Skip to main content
Selected work

Local-first dictation

yap

Speak freely. Keep it local.

Local Windows dictation with on-device speech models, searchable history and optional local text models for cleanup, prompts and chat.

Year
2026
Platform
Windows first
Source
Open source / MIT
Record with a shortcut, transcribe speech on your computer, then paste into the active app.

The idea

I built Yap to experiment with speech models and local LLMs while creating my own alternative to apps like Superwhisper. It turns speech into text you can clean up, summarize or reshape with custom prompts. The entire workflow runs on your computer: no accounts, no subscriptions, and no recordings or transcripts sent to a datacenter.

How it works

  1. Set up once

    Download a speech model and test your microphone. Choose Whisper, NVIDIA Parakeet or Moonshine to suit your language and computer.

  2. Record with a shortcut

    Place the cursor in another app. Hold the global shortcut to record, then release it to transcribe and paste. In toggle mode, press the shortcut once to start recording and again to stop.

  3. Work with your transcripts

    Import audio files for transcription, or work with your saved dictations. Use prompts and chains to clean up, summarize or reshape the text, then edit or export the result.

Under the hood

  • Go
  • Wails
  • React
  • TypeScript
  • Whisper
  • SQLite
  • Tailwind CSS

Speech recognition on the device

Go coordinates native audio capture and local model runtimes, while Wails hosts the React interface. Models and runtimes download during setup; dictation works offline afterward. No Python installation is needed.

Profiles for different use cases

Create profiles with a speech model, language and prompt choices for different tasks. Profiles can switch automatically per app, while global shortcuts, a floating recording indicator and tray controls keep dictation available wherever you are working.

A choice of local text models

Connect to LM Studio or Ollama to use models you already have installed, or use Yap's built-in runtime with Gemma 4 models. Custom prompts and chains let you process transcripts in several steps and keep the intermediate results.

History you can work with

SQLite stores searchable history locally. Saved transcripts support editing, export and optional chat to ask questions about the text. History saving can be turned off for temporary dictation; retaining audio is off by default.

Good to know

Windows is the main platform for now. I plan to follow with macOS once I have a MacBook again for development and testing. Linux support is a work in progress. Initial model downloads need internet access.