AriaType - Local first & Privacy first voice input built on the top of Tauri and llama.cpp

#ai #programming #productivity #rust

The Backstory

This Spring Festival, I was working on a web coding project. About 3 days in, I had an epiphany:

The biggest challenge in web coding isn't the cost of API calls. It's my stamina.
The constant cycle of writing prompts, communicating with AI models, and iterating corrections — that mental load is exhausting.

So I decided to build myself a voice input tool. Not because other options were bad, but because I wanted something that felt native.

I looked at TypeLess. The subscription fee was steep — almost matching my AI Sub plan. I wasn't willing to pay that much for something I could build for myself.
So I did.

Two hours. MVP.

LLaMA CPP + Whisper. Rough around the edges, but functional. Good enough for personal use.

Then the feedback loop started.

After showing it to colleagues post-holiday, I got real feedback.

Some of it stung, but all of it was valuable. So I spent the following weekends and holidays polishing it.

Today, AriaType 0.1 is officially released.

What AriaType does:

Local STT models — runs whisper-based models on your machine. For English, I recommend Vesper. For Chinese/CJK users, Sense Voice (by Alibaba) has noticeably better accuracy and speed
Polishing — local small models (under 2B params) for grammar correction, filler word removal, and formatting
Cloud service mode — use your own AI subscription for Polish. No separate payment required
Noise reduction & silence detection — skips silent audio chunks to save costs
100+ languages supported
Privacy by default — your voice data never leaves your machine unless you explicitly use cloud services

The tech stack:

I did zero "traditional" coding. Everything was AI-assisted:

Primary: GLM-5.1 and MiniMax-M2.7
Complex problem-solving: Claude OPUS 4.6 and ChatGPT 5.4 (used sparingly)

The real challenge no one talks about:

The gap between an AI-generated MVP and a real product is 80% of the work.
As the codebase grew, the challenge shifted from "can AI write this feature" to "can AI accurately modify and iteratively improve a complex, growing codebase without breaking things."

That's the engineering problem I now call harness engineering — building the systems and patterns that let AI reliably extend and maintain a project over time.

Why open source?

I enjoy the process. It's challenging, but the feedback loop is rewarding. I figured others might have similar needs — wanting local/offline STT, wanting their data to stay private.

So I registered a domain, built a website, and open sourced the project.