Edge Computing Solutions Supporting Voice Recognition Services for Speakers with Dysarthria

In the framework of Automatic Speech Recognition (ASR), the synergism between edge computing and artificial intelligence has led to the development of intelligent objects that process and respond to human speech. This acts as a key enabler for multiple application scenarios, such as smart home automation, where the user's voice is an interface for interacting with appliances and computer systems. However, for millions of speakers with dysarthria worldwide, such a voice interaction is impossible because nowadays ASR technologies are not robust to their atypical speech commands. So these people, who also live with severe motor disabilities, are unable to benefit from many voice assistant services that might support their everyday life. To cope with the above challenges, this paper proposes a deep learning approach to isolated word recognition in the presence of dysarthria conditions, along with the deployment of customized ASR models on machine learning powered edge computing nodes. In this way, we work toward a low-cost, portable solution with the potential to operate next to the user with a disability, e.g., in a wheelchair or beside a bed, in an always active mode. Finally, experiments show the goodness (in terms of word error rate) of our speech recognition solution in comparison with other studies on isolated word recognition for impaired speech.