Martijn Stegink

I now dictate a large part of my work instead of typing it, because it is three to four times faster. But there is more. Speaking and thinking do not ask the same attention of you that typing and thinking do. While a sentence is still coming out of your mouth, part of your head is already forming the next one. You can talk as fast as you think, something typing never lets you do.

For years I have used voice commands to run my smart home, and that worked well. I wanted the same on my computer, but more than commands: integrated into everything I use, not just one app. Siri is fine for short, plain sentences, but falls back noticeably with accents, background noise, jargon or longer speech. Fireflies and Granola turned out to be built for meetings, they are great but not what I was looking for. Wispr Flow, Aqua Voice and SuperWhisper came closer and I liked them, but I ran into the limits of the free plan fast.

Then a colleague recommended Spokenly, which you can use for free, offers custom and local prompt cleaning and allows you to bring your own model. I decided to try it and copied his setup to my own machine, and it was slow as hell.

He runs it on a fast MacBook Pro with plenty of memory. I use a MacBook Air M2 with 8GB day to day, and an iMac M1 with 16GB at my desk. I figured memory was the difference. So I started testing myself instead of taking over his configuration.

I kept Parakeet for speech recognition and went through cleaning models through oMLX one by one to see what actually fits in 8GB. I tested them on the MacBook Air with 8GB and on the iMac M1 with 16GB, and the measured speed between those remains close: in daily use it is too slow. With more memory it will work better, but even with more memory I advise against local cleaning. The privacy argument rarely holds, because most of the data goes to the cloud anyway, and it costs memory that you probably bought for something else.

Whether the cleaning prompt could see which app was active made more difference than the model itself. It pushed the cycle time on cloud from 1.16 to 1.7 seconds. Half a second sounds like nothing, but it happens with every dictation, all day long. And it is a 47 percent increase on a baseline that was the whole point of dictating. On the local model that same setting made the prompt cache hit rate collapse from above 90 percent to zero, the moment I switched windows between dictations. One setting, and the entire speed advantage of dictating was gone.

So I decided to go with cloud cleaning.

I wrote my own cleaning prompt and improved it step by step, normalisation, name correction, abbreviations, injection protection, and I tested every part on its own, so that I knew which part was responsible the moment something went wrong. I also built a deterministic word replacement layer as an extra safety net for names, tested it thoroughly, and now advise against it.

The real problem was never the model or the prompt. It was how fast and how naturally I spoke. A relaxed, complete sentence almost always comes out of the cleaning well; a hurried or stumbling one does not, whatever model sits behind it and whatever cleaning layer you use. The temptation is to speak fast, because dictating already feels so much faster than typing that speed becomes the reflex. But even slow, calm speech is still much faster than typing. I type fast myself, and slow dictation still wins by a wide margin.

Names are one of the things none of the models are good at, cloud or local, so I now correct a name myself instead of looking for a setting that does it for me. Words that resemble each other run into the same wall. Claude and Cloud, for instance: no prompt keeps those two reliably apart.

I tested with different cloud models through Anthropic and OpenAI. The cheaper, smaller model I had picked as my default turned out, in daily use, to read the cleaning text itself as an instruction now and then instead of as text to clean up. Not often, but often enough to be a nuisance.

Talking to a computer also changes a room. On your own it is simply faster. With colleagues who dictate as well, nobody notices. With people who do not, one finds it annoying and another feels watched. And some people find it impressive that you can make a computer do what you want by talking to it. Certainly when you are steering AI with it that carries out complicated tasks for you. This does not appear to be happening only at my house. Offices seem to be going through the same switch.

Based on my own test rounds and the setup guide that came out of them.