DEV Community

Franklin AI
Franklin AI

Posted on

Open Mic: Building Local Voice Input Across the Desktop

Open Mic: local voice input across your operating system

Open Mic is an open-source desktop tool for turning voice input into cleaned-up text in the application currently being used.

A shortcut starts recording. The spoken input is sent through a local language model, then the formatted result is inserted into the active text field. The video demonstrates this with a grocery list that is corrected and formatted automatically.

Local model and capabilities

The project uses the open-source EdgeRock 4B model. The stated goal is fully local, offline processing, so voice and related data remain on the user’s machine.

The tool can:

  • Clean up grammar and filler words
  • Rewrite or translate selected text
  • Summarize highlighted content
  • Apply a custom rewrite instruction
  • Capture a region of the screen
  • Analyze image and text input together
  • Continue a conversation using earlier context

The screen-capture workflow is particularly different from basic dictation. A user can select a chart, image, or code region and ask the model to explain it. The model then accepts follow-up questions about the same context.

Context handling

The video describes EdgeRock as using a recurrent state to summarize global context, a local window for recent tokens, and a fixed-size anchor bank for important long-range details. These “memory anchors” are intended to preserve high-value information after it falls outside the recent-token window.

The video also discusses a reported 79.8 score on MMLU Pro and a diffusion-style decoding approach that generates and refines groups of tokens rather than decoding only one token at a time. Those benchmark and performance details come from the model materials discussed in the video; this post does not independently reproduce them.

Open Mic and the related model resources are linked here: https://youtu.be/Yrx2BZ6YtIs

Top comments (0)