Voice input causes more damage to LLM accuracy than typing because transcription restructures meaning, not just adds noise—and this can't be fixed with simple fine-tuning or extra thinking time.
This paper studies how voice and keyboard input affect LLM performance by introducing HIVE, a tool that simulates real-world perturbations from both channels. Voice transcription errors hurt accuracy more than typos because they restructure sentence meaning, while keyboard errors are absorbed better. The gap only matters for tasks requiring reasoning, not multiple choice.