Typd

I’m building Typd: a local dictation app. You hold a key, talk, and finished, formatted text appears wherever your cursor is. It runs entirely on your computer in under 2 GB of VRAM or unified memory, works offline, and it’s free.

Why: today’s best dictation apps need the internet, send your voice to the cloud, and ask you to trust whoever’s on the other end with it. And they’re genuinely great. Wispr Flow takes the messy things you say out loud and turns them into clear, well-formatted text. I wanted that without the cloud.

My first attempt was a basic Wispr Flow clone. NVIDIA’s Parakeet (a 0.6B speech model) transcribed locally, then a free cloud API cleaned up the text: removing fillers, fixing structure. Only half of it was local. The transcript still left my machine, and everything depended on my connection. Running the cleanup LLM locally didn’t work either: it needed more VRAM than my laptop had, spilled into RAM, and got slow.

Then I asked: why do dictation apps use two models at all? Every one that explains how it works, Wispr Flow included, describes the same pipeline: a speech recogniser, then a fine-tuned LLM that removes the “um”s, works out what you meant and writes it properly. My guess is that it’s just history. Good speech models and good LLMs both already existed, so apps glued them together.

From first principles, though, the job is simple: audio goes in, finished text comes out. A second model only adds latency, memory and failure points. So why not one model that hears you and writes the final text directly?

That’s Typd. I took an open speech model (Qwen3-ASR) and trained it so it doesn’t just transcribe, it writes: fillers and false starts gone, self-corrections resolved, lists and paragraphs where they belong. All the training so far ran on free GPUs. In blind tests on 66 of my own real dictations (judged by AI judges that didn’t know which app wrote what), Typd was preferred over Wispr Flow 27 times to 12, with 27 ties. It returns text about 0.4 s after I stop talking.

What I’m working on next: structuring long, rambling speech better, and never losing something you said when you correct yourself mid-sentence. The hard part is that big LLMs are good at this because they know so much. A model that can explain anything from the height of the Eiffel Tower to photon spectra can work out what you meant. It’s like having a professor help you write a letter. But you can’t run a professor on a laptop. So Typd teaches a small model just that one skill, distilling what big models do for dictation into something that uses a sliver of your computer.

Typd runs on Windows and macOS, 100% private, no internet needed. It comes in two sizes: a 1.7B-parameter version (about 2 GB) for machines that can spare it, and a lighter 0.6B version (under 1 GB) for everything else.