A Whisper dictation app built for everyday desktop writing
Secretary42 packages whisper.cpp, FFmpeg, model management and system-wide paste into a desktop app. You choose the local model; no Python runtime, cloud API key or transcription account is required.
From microphone audio to text
- The global shortcut starts and stops an explicit recording.
- Bundled FFmpeg converts the recording to 16 kHz mono audio.
- The selected quantized Whisper model runs locally through whisper.cpp.
- The transcript is pasted, copied, or saved according to your output setting.
Which Whisper model should you choose?
| Model | Approx. download | Best fit |
|---|---|---|
| Tiny English | 31 MB | Fast tests and lower-powered machines |
| Base English | 57 MB | Fast English dictation with improved accuracy |
| Small English | 182 MB | Recommended balance for everyday English writing |
| Medium multilingual | 515 MB | Higher accuracy and multilingual dictation |
| Large v3 multilingual | 1.5 GB | Maximum available accuracy when speed and storage permit |
These are approximate download sizes for the quantized models in the shipping registry. Actual speed depends on the computer, recording length, chosen language and model size.
Whisper without a developer setup
whisper.cpp is a command-line engine. Secretary42 adds the surrounding desktop workflow: microphone selection, permissions, global shortcuts, model downloads, local history, clipboard/file output, and cleanup of temporary audio. The engine and FFmpeg ship with the application, so users do not install Python or assemble a transcription pipeline.
Faithful transcription, not AI rewriting
Secretary42 returns the local transcription result. It does not send that result to a cloud language model for rewriting, remove ideas it considers unnecessary, or run agent commands. That keeps the processing boundary clear, but it also means you should expect to edit punctuation, names or domain-specific vocabulary when the speech model gets them wrong.
Start with the balanced model
Install Secretary42, download Small English, and compare it with a larger model using your own microphone and vocabulary.