Towards Speaker-adaptive Speech Recognition based on Surface Electromyography

We present our recent advances in silent speech interfaces using electromyographic signals that capture the movements of the human articulatory muscles at the skin surface for recognizing continuously spoken speech. Previous systems were limited to speaker- and session-dependent recognition tasks on small amounts of training and test data. In this paper we present speaker-independent and speaker-adaptive training methods which for the first time allows us to use a large corpus of data from many speakers to reliably train acoustic models. On this corpus we compare the performance of speaker-dependent and speaker-independent acoustic models, carry out model adaptation experiments, and investigate the impact of the amount of training data on the overall system performance. In particular, since our data corpus is relatively large compared to previous studies, we are able for the first time to train an EMG recognizer with context-dependent acoustic models. We show that like in acoustic speech recognition, context-dependent modeling significantly increases the recognition performance.