Fellowship project title: Improving a speech recognizer to detect speech distortions.
The goal: In this project, we improve a popular speech recognizer (Kaldi) so that it will recognize distortions, inconsistent misarticulations. We modify the source code to enable the recognition ofdistortions, and we organize an initiative to draw the attention of other researchers in this domain to this specific problem in particular, and, in general, to the usability of speech processing software for the research of language decline caused by brain damage.
The why: Speech recognition software can be very useful in the research of language of persons that have apraxia of speech, a neurological disorder that causes problems with speech articulation. However, before software can recognize disordered speech, the underlying statistical speech models must first be adapted, because these are typically trained with the speech of healthy participants.