PROJECT SUMMARY / ABSTRACT The goal of this proposal is to improve our fundamental understanding of speech production, and to translate this knowledge into medical devices called intracortical brain-computer interfaces (iBCIs) that will enable people who have lost the ability to speak to fluently communicate via a computer just by trying to speak. The study will enroll participants who have lost or are actively losing their ability to as part of the ongoing BrainGate2 clinical trial (a multi- site trial investigating the feasibility and safety of iBCIs in people with paralysis.) We will place four chronic 64- channel silicon microelectrode arrays in speech areas of cortex and use simultaneously recorded speech (or attempted speech) data and action potential-resolution neural data to characterize and decode in real time the link between neural activity and speech production. A key advance of this approach is our ability to record simultaneously from hundreds of individual neurons. This will provide a much more detailed view of the underlying neural computations at their fundamental resolution – spikes – and much higher signal-to-noise ratio signals for decoding attempted speech. The project is divided into two research areas. First, we seek to understand how networks of neurons in two closely interacting cortical areas – ventral (speech motor) precentral gyrus (PCG) and superior temporal gyrus (STG) – generate speech. These areas have not previously been recorded from at scale with single neuron resolution and characterizing their dynamics will be critical for designing an effective speech iBCI. We will examine “what” parameters of speech neurons in these two areas encode, and “when” during the preparation and production of speech different neural computations occur. We will apply cutting-edge dynamical systems and “neural population doctrine” analysis approaches to disentangle various components of neural activity (e.g., motoric, sensory feedback, error-processing) that are distributed across neurons. Second, we will apply state-of-the art artificial intelligence (AI) and control theory- inspired iBCI methods to translate the neural activity that accompanies the person’s attempt to speak into words that appear on a computer (“Brain-to-Words”) and directly into sound (“Brain-to-Voice”). Brain-to-Words, in which speech units such as phonemes are decoded from a slightly delayed window of neural activity, is more constrained because the final output is whole words (which could then be spoken by the computer), without the full richness of the person’s voice. However, it allows for powerful AI techniques to be applied to automatically correct for errors based on the known statistics and rules of language. Brain-to-Voice can potentially restore the full expressive range of speech but requires higher accuracy for intelligible real-time voice synthesis. However, we hypothesize that control theory-based iBCI design and the brain’s immense ability to learn ...