Voice Box

Often while reading long documents, technical articles, or forum threads on screen, my eyes get really tired and it becomes difficult to focus.

I thought the natural solution was a text-to-speech extension so I could listen instead of reading. But when I evaluated existing tools, I realized almost all of them upload your data to external servers — including the text you highlight, the websites you visit, and your reading habits. Not knowing what happens to that data, and finding other offline alternatives hit-or-miss, I decided to build one suited for my needs.

My Requirements Were Simple

Conversion should happen completely locally.

No data should be collected.

Multiple natural voices supported.

And I was able to achieve that. The extension runs two local engines directly in the browser: Kokoro and Piper. Depending on your machine's hardware and whether WebGPU is available, it uses Kokoro for expressive, neural-quality speech or Piper for fast, low-overhead synthesis.

Both engines support 12 distinct male and female voices across American and British accents. I also built in instant sample previews, so you only download the voice models you actually want.

The experience is completely friction-free: you highlight text, click Read Aloud, and it begins reading immediately. It synchronizes highlighting sentence-by-sentence on the page so you never lose your place. Inside the side panel, you can adjust reading speed, click any sentence to scrub back and forth, collapse into a mini-player, or replay instantly from cache.

Use of AI in This

We used AI to build this, and it was a great learning experience.

Instead of just having AI write raw code, the AI actually interviewed me about what I wanted to build. It challenged my reasoning and spotted several blind spots I had completely overlooked, like fallback handling for hardware acceleration and tab-switching behavior.

Most importantly, it helped me learn how to break a project down into smaller, manageable sections. In the past, I usually jumped straight into coding without a clear plan. Here, the AI taught me the discipline of vertical slicing — building one small, verifiable layer at a time. It gave me both a useful private tool and a much stronger engineering workflow.