The idea of using speech recognition in Computer-Assisted Language Learning is becoming increasingly popular. Early systems focussed exclusively on pronunciation practice; now, however, applications are starting to emerge designed to carry out simple conversations with students and give them the opportunity to practise spontaneous speaking skills. Typically, a system of this kind will be deployed on the web, use small-vocabulary recognition, and structure interaction in some kind of game-like way. The novelty of the field, however, means that few solid results have emerged. Despite positive anecdotal studies, it still remains to be demonstrated that systems with these characteristics are in fact able to help students acquire language skills, and, if so, which features influence learning outcomes.
Under funding from the SNF, and beginning in May 2014, we are carrying out a three-year project aimed at systematic investigation of the issues involved. Our starting point is the CALL-SLT system which we built under an earlier SNF-funded project. We are developing the current CALL-SLT prototype in several directions; the goal is both to extend the state-of-the-art in speech-enabled CALL, and to enhance the system to the point where we can perform the experiments we have in mind. Specifically, we will implement and compare a range of different speech recognition and recognition feedback strategies, further develop our existing framework for specifying script-based interactive dialogue games, use this framework to develop a substantial quantity of course content, and embed the system inside a social network. Content development will focus on English and French as L2s, with a variety of L1s including at least French, German, Italian and English. During the second half of the project, we will use these resources to carry out a series of evaluations in secondary schools, universities and on the web. The web-based evaluations will be designed to contrast different versions of the system against each other and will be performed using crowd-sourcing methods, making it feasible to recruit large enough sets of subjects to perform controlled experiments.
The resulting software base, including lingware and courseware, will be made generally available to the community in fully documented Open Source form. We will also release annotated learner speech corpora collected during the experiments.