An ASR-Based Interactive Game for Speech Therapy

The demand for intensive and costly speech therapy to patients impaired by communicative disorders can potentially be alleviated by developing computer-based systems that provide automatized speech therapy in the patient’s home environment. In this paper we report on research aimed at developing such a system that combines serious gaming with automatic speech recognition (ASR) technology to provide computer-based therapy to dysarthric patients. The aim of the serious gaming environment is to increase the patients’ motivation to practice, which tends to decrease over time with conventional speech therapy, as progress in dysarthric patients is often slow. Additionally, some speech exercises (e.g. drills) are not particularly motivating due to their repetitive nature. The ASR technology is aimed at providing feedback on speech quality during training to improve speech intelligibility. Different types of acoustic models were trained on normal speech of adults and elderly people, and tested on dysarthric speech. The results show that speaker-adaptive training and Deep Neural Networks (DNN)based acoustic models substantially improve the performance of ASR in comparison to traditional GMM-HMM-based methods. In this specific case, the ASR-based game is developed to provide speech therapy to dysarthric patients, but this approach can be adapted for use in other types of communicative disorders.