I built Blabb because
I wanted to use it.

I'm Trevor. By trade I'm a data and cloud engineer, and by habit I'm Blabb's heaviest user — a hundred-plus dictations a day, straight into the locked-down RDP sessions and Teams chats where I work. I didn't build Blabb to have something to build. I built it to use — and nothing I tried did the two things I needed.

Hear me accurately

The first was accuracy. I'd used a few speech tools and ASR models — OpenAI's Whisper among them — and none of them felt accurate enough for real work. So I auditioned openly licensed speech models against the two benchmarks my house actually runs: dictating over the robot vacuum by day, and dictating quietly at night so I don't wake the family. The model that topped the Open ASR Leaderboard passed both — so that's the one that ships.

Clean up the blabbering

The second was cleanup. My problem isn't volume — I don't talk all that much — it's that I blab. I talk at length without pausing, and I can blab on for days. What comes out of my mouth isn't what I want on the page. That's where a language model earns its place: take the rambling transcription and hand back clean text.

When the model tries too hard

But language models have a habit of being too helpful. They hallucinate. They add context I never gave. They answer a question I only mentioned, or carry out an instruction I only stated — instead of just typing my words. That was never what I wanted. So Blabb treats everything you say as text to clean, never a task to perform, and a validation layer compares the output to what you actually said. If the model ever does something other than type your words, the rewrite is thrown out and your original goes through.

Built for the limits, not against them

Models have hard limits too — token limits, technical limits — and those limits surface in the output as errors: a dropped sentence, a number that's off, text that trails away mid-thought. I built safeguards for those. Some are still in place; some I've been able to switch off, because IBM Granite 4.1 has gotten that good. What it delivers is faithful to the original transcription and reliable, and the guards that remain are the ones still earning their keep.

Built so I can't lose your data

Everything runs on your PC — and honestly, that's self-interest too. I'm one person; I have no desire to run servers full of other people's words, secure them, or be capable of losing them. The best way to protect your data is to never have it — so Blabb is built so that I can't.

Why I notice this stuff

I'm not a data scientist, but I follow where machine learning is going — and you don't need a PhD to do the useful part: test the inputs and the outputs, and probe for the failure modes you've read about. Hallucinations among them; I've run into one, and tested for it. A data and engineering background makes you a little paranoid in the right ways. You look at a result, ask what could go wrong before it does, then build the check that catches it. That's most of what Blabb is.

The subscription is what it looks like: if Blabb earns a place in your day, shout me a coffee and I'll keep making it better. That's why the trial comes first — I'd rather earn it. And if anything's off, tell me: support@blabb.io. I read everything.

— Trevor    dictated this instead of typing it, obviously