Rewrites raw dictation into clean, punctuated written text

You paste in a raw transcript, the kind every speech-to-text tool produces: lowercase, unpunctuated, full of ums and self-corrections. The model rewrites it as text you would actually send. "so um i need to send the the report by uh friday no wait make that thursday" comes back as "I need to send the report by Thursday." It keeps the value you landed on when you correct yourself, renders spoken numbers, dates, prices, and email addresses in written form, and applies punctuation and capitalization.

A short control line at the top of the input steers the result: four registers from casual to formal, a list mode that turns real enumerations into bullets, and an email mode that lays out a greeting, body, and sign-off. It does exactly one job; it is not a chat model and will not follow general instructions.

The whole thing is 0.6 billion parameters, small enough that the compressed build is a 462 MB file running comfortably on a laptop CPU. This first release covers English only.

Open the transcript-cleanup demo on Hugging Face and paste real dictation, the messier the better. For a first test, say a sentence where you change your mind halfway, like a date you correct a moment later, and check which value survives; resolving self-corrections is the part ordinary cleanup scripts cannot do. If it earns a place in your workflow, ready-made builds run offline in common desktop tools like Ollama and LM Studio.

It is strict about its input format: the exact system prompt and control line it was trained on are required, and skipping either can produce garbled output. It also only speaks English in this release, so transcripts in other languages will not work.

• Clean up dictated notes before pasting them into a document or chat.
• Dictate a whole email and let email mode shape the greeting and sign-off.
• Turn spoken figures into written ones: "twenty three thousand four hundred and fifty dollars" becomes "$23,450".
• Batch-polish transcripts from a recorder or meeting tool into readable minutes.
• Add a cleanup pass to your own dictation app; the license is permissive with a naming clause, so read it first.

Try the demo

View model page

Read this issue on neuralexpedition.com →