UC San Diego

Losing the ability to speak—whether from stroke, traumatic brain injury, or other neurological disorders— significantly reduces a person’s quality of life. Research studies demonstrate proof-of-concept Speech-Synthesis Brain-Machine Interface (BMI) systems, but several limitations impede their clinical viability. A major rate limiting factor impeding progress in developing speech prosthesis is the lack of an established animal model to ask basic science questions regarding the neural encoding of vocal communication. This dissertation aims to address this gap by establishing songbirds as an animal model for a human speech prosthesis. Songbirds are a well-established model for vocal learning, and their motor nuclei are homologous to the human motor cortex with respect to function and gene transcription. This work builds upon this basis to demonstrate songbirds’ suitability as a preclinical model to accelerate the development of speech-prosthesis technology.First, we answer basic science questions regarding nuclei important for the song system by characterizing neurophysiological similarities and differences with respect to human motor areas during vocalization. In analyses of data recorded with electrodes chronically implanted in the premotor region HVC of awake free-behaving zebra finches, we detail novel Local Field Potential (LFP) signatures correlated to vocal behavior. These LFP signatures are decomposed using signal processing methods to characterize their relation to vocal production. This work found that HVC exhibits many remarkably similar spectral characteristics to LFP in human motor cortex during speech. Next, we developed proof-of-concept systems that demonstrate algorithms feasible for real time vocal BMIs. Utilizing simple algorithms, we show that HVC LFP features can be leveraged to predict vocal activity. These algorithms can be run in real time to predict both the identity and onset of syllable production. Leveraging these simple algorithms, we analyze preliminary system requirements necessary for decoding vocal elements. The methods developed to leverage these LFP features to predict vocal behavior can be implemented in real-time and suggest a path for developing a similar system for humans. Finally, this thesis details both software and hardware designs to enable reproduction and wider adoption of the songbird animal model by the speech prosthesis research field. We developed novel methods to partition freely produced vocal behavior data based on the subjects’ behavior, which are provided to the field as open-source software. We also designed an integrated counterweight and tether management system that dramatically lowers the stress on chronically implanted small animal subjects. Collectively, these works enrich the literature connecting human and avian vocal-motor production, and we believe strengthen the argument for utilizing songbirds to supplement human speech prosthesis research.