speakcursor

Hold a key, speak, release — the text appears wherever your cursor is. Speech recognition runs locally, no internet required.

Русская версия

• Hold F8 — plain dictation: recognize, fix, paste.
• Shift + F8 — smart dictation: a stream of thought becomes a structured Markdown document (specs, notes, README drafts).
• Tray icon: grey means idle, red means recording, orange means processing.
• Tray menu: History (last 20 transcripts, click to paste again), Settings, About, Exit.

The app has no main window — only a tray icon and a settings dialog.

Smart dictation

The feature that sets this apart from other Whisper dictation tools. Hold Shift+F8, think out loud, and the rambling gets turned into a clean document instead of a wall of text.

You say:

so basically we need a settings screen where the user can put in the api key and also pick a model and uh also there should be a hotkey field and it should save automatically and yeah the autostart checkbox too

You get:

## Settings screen

- API key input
- Model selection
- Hotkey field
- Autostart checkbox

Settings are saved automatically.

Requirements

• Windows 10/11, Python 3.12
• A microphone
• NVIDIA GPU — recommended, but not required

On a GPU (tested on an RTX 5070 Ti) the large-v3 model transcribes 11 seconds of speech in about one second. Without a GPU the app falls back to CPU, where large-v3 is noticeably slow — pick a smaller model (small or base) in the settings.

Install

git clone https://github.com/VllSunday/speakcursor.git
cd speakcursor
python -m venv .venv
.venv\Scripts\pip install -r requirements.txt
.venv\Scripts\pythonw.exe main.py

The first run downloads the Whisper model (large-v3 is about 3 GB) into the HuggingFace cache. The tray icon shows up immediately, but recognition only starts working after ~15 seconds.

Optional GPT cleanup

Raw Whisper output can mangle technical terms — it happily turns "Whisper" into something else entirely. Sending the transcript through OpenAI fixes terms, grammar and punctuation.

Paste an API key into Settings, or set the OPENAI_API_KEY environment variable. The model is configurable (gpt-5-nano, gpt-5-mini, gpt-4.1-mini, gpt-4o-mini, or anything you type in). Cost per dictation is negligible.

Without a key — or with the "Use GPT" checkbox off — the raw Whisper text is pasted and nothing ever leaves your machine.

Administrator rights

If you dictate into a window that runs as administrator (an elevated terminal, Registry Editor, Task Manager), the app has to run as administrator too. Windows blocks input sent from a normal process into an elevated window, so the text simply will not paste.

The "Start with Windows" checkbox creates a Scheduled Task with highest privileges: it starts silently at logon with no UAC prompt. To create it, enable the checkbox while running the app as administrator.

If you never dictate into elevated windows, a normal launch is fine.

Pasting

Default mode is "Auto": terminals (Warp, Windows Terminal, cmd, PowerShell, WezTerm, Alacritty…) get Ctrl+Shift+V, because Ctrl+V is taken by the terminal itself; everything else gets Ctrl+V. The clipboard contents are saved and restored afterwards.

If some app understands neither shortcut, switch to "Type text" mode in the settings — it emulates keystrokes, works everywhere and never touches the clipboard, but is slower for long text.

Freeing memory when idle

The large-v3 model holds about 3 GB of VRAM the whole time the app sits in the tray, which is annoying when the GPU is needed for something else. "Unload model from memory" in the settings frees it after 5, 10 or 15 minutes without dictation; "Never" (the default) keeps it loaded.

The trade-off: the first dictation after an unload takes a few seconds longer, because the model is read from disk again. Everything after that is as fast as usual.

Build an .exe

.venv\Scripts\pip install pyinstaller
.venv\Scripts\pyinstaller build.spec

dist\VoiceInput\VoiceInput.exe runs without a console window. The build is about 3 GB, almost entirely CUDA libraries (cuDNN and cuBLAS).

Data location

%APPDATA%\VoiceInput\ holds settings (settings.json), history (history.json) and the error log (error.log).

Layout

File	Purpose
main.py	entry point, wires everything together
tray.py	tray icon, menu, settings dialog
hotkeys.py	hotkey hold detection
audio.py	microphone recording
whisper_service.py	local recognition (faster-whisper)
gpt_formatter.py	optional OpenAI cleanup
clipboard.py	pasting into the active window
settings.py	settings and autostart
history.py	last 20 transcripts