A Probabilistic Decoding Approach to a Neural Prosthesis for Speech

Neural prosthetic systems for motor control and communication have produced striking results in recent studies with non-human primates and human volunteers. We describe a new approach in our ongoing work toward developing an intracortical neural prosthesis for speech restoration with a 26 year old human volunteer with tetraplegia (including loss of vocal and facial muscle control). We propose to use hidden Markov models (HMMs) to decode neural firing activity in speech motor cortex. We show how classical and recent approaches to automatic speech recognition (ASR) apply directly to the decoding stage of a neural prosthesis. We outline a series of experiments in collecting cortical neural firing data from our human volunteer, and discuss important challenges and considerations in implementing an HMM framework for a neural speech prosthesis.