Show HN: Umless, open-source on-device dictation that writes what you meant (github.com)

Umless has introduced a revolutionary open-source on-device dictation tool that allows users to dictate text in a more conversational style. Unlike traditional dictation software that transcribes precisely what is said, Umless leverages a cleanup language model (LLM) to convert spoken phrases, including fillers and corrections, into polished text tailored to the application being used—be it a casual chat in Slack or a formal email in Outlook. This functionality ensures that the final output reflects the user's intent and context, enhancing the dictation experience. Significantly, Umless operates entirely on-device, meaning no data is sent to the cloud, preserving user privacy. The tool is designed for macOS, Windows, and Linux, and offers different tiers of performance based on user needs—ranging from a lightweight solution with a 5.7% word error rate to a more robust version with as low as 1.8% error. Umless supports real-time spelling of names shown on-screen and adapts its tone based on the context of the conversation. As an open-source project, it invites further community development, offering a flexible platform where users can inspect, swap, or fine-tune the underlying models, thereby pushing the boundaries of personalized AI-driven communication.