Dictata: Local Whisper voice dictation for Windows, no data leaves machine

Dictata is a Windows native application written in Rust that enables system-wide voice dictation without uploading audio or text to external servers. The core workflow is simple: a user presses a global hotkey (default Ctrl+Alt+Space), speaks, presses again, and the transcribed text is automatically pasted into the active window. The transcription itself is handled by whisper.cpp, a local inference engine for Whisper. The tool supports three interaction modes: standard hotkey dictation (press-to-record), continuous streaming (text appears as the user speaks, pausing at detected silence), and file transcription (audio or video files in any ffmpeg-supported format). Optional voice-activity detection (VAD) reduces computational load and hallucinations by skipping silent sections, downloading a small ~2 MB model on first use.

For GPU acceleration, Dictata leverages Vulkan to support AMD, Intel, and NVIDIA graphics; alternatively, users can run transcription on CPU alone. The application offers multiple audio sources via cpal and WASAPI loopback: microphone, system audio, or a mix (useful for meeting transcription). A floating dock displays the audio waveform in real time and can be configured for size, opacity, and position. Output can be raw transcription or post-processed through a local OpenAI-compatible LLM using customizable prompts (for cleanup, email formatting, message formatting, list generation, etc.). Users can inject custom vocabulary and term replacements as the initial prompt to the LLM.

The application includes a built-in ggml model library that lets users download, delete, and receive hardware-aware recommendations; it also supports direct installation of any ggml .bin model from HuggingFace via URL, repository name, or keyword search. The user interface supports French, English, and Spanish, includes a dark theme, system tray integration, and a transcription history page. Configuration is stored in a config.json file next to the executable or in a directory specified by the DICTATA_HOME environment variable. The developer notes that the file must be UTF-8 without BOM.

The codebase is modular, with 64 unit tests covering configuration, resampling, mixing, output modes, settings logic, hardware detection, HuggingFace query parsing, and safety. The executable is self-contained, requiring only ffmpeg in PATH for file transcription and optionally the Vulkan SDK and Visual Studio Build Tools for GPU builds. The tool is licensed under MIT with Commons Clause, permitting free use, modification, distribution, and self-hosting for personal or non-commercial purposes, but prohibiting commercial resale or paid services derived substantially from the software.