The boundary between levels B1 and B2 is often characterised by fluency of expression and the ease with which the speaker can be understood. Research has shown that speech rhythm, particularly lexical stress and pauses, play an important role in English (Jenkins et al., 2018; Frost & O’Donnell, 2018), and these factors are often mentioned by raters at B1/B2 levels. The goal of this work is to develop an automatic tool to measure these phenomena, and to first describe how English actually “sounds” like in the speech of French and Japanese learners of different global proficiency levels: do they stress words? Which ones? On which syllable? Do they make it higher? Longer? Louder? Where do they slice the flow of speech? At this stage, nothing tells us what is “good” or “bad”. It may be that stressing the wrong syllable is a barrier to comprehension, or not. These prosodic measurements allow us to cluster learners into profiles of speakers. The next step will consist of asking native and naïve English listeners to judge, while listening representative excerpts of learners’ speech, when they struggle to understand, through an incremental judgment test (de Kok, 2013). This should allow us to verify the actual impact of stress and pause position for comprehensibility, as well as to make a qualitative judgment about different type of non-native pronunciation. The ultimate goal is to end up with an online tool which enables learners to identify specific elements of their speech that need improvement and to aid teachers in addressing the difficulties of their students.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |