Therapeutic Spoken Dialogue System in Clinical Settings: Initial Experiments

The article presents initial experiences using Terabot, a task-oriented spoken dialogue system, in a psychiatric clinical study. Seven patients conversed with the system, having five therapeutic sessions each. This allowed the collection of 600 recordings, which were then analyzed. We observed speech and intent recognition accuracy and compared it with the results achieved with healthy testers. We demonstrated that speech and intent recognition decreased with the patient testing, and discuss potential reasons for this. We also showed that by adding words that were misrecognized to the training data for the dialogue system, we managed to increase recognition accuracy for some intents. Six out of seven patients confirmed that the relaxation exercises offered by Terabot were helpful, and five even said they liked talking with the dialogue system. These results encourage us to continue the study and further develop the dialogue model.