Screen Reading Enabled by Large Language Models

Large language models (LLMs), such as the pioneering GPT technology by OpenAI, have undeniably become one of the most signifcant innovations in recent history. They have achieved phenomenal success across a broad spectrum of applications in numerous industries, transforming how we interact with the digital world. Notwithstanding these remarkable successes, applying LLMs within the realm of accessibility has largely been unexplored. We introduce Savant, as a demonstration of the potential of LLMs for accessibility. Specifcally, Savant leverages the impressive text comprehension abilities of LLMs to provide uniform interaction for screen reader users across various applications, mitigating the signifcant interaction burden imposed by the heterogeneity in user interfaces for blind screen reader users. Savant automates screen reader actions on control elements like buttons, text felds, and drop-down menus via spoken natural language commands (NLCs). Interpreting the NLC, identifying the correct control element, and formulating the action sequence are facilitated by LLMs. Few-shot prompts supply context and guidance for the LLMs to produce appropriate responses, specifically converting the NLC into a correct series of actions on the user interface elements, which are then performed automatically. The demonstration will exhibit Savant's capability across a variety of exemplar applications, emphasizing its versatility.