JAM: A Unified Neural Architecture for Joint Multi-granularity Pronunciation Assessment and Phone-level Mispronunciation Detection and Diagnosis Towards a Comprehensive CAPT System

Computer-assisted pronunciation training (CAPT) systems are designed for second-language (L2) learners to practice their pronunciation skills by offering objective, personalized feedback in a stress-free, self-directed learning scenario. Both mispronunciation detection and diagnosis (MDD) and automatic pronunciation assessment (APA) are indispensable components in the CAPT systems. The former is responsible for pinpointing phonetic pronunciation errors and providing the corresponding diagnostic feedback, while the latter manages to evaluate oral skills across various linguistic levels with disparate aspects. Most existing efforts typically treat APA and MDD as independent tasks, where the correlations between the assessment scores and the phonetic pronunciation errors are nearly sidelined. In light of this, we introduce JAM (a Joint neural model for APA and MDD), a novel end-to-end neural model for CAPT that streamlines the components of APA and MDD into a unified structure with a parallel pronunciation modeling architecture. To capture fine-grained pronunciation cues from L2 learners' speech, electromagnetic articulography (EMA) features are introduced for the proposed model, which portrays the movement of articulatory structures, such as the jaw, lips, and tongue. A series of experiments conducted on the speechocean762 benchmark dataset demonstrate the feasibility and effectiveness of our approach compared to several competitive baselines. Additionally, an ablation study is presented to assess the contributions of different input features and training strategies in the proposed model.