Automatic speech recognition for users with a voice impairment

Speech disorders affect many people in our society. They cause communication difficulties and social isolation. Speech disorders have a variety of causes: malformation of organs (e.g. cleft palate, cheiloschisis, …), surgery (throat cancer), neurological diseases (multiple sclerosis, Parkinson’s, …) or the consequence of a stroke (CVA). Modelling these forms of pathological speech is relevant for the medical field, e.g. diagnosis, quantification of the severity of the condition and in therapy. Automatic recognition of pathological speech is useful for building communication devices that transform disordered speech into normal speech and for building assistive command-and-control devices (e.g. home automation) to increase autonomy, comfort and safety in case the speech impairment co-occurs with physical impairment of the limbs (often the case in neurologically induced speech pathology). The primary aim is to advance automatic speech recognition (ASR) for disordered speech, i.e. extraction of the verbal content into text or semantics. The state of the art requires users to enroll their voices and the goal is to minimize the duration of this phase, i.e. showing better accuracy under shorter enrollment (Ons, 2014). Also, thanks to the collaboration with the Technical University of Budapest (under CELSA financing) we will be able to incorporate computational representations of speech features that have shown to show relevance for medical diagnosis of voice disorders (Kazinczi et al., 2015; Vicsi et al., 2012) but also to contribute the new feature extractors based on prior work (Stouten, 2009).