Skip to content

Releases: iFrosta/localSTT

Release list

LocalSTT 1.0.0

LocalSTT 1.0.0 Latest
Latest

Choose a tag to compare

View all tags
@iFrosta iFrosta released this 04 Sep 23:06
v1.0.0
bbe8a64

Hold a hotkey, talk, let go — the text is typed where your cursor already is. Speech
recognition that runs entirely on your own GPU, with a Windows 11 tray app, an optional
AI cleanup pass and an OpenAI-compatible HTTP endpoint. Nothing is uploaded.

Which download

Size Use it when
LocalSTT-1.0.0-win64.zip 1.4 GB You just want it to work. Python, every dependency and the CUDA 12 runtime are already inside. Unzip and run.
LocalSTT-1.0.0-win64-nocuda.zip 120 MB You already have a CUDA 12 runtime (cuBLAS 12, cuDNN 9) on the DLL search path. Everything else is included.
Source — You want to follow changes with git pull. Clone the repository and run install.ps1.

The two archives are the same application. The difference is the NVIDIA runtime
libraries — 1.3 GB of cuBLAS and cuDNN, which most people do not already have.

Running it

  1. Unzip anywhere — the folder can be moved or renamed later.
  2. Run start-localstt.vbs. A tray icon appears; green means ready.
  3. Hold Ctrl+Win, talk, let go.

There is no Python to install, no pip to run and no virtual environment. The first
dictation downloads the Whisper model (~1.6 GB) and is slow while it does; every one
after that is not.

What you need

  • Windows 10 version 2004 (build 19041) or newer, 64-bit
  • An NVIDIA GPU with compute capability 7.0 or newer (RTX 20-series, GTX 16-series and
    up), driver 527.41 or newer, and about 4 GB of free VRAM
  • A microphone

There is no CPU fallback. Without a usable CUDA device LocalSTT refuses to start and
says why — it does not fail silently at the first dictation.

Hotkeys
Ctrl+Win Dictate. Hold and release, or tap to start and tap again to stop.
Ctrl+Shift+Win Dictate with AI cleanup — needs Ollama running.
Ctrl+Alt+Win Voice command, matched against commands.json. Nothing is typed.
Esc Cancel. Nothing is transcribed, delivered or run.

All four are rebindable in Settings → Hotkeys: click the field, press the combination,
let go. Any key, not just modifiers.

Notable

  • The clipboard is left alone by default — dictation types the text rather than pasting
    it, so whatever you had copied is still there.
  • A self-test checks the driver, VRAM, CUDA libraries, microphone and port on first run,
    and tells you what to do about each failure.
  • Dictation history is off by default.

Full documentation: README ·
по-русски

Checksums

Verify a download with Get-FileHash <file> -Algorithm SHA256.

SHA256  LocalSTT-1.0.0-win64.zip         E5A184D941E0CBF5F09D0B42E3204F2A4A814D1BF9FACF41DE237A69303CD34B
SHA256  LocalSTT-1.0.0-win64-nocuda.zip  E6971E1F3D41DCA2C0B9C40E7F798A789FAAA84967BDA985CE95A468A25BE595
Assets 4
Loading