🎙️ VoskFlow Devlog — Offline Speech Intelligence Desktop App

Building VoskFlow — an entirely offline, privacy-first voice dictation engine with an Apple/VOID-inspired minimalist HUD that works across any Windows app.

Here’s everything built, debugged, and polished over the latest development sprint! 🚀

⚡ What We Built & Accomplished

1. 🌌 The “VOID” Minimalist HUD & Pill Widget

• Idle State: A discreet, 4px Warm Ochre (#DCA54C) dash docked right above the Windows taskbar.
• Active State: Hold the global push-to-talk hotkey (Ctrl + Space), and the dash smoothly expands into a floating pill with live symmetrical audio waveform visualizer bars.
• Hands-Free Auto Mode: Double-tap Ctrl to lock into hands-free streaming dictation.
• Zero Distraction: When released, it transcribes offline in milliseconds and types your text directly into whatever window/app has focus.

2. 🛠️ Frameless 6-Step Onboarding & Calibration Wizard

• Custom frameless Tauri + Webview2 window with dynamic particle starfield canvas.
• Hardware RAM auto-detection to automatically recommend the right offline model (tiny.en, base.en, small.en).
• Interactive Microphone Tuning & Voice Calibration step with live audio peak meters and immediate dictation verification.
• Seamless transition to the full system Dashboard upon completion.

3. 🧵 Bulletproof AI STT Worker & Windows Engine Architecture

• CTranslate2 & OpenMP Multi-threading Fix: Solved the infamous OpenMP CPU inference deadlock on Windows by isolating model initialization and transcription into a dedicated Python Queue worker loop.
• float32 Universal Compute Engine: Standardized compute format to ensure compatibility across all CPU architectures (Intel & AMD) without quantization crashes.
• Thread-Safe Qt Signal Dispatching: Re-architected background thread-to-GUI communication to eliminate GetDC (Invalid window handle) errors.

4. 🔕 Stealth Tray Engine & Lifecycle Management

• Persistent Background Daemon: Closing the Dashboard now cleanly minimizes VoskFlow to the system tray so speech dictation is always on standby.
• Auto-Cleanup for Stale Processes: Built automated socket & process-mutex cleanup to prevent orphan zombie backends when restarting or developing.
• Tray Bridge: Open/close Dashboard anytime directly from the Windows notification tray menu.

🧠 Tech Stack

• Frontend: Tauri v2, Rust, HTML5/CSS3 Canvas (Particle Physics), JetBrains Mono
• AI Core: Faster-Whisper (CTranslate2 offline speech models)
• Audio & System: PyAudio, NumPy, PyQt6 (Translucent GUI overlays), Pynput (Low-level global hotkeys)
• Platform: Windows 11 / 10

🔮 What’s Next

• Exportable custom voice shortcuts and automation macros.
• Local custom vocabulary & dictionary substitution rules.
• Packaging into a self-contained one-click .exe installer.

100% offline. Zero telemetry. Your voice stays on your machine. 🔒

🎙️ VoskFlow Devlog: Ready for V1 Ship! 🚀

Spent the last sprint squashing bugs and polishing VoskFlow (100% offline, privacy-first voice-to-text) for its V1 release!

What got done:

• 🛡️ True 100% Offline Mode: Patched faster-whisper from pinging HuggingFace API on startup. If local weights exist, it loads straight from disk—zero network required.
• 🎨 Midnight Study UI: Overhauled the web dashboard with our official palette (Espresso #1C1917 + Warm Ochre #DCA54C), serif typography, and a live status indicator.
• ⌨️ Bulletproof Push-to-Talk: Replaced basic key listeners with a low-level global hook (keyboard.hook), fixing combo hotkeys like Ctrl+Space without crashes.
• 📦 PyInstaller Monolith: Built a custom .spec bundling the models, CTranslate2 DLLs, and web UI into a clean, windowed .exe (no ugly terminal popup).

Next up: final .exe compilation and shipping! Don’t type, just speak. ✨

Finaaly trinaed a llm and connetced it

Today we overhauled the core engine for commercial-grade speed and accuracy. We migrated to Faster-Whisper, which gives us flawless auto-punctuation and capitalization. To achieve a zero-latency feel, we built a background worker that streams and processes audio chunks every 500ms while you hold the hotkey, meaning the final text drops instantly on release. Finally, we engineered a dynamic disfluency filter that automatically strips out filler words like umms and ahhs before typing. Up next: connecting a local LLM to turn spoken commands into OS-level keyboard shortcuts!

[Update] - Landing Page is Live fr 🚀

Ngl, the landing page we just dropped absolutely ate and left no crumbs. Added screenshots of our black glassmorphism pill and the bouncing voice bubbles, and the aesthetic is heavily based. Wispr Flow is literally shaking rn. No cap, the UI is serving looks. Now that the site is up, we’re ready to plug in the Vosk ML model and actually let it cook. Huge W. 🔥 https://evinjsubin.github.io/VoskFlow/landing/