Sargosh
A Windows dictation app that turns speech into clear text and inserts it into supported apps.
- Electron
- React
- TypeScript
- Speech Recognition
- LLM Integration
- Fastify
The problem
Speaking is often easier than typing, but spoken words include pauses, fillers, and corrections. A useful dictation app needs to turn that speech into readable text and put it where the user is working, without making them switch apps and paste it manually.
I wanted to build that experience for Windows, including people who switch between Hindi and English while speaking.
What I built
Sargosh runs in the background with a small floating overlay. Users hold a custom shortcut to dictate, or double-tap it for hands-free recording. When they finish, the app transcribes the recording, cleans up the text, and inserts it into the selected text field.
The app supports English, Hindi, and Roman-script Hinglish. Cleanup removes fillers, handles spoken corrections, and formats clear lists. Users can replay recordings, copy transcripts, retry transcription, export audio, and view their speaking statistics and remaining allowance.
Architecture
Electron runs the desktop app and overlay, while React and TypeScript handle the interface and recording flow. A bundled Python helper uses Windows APIs to detect shortcuts, track the selected input, and insert the finished text.
The desktop app sends recordings to a Node.js and Fastify backend on Render. The backend verifies the account, checks usage limits, sends audio to Deepgram, and passes the transcript to an LLM for cleanup. Provider keys stay on the server.
Supabase handles Google and email OTP login, with PostgreSQL storing account-linked usage records. SQLite and local audio files keep recording history on the user's computer, separated by account.
Bugs I actually debugged
Bug 01
Putting text into the right input
A user can change text fields while speaking, and clicking the overlay should not replace the intended destination. Sargosh tracks the latest selected input during recording and freezes that destination when recording finishes. If insertion is unavailable, the overlay offers the finished text for copying.
Bug 02
Keeping repeated dictations reliable
Text insertion worked on the first attempt but could lose characters during repeated recordings. The Windows helper needed careful handling of shortcut release, keyboard state, focus, and character timing. I tested separate hold and hands-free recordings in different orders to check that users could continue writing in the same input.
Bug 03
Enforcing limits outside the desktop app
Cloud transcription and cleanup have a cost, so the backend checks and records usage instead of trusting a counter on the user's computer. The allowance is shared across devices signed into the same account. Setup practice is tracked separately so it does not reduce the regular allowance or add to dictation statistics.
Bug 04
Getting the app onto another Windows PC
The first installer worked on my computer but was blocked by a Windows code-integrity policy on another PC. I traced the failure through installation and Windows event logs, then packaged the app as MSIX and published it through Microsoft Store.
Results
- Published a Windows desktop app on Microsoft Store, with a separate website explaining the app and guiding users through installation.
- Users can dictate into supported apps through a shortcut without manually copying and pasting the result.
- English, Hindi, and Hinglish share one recording and cleanup flow, including spoken corrections and list formatting.
- Recording history stays local and separate by account, while the backend enforces shared usage limits across devices.
Stack
- Desktop & Interface
- Electron · React · TypeScript · MediaRecorder
- Speech & AI
- Deepgram · LLM API Integration
- Backend & Data
- Node.js · Fastify · Supabase Auth · PostgreSQL · SQLite
- Windows Integration
- Python · Windows UI Automation · MSIX
- Hosting & Distribution
- Render · Vercel · Microsoft Store