Dictata: Privacy-Focused Local Voice Dictation Software
Complete offline speech transcription with Whisper and LLM
1Fully Local Voice Dictation Software
Dictata is a Windows application enabling system-wide voice dictation. Activated by hotkey, it transcribes speech locally and pastes into the active app. Data never leaves the machine. Uses whisper.cpp as default transcription engine with optional ONNX backends (Parakeet, SenseVoice, Moonshine, Zipformer). Built in Rust for performance and security.
2Advanced Features and Flexibility
Offers user-friendly features: toggle or push-to-talk dictation modes, auto-paste, customizable floating dock with waveform. Continuous (streaming) mode inserts text as you speak. VAD (Voice Activity Detection) reduces compute load and hallucinations during silence. Downloads a small (~2MB) model on first use, configurable in settings. Supports Vulkan for GPU acceleration across AMD/Intel/NVIDIA hardware.
3Multi-Backend Engine Architecture
The core engine automatically selects backend based on model format (ggml .bin → whisper.cpp, ONNX bundle → matching backend). Optional ONNX backends can be enabled at build time via Cargo features. whisper-rs with vulkan support enables GPU transcription. Each ONNX backend links statically via ort, requiring no external DLLs. Backends include NVIDIA Parakeet, SenseVoice CTC, Moonshine, and Zipformer transducer.
4Input/Output Options and Processing
Supports microphone, system audio (WASAPI loopback), or mixed input (meeting mode). Output options include raw transcription or LLM post-processing (cleanup, email formatting, messaging style, lists) via local OpenAI-compatible API with customizable prompts. Can transcribe any audio/video file format supported by ffmpeg. Custom vocabulary and replacements can be injected as initial prompt.
5Model Management System
Built-in ggml model library handles downloads, deletions, and hardware-aware recommendations (detects GPU/VRAM on Vulkan, CPU/RAM otherwise). Provides one-click verified ONNX bundles (Parakeet, SenseVoice, Zipformer, Moonshine). HuggingFace integration allows installing any ggml model or ONNX bundle via direct URL, repo (owner/name), or keyword search. Maintains transcription history and supports multiple languages.
6Developer and Build Details
Open-source Rust project requiring Vulkan SDK for GPU support. Includes workaround for Windows MAX_PATH limitation during build. CPU-only builds use --no-default-features --features whisper flags. ONNX backends are optional build features. PowerShell scripts automate testing, building, and packaging (producing self-contained dictata-version-windows-x86_64[-cpu].zip). Binary embeds icon and version metadata via winresource.
7User Experience Features
Offers French/English/Spanish UI with dark theme and system tray integration. Runs as background process with transcription history. Requires Windows 10/11 (Linux support limited - see LINUX.md). File transcription needs ffmpeg in PATH. GPU builds need Visual Studio Build Tools (CMake + Ninja). Configuration file (config.json) is generated on first run with default settings, avoiding personal data in distributed builds.
8Privacy-Centric Design Philosophy
Dictata's key advantage is complete local processing - no data ever leaves the user's machine. This addresses privacy concerns for individuals and enterprises handling sensitive information. Open-source nature enables transparency and trust. Ideal for industries with strict data processing regulations. Self-contained binary with no external dependencies enhances security and simplifies deployment in controlled environments.
Original article: Hacker News: Show HN — Dictata: Yerel ve Gizlilik Odaklı Sesli Yazılımı