01
The orb answers first.
A rising beep starts the session. The orb turns red and pulses with your voice, even when the window is closed.
Speak anywhereon Windows.
Text on your clipboard.
Nothing leaves your machine. Free, one executable, no installer. Windows 10 and 11.
Your voice never leaves this machine.
Whisper.cpp runs the transcription where you recorded it: on your computer.
- Network requests
- One optional model download, only when you press Get.
- Storage
- Plain files under %LOCALAPPDATA%\talk2dee\, yours to delete.
- Transcription
- Local whisper.cpp, including supported CUDA builds.
Know what it is doing without opening a window.
The visual state follows the work. Red records. Honey transcribes. The toast confirms the copy.
Recording
Esc cancels
Listening
The tray icon and orb turn red. The orb pulses with your voice.
1 / 3
A tray app when you need it. Gone when you do not.
Current recording, history, dictionary, and model screens from the Windows app.
02
Every transcript stays findable.
Search the last 200 transcripts, pin the useful ones, filter them, or copy any line again.
03
Teach it the words that matter.
Dictionary packs keep correction rules and vocabulary together, with a tester before anything reaches your clipboard.
04
Choose the model yourself.
Five curated models show language, size, accuracy, and speed before you download. Custom ggml models work too.
The useful parts stay close.
Dictate, correct, find, and paste without handing the work to another service.
One hotkey, two ways to use it.
Press Ctrl+Alt+Space to start and stop, or hold the chord for push-to-talk for up to 3 minutes. Release to finish. Esc cancels without saving or copying anything.
English and Arabic belong in the same app.
Choose English, Arabic, or auto-detect. Arabic is right-to-left everywhere, and English-only models cannot be selected by mistake.
History you control.
Pin, tag, search, filter, copy again, or delete. The last 200 transcripts live in a local SQLite file.
Fix names once.
Named dictionary packs hold regex-capable corrections and vocabulary terms. Enable only the packs you need.
Pick the engine yourself.
Five curated models, sha256-verified downloads, or any ggml model you place in the folder. CUDA builds work too. The same engine is available through t2dcli.
Before the first recording.
Talk2dee is one executable. It expects two free command-line tools and one Whisper model.
- 01Download v0.5.0
Windows 10 or 11, x64
Download the executable and put it wherever you keep portable apps. There is no installer.
- 02
FFmpeg on PATH
Install the free audio tool with winget.
winget install yt-dlp.FFmpeg - 03
whisper-cli on PATH
Install whisper.cpp, then download a model inside Talk2dee by pressing Get.
scoop install whisper-cpp
Good to know before downloading.
The short answers, including the limits.
Is talk2dee really free?
Yes. It is free, open on GitHub, and ships as one executable with no installer or subscription.
Does it work offline?
Yes. Recording and transcription are entirely local through whisper.cpp. The only network request is an optional model download, made only when you press Get.
Does it run on Mac or Linux?
No. The current app is for Windows 10 and 11 on x64 machines.
Which languages are supported?
English, Arabic, and automatic detection with a multilingual Whisper model. Arabic is rendered right-to-left throughout the app.
Where is my data?
Under %LOCALAPPDATA%\talk2dee\ as plain config, SQLite history, and a log. Delete that folder and the local data is gone.
Free. Local. Windows only.