Speaking with LLMs: Co-Designing Voice Interaction Patterns with Blind and Visually Impaired Users

The widespread availability of chat-based interfaces for LLMs, such as ChatGPT, has brought AI into everyday professional and personal contexts. Ensuring the inclusivity of this prompt-based interaction paradigm is therefore essential. This paper responds to this need by presenting a human-centered design study involving 119 blind and visually impaired (BVI) individuals. Through questionnaires, interviews, and focus groups, we identified key challenges and derived design patterns for voice interactions with LLMs. These patterns specifically address tasks that remain difficult for BVI users, such as accessing and controlling lengthy model-generated responses, refining prompts iteratively, and retrieving content from past chats. We implemented the identified patterns in a web platform and validated them through additional focus groups. With these contributions, we aim to provide a foundation for the inclusive design of voice interactions with chat-based LLM systems.