Linguistic synchrony is a potentially critical factor in developing AI tutors, but its computational modeling using Transformer-based embeddings remains underexplored. This study evaluates the reliability of synchrony measures from four Transformer embeddings–SBERT, MBERT, DistilMBERT, and XLM-R–across 128 one-on-one online L2 tutoring sessions. We implement a bi-encoder architecture to measure synchrony at both global and local levels. Results show that models with similar pre-training architectures produce more consistent synchrony measures, while multilingual models introduce variability, particularly in bilingual contexts with code-switching. Our findings emphasize the importance of model selection in the educational context. These insights advance AI-driven linguistic modeling and have implications for personalized AI tutoring systems.