Stop paying for AI transcription! 🎙️ WritHer: 100% Local Voice Assistant for Windows. Privacy-first, Whisper + Ollama powered. Open Source on GitHub!

WritHer — Fri, 01 May 2026 09:09:47 +0000

Hey everyone! I wanted to share a small tool I’ve been building called WritHer.(Free and open source alternative to Wispr Flow)

The idea is simple: it lives in your system tray and gives you two things.

Hold AltGr anywhere (any app, any text field) and just speak. It transcribes your voice with Whisper and pastes the text right where your cursor is. No clicking, no switching apps.

Hold Ctrl+R and you get a voice assistant that understands natural language. You can say things like “remind me to call Marco in one hour” or “appointment with the dentist tomorrow at 3pm” and it handles the rest. Notes, to-do lists, shopping lists, reminders with toast notifications, all stored locally in SQLite.

The part I’m most proud of: everything runs 100% offline. Speech recognition via faster-whisper, intent parsing via Ollama, no cloud, no API keys, no telemetry. Once you download the models it works with no internet at all.

There’s also a little animated floating widget with eyes that react to what it’s doing (listening, thinking, error…) which is silly but I kind of love it.

It’s Python, MIT license, Windows 10/11 only for now.

GitHub: https://github.com/benmaster82/writher

https://getwrither.com

Would love feedback, especially from anyone who uses voice input regularly. Still early days but it works well for my daily workflow!

I Made Tkinter Look Like a Modern Glassmorphic App — Here's the Dark Magic I Used

WritHer — Thu, 26 Feb 2026 22:29:53 +0000

If you've ever built a desktop app in Python, you've probably used Tkinter. And if you have, you probably think it's doomed to look like a clunky, grey Windows 95 application.

I thought so too.

But recently, I needed a lightweight, floating UI for a 100% local, offline voice assistant I was building. I absolutely refused to bundle an entire Chromium instance (Electron) just to render a small widget.

So I decided to push Tkinter to its absolute limits. The result is Writher: an open-source, privacy-first voice assistant and dictation tool powered by faster-whisper and Ollama.

In this post, I'll break down the tricks I used to make a legacy Python GUI look modern, and the architecture behind a fully local AI desktop app.

🎨 The UI Hack: Making Tkinter Beautiful

To get that modern, glassmorphic floating pill shape with glowing borders, I completely bypassed Tkinter's standard widgets.

Instead, I used PIL (Pillow) to render high-resolution graphics dynamically on a transparent Tkinter Canvas.

The Trick: Borderless & Transparent Windows

# Remove the window frame entirely
root.overrideredirect(True)

# Chromakey hack: make a specific color fully transparent
root.wm_attributes("-transparentcolor", "#000001")

This gives you a frameless, floating window — the foundation for any modern-looking widget.

Dynamic Glow Rendering

Every frame of the animation (the bot's eyes changing expressions: listening, thinking, happy) is drawn on-the-fly using ImageDraw and ImageFilter.GaussianBlur to create a glowing effect that mimics SVG filters:

from PIL import Image, ImageDraw, ImageFilter

def draw_glow(size, color, blur_radius=10):
    img = Image.new("RGBA", size, (0, 0, 0, 0))
    draw = ImageDraw.Draw(img)
    draw.rounded_rectangle(
        [blur_radius, blur_radius,
         size[0] - blur_radius, size[1] - blur_radius],
        radius=20, fill=color
    )
    return img.filter(ImageFilter.GaussianBlur(blur_radius))

The Ghost Window

Since Writher is a dictation tool, clicking it shouldn't steal focus from your active app (like VSCode or a text editor). I used Win32 ctypes to apply the WS_EX_NOACTIVATE style:

import ctypes

GWL_EXSTYLE = -20
WS_EX_NOACTIVATE = 0x08000000
WS_EX_TOOLWINDOW = 0x00000080

def make_ghost_window(hwnd):
    style = ctypes.windll.user32.GetWindowLongW(hwnd, GWL_EXSTYLE)
    ctypes.windll.user32.SetWindowLongW(
        hwnd, GWL_EXSTYLE,
        style | WS_EX_NOACTIVATE | WS_EX_TOOLWINDOW
    )

💡 Pro tip: WS_EX_TOOLWINDOW also hides the window from the Alt+Tab menu, making it behave like a true desktop widget.

🧠 The Brain: 100% Local AI

I'm tired of sending my voice and private notes to the cloud. Writher operates entirely on your local machine.

Speech-to-Text — I used faster-whisper (CTranslate2). It runs flawlessly on CPU or GPU and transcribes voice in near real-time.

The LLM — I hooked the app to Ollama. Press Ctrl+R, Writher listens, transcribes, and passes the text to a local model for processing.

Function Calling — This is where it gets interesting. Instead of just chatting, I configured the LLM with tool definitions. It converts your voice commands into structured function calls — saving notes, scheduling appointments, or setting reminders in a local SQLite database:

# Thread-safe SQLite with WAL mode
conn = sqlite3.connect("writher.db")
conn.execute("PRAGMA journal_mode=WAL")

The LLM doesn't just understand you — it acts on what you say.

⚙️ The Muscle: Seamless OS Integration

Most dictation tools copy text to your clipboard and force you to paste manually. Writher acts like a phantom keyboard.

I initially used pyperclip, but it suffers from race conditions when other apps lock the clipboard. To make it bulletproof, I wrote a custom clipboard injector using the Win32 API (OpenClipboard, SetClipboardData).

The flow:

Save your current clipboard contents
Inject the transcribed text and simulate Ctrl+V via pynput
Restore your original clipboard — all in milliseconds

And just in case Windows blocks the clipboard? I implemented a fail-safe that automatically appends your dictation to a recovery_notes.txt file. You never lose a single word.

🔑 What I Learned

Building Writher taught me a few things I wasn't expecting:

You don't need Electron for everything. A transparent Tkinter canvas + Pillow can produce surprisingly polished UIs. The binary is tiny compared to any web-based alternative.
Local AI is ready. With faster-whisper + Ollama, you can build genuinely useful AI tools that never phone home. Privacy doesn't have to mean sacrificing quality.
Win32 APIs are your secret weapon on Windows. Ghost windows, clipboard control, focus management — they unlock capabilities that pure Python can't reach alone.

🚀 Try It Yourself

The whole project is open-source. You can check out the UI implementation, the AI pipeline, and run it on your own machine:

👉 Writher on GitHub

I'd love to hear your thoughts. Have you ever pushed a legacy GUI framework beyond what it was meant to do? What local LLM setup are you running? Drop a comment — I read all of them.

And if you find the project useful, a ⭐ on the repo helps more than you'd think. 🙏

DEV Community: WritHer