PG-MDD: Prompt-Guided Mispronunciation Detection and Diagnosis Leveraging Articulatory Features

Mispronunciation detection and diagnosis (MDD) manages to pinpoint phonetic errors of L2 (second-language) learners and then provides timely and informative diagnosis on erroneous pronunciation segments. Recently, dictation-based neural methods have emerged as an appealing modeling paradigm for MDD, which simultaneously identifies pronunciation errors and provides diagnostic feedback by aligning the recognized phone sequence to the corresponding canonical phone sequence of a given text prompt. Despite their decent performance in terms of F1-score, dictation-based models still struggle to accurately detect pronunciation errors with balanced precision and recall evaluations, resulting in inferior learning efficiency for L2 learners. In view of this, we propose a novel prompt-guided dictation-based MDD model, dubbed PG-MDD, that can efficiently strike a balance the precision and recall rates while maintaining a high-performing F1-score. PG-MDD first jointly optimizes the mispronunciation detection and diagnosis processes during the training phase, while aptly guiding the diagnosis process with phone-dependent thresholds in the inference phase. In addition, a novel multi-view audio encoder is introduced to render the fine-grained articulatory cues within learners' speech. A comprehensive set of empirical experiments conducted on the L2-ARCTIC benchmark dataset suggests the practical feasibility of our method in relation to several competitive baselines.