SpeakoFlow Mini - Tiny local model that fixes dictation without rewriting it

Hi everyone! 👋 I make SpeakoFlow, a free offline voice assistant. This is the small model that cleans up what you dictate, and I trained it because nothing else would leave my text alone. Speech recognition is good now. Punctuation, capitals, fillers and numbers are already handled. So the only job left is fixing what you actually got wrong. Say "let's ship it Thursday, no, Friday" and you want "let's ship it Friday". Say something that was already fine and you want it back untouched, character for character. That second half is where everything I tried fell over. A capable model handed correct text will improve it. A comma becomes a full stop, a paragraph becomes bullets, a word becomes a smoother synonym. Each edit looks reasonable on its own and each one is wrong, because I said the other thing and now I am undoing it by hand. I did get it working with a large hosted model behind an API. That meant a network round trip and a bill every time someone dictated a sentence, for a feature that runs while you wait for text to appear. So I trained a 0.8B to do the narrow version of the job instead. It is 833 MB, Apache-2.0, and runs on an ordinary CPU with no internet. Already-correct text comes back untouched 92.6% of the time. Being straight about the limits: on the strange cases real dictation throws up, a big hosted model still reads the situation better than this one does. What this does is get close on the cases that keep happening, at a size you can ship inside an app. Full results, limitations and a one-line Ollama command are on the model page. If you dictate something and it changes a line you wanted kept, or misses one you wanted fixed, send me the text. That is the feedback I want most.