Winpepper
https://github.com/obra/winpepper- Category
- AI Tools
- Rank
- No. 1527Tools index
- Pricing
- Open Source
- Type
- TOOL
- Builder
- obra
- GitHub
- 18 stars
- Latest release
- v0.7.4-alpha
- Date
About
Windows-native local dictation. Hold a hotkey, speak, release — cleaned-up words appear in the focused app. Parakeet TDT v3 ASR + LlamaSharp cleanup, all local.
What it does
Winpepper turns spoken notes into editable text inside ordinary Windows programs. It captures microphone audio, recognizes speech on the computer, optionally fixes punctuation and filler words, applies saved vocabulary corrections, then inserts the result at the active cursor. It also stores dictation history for review and reruns.
Why it's ranked here
This is a credible privacy-first desktop utility with unusually strong operational detail. Local processing, customizable vocabulary, resumable model downloads, diagnostics, and extensive automated and real-hardware testing support serious use. The alpha label, unsigned installer, hardware-sensitive cleanup latency, and substantial model footprint keep the verdict measured.
What's good
Audio and transcripts remain local unless the user explicitly enables AssemblyAI with a personal key. Streaming recognition reduces the delay after speaking, while raw transcription can bypass slower language-model cleanup. Corrections handle names and jargon. History and the Lab support comparison and reprocessing. Diagnostic bundles exclude audio and transcripts.
Tradeoffs
It requires 64-bit Windows 11 version 22H2 or newer, roughly 2 GB of storage, and an initial model download. The unsigned installer triggers a Windows warning. Streaming needs the Visual C++ runtime and retains about 1 GB of memory after first use. Cleanup can take several seconds on integrated graphics. The project explicitly calls itself alpha software.
How to use it well
Use it for frequent drafting into email, chat, documents, browsers, or code editors when local speech processing matters. Add recurring names and technical terms to corrections, keep streaming enabled for faster results, and disable cleanup when latency matters more than polish. It is not a cross-platform dictation service or a cloud transcription platform.
Technical notes+
README.md describes a single .NET 9 and WinUI 3 process using WASAPI capture, ONNX Runtime DirectML with CPU fallback, NVIDIA Parakeet TDT v3, Qwen 2.5 0.5B through LlamaSharp and Vulkan, and SendInput injection. docs/DEVELOPMENT.md separates UI, keyboard-hook, audio, ASR, cleanup, and optional window-context workers around an Idle → Recording → Transcribing → CleaningUp → Injecting state machine. It documents nine projects under tests/, including Windows-only coverage for WASAPI, UI Automation, OCR, model integration, and keyboard hooks. docs/release.md describes a self-contained win-x64 WiX v5 MSI, SHA-256 release artifacts, and an automated install, self-test, autostart verification, and uninstall gate. scripts/windows-sandbox/ supports disposable installation checks, while docs/windows-smoke-test.md defines stateful real-hardware verification.
Observed
- License
- Apache License 2.0
- Primary language and framework
- C# on .NET 9 with WinUI 3
- Packaging
- Self-contained 64-bit Windows MSI built with WiX v5 and installed per user
- Interface
- Windows desktop application with tray controls, global hotkeys, settings pages, and text insertion into focused applications
- Platform support
- 64-bit Windows 11 version 22H2 or newer
- Privacy mode
- Local speech and cleanup processing by default, with optional AssemblyAI support using the user's API key
- Test structure
- Nine test projects plus Windows Sandbox checks and a stateful real-hardware smoke-test procedure
Read from README.md, docs/release.md, docs/DEVELOPMENT.md, docs/manual-test.md, docs/automation-ids.md, docs/windows-smoke-test.md, docs/testing-windows-from-wsl.md, docs/plans/2026-07-21-pending-paste.md, docs/plans/2026-08-01-models-page-ux.md, docs/plans/2026-07-24-silence-trimming.md.
What it can do
Convert speech to text via hotkey activation
Spoken words while holding hotkey → Typed text in focused application
Process audio using local speech recognition
Audio from microphone → Raw transcribed text
Clean up and improve transcribed text
Raw speech-to-text output → Cleaned and formatted text
Insert text into any Windows application
Processed text from dictation → Text appearing in currently focused app
Perform offline speech recognition
Spoken audio → Text transcription without internet connection
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.