A method, computer program, and computer system are provided for real-time caption generation in an edge computing environment. Contexts related to one or more participants using a caption service in a web conference service are monitored. Personal characteristics associated with each of the participants are determined based on the monitored contexts. An edge device is selected from among a plurality of edge devices based on the determined personal characteristics. The selected edge device is configured to perform caption conversion a participant from among the participants in the web conference service. A lightweight user accent-oriented caption edge module associated with the selected edge device is customized for the participant. The customized lightweight user accent-oriented caption edge module is deployed to the selected edge device based on a corpus of captions associated with the user accent-oriented caption edge module most closely matching the determined personal characteristics.