Svara: Private On-Device Voice Dictation for Windows and Linux

AI products · Creator, sole engineer · released

System-wide dictation for Windows where recognition runs entirely on your own machine - CPU or GPU. Press a hotkey, speak, and the text appears at your cursor in whatever application you are using. Nothing is uploaded.

  • 54 ms to first word (Linux, CPU)
  • 5.56% word error rate

Code: https://github.com/vasu-devs/Svara

How it works

Recognition never leaves the machine. CPU is a first-class path, not a fallback.

  • Hotkey → Pre-roll buffer
  • Pre-roll buffer → On-device STT
  • On-device STT → Cleanup
  • Cleanup → Your cursor
  • On-device STT → Your cursor

The detail

Cloud dictation tools are excellent and unacceptable for anything sensitive - they stream everything you say, including client names and credentials read aloud, to someone else's servers. Svara is the local answer: no account, no subscription, works offline. CPU inference is a first-class path rather than a fallback, so it runs on a laptop without a GPU.

It is built to survive real use: a pre-roll buffer so the first syllable is never clipped, streaming transcription, and insertion that knows what app it is typing into - terminal-safe handling, editor-aware paste, and a clipboard fallback when Windows privilege boundaries block input. Unfinished dictation is protected with temporary recovery audio, and transcript content is redacted from ordinary logs.

Personal vocabulary, local history, optional LLM cleanup through a local model, tray controls, and a packaged release with an update flow that waits for you.

A benchmark harness measures latency and accuracy on your own hardware rather than quoting numbers from mine, and regression tests cover cleanup, insertion, configuration and log redaction.

A Linux version (a preview release, October 2026) types live while you speak, polishes each sentence in place with a local model, and takes spoken edits such as "scratch that". Its engine was chosen by measurement on the same machine: 54 ms to the first word and 5.56% word error rate, against 290 ms and 8.33% for the previous engine. A short reading set adapts the model to your voice, and the adapted model is kept only if it wins on held-out lines.

Python, faster-whisper, Moonshine, CTranslate2, Windows APIs, Wayland, SQLite, Ollama. Open source under AGPL-3.0.

What I did

  • Built the whole application - capture, streaming transcription, insertion, tray UI and packaging
  • Added a pre-roll audio buffer so the first word of a dictation is never lost
  • Built app-aware insertion: terminal-safe handling, editor-aware paste and a clipboard fallback across Windows privilege boundaries
  • Made CPU inference a first-class path so it runs on laptops without a GPU
  • Protected unfinished dictation with temporary recovery audio and redacted transcripts from ordinary logs
  • Added personal vocabulary, local history and optional LLM cleanup through a local model
  • Built a benchmark harness that measures latency and accuracy on the user's own hardware
  • Wrote regression tests covering cleanup, insertion, configuration and log redaction
  • Ported it to Linux (Wayland, Hyprland): live typing at the cursor, in-place polish, spoken edits, and an engine picked by benchmark (54 ms first word, 5.56% WER)

Stack

Python, Speech Recognition, Machine Learning, Desktop Application, Whisper, CUDA, PyInstaller, Audio Processing, Open Source, Benchmarking, CTranslate2, SQLite, Ollama, Windows APIs, Local AI, Linux, LoRA Fine-Tuning

Case study: https://www.siddhvasudev.com/projects/svara