S1-ASR-Cleanup: cleaning up dictation with a fine-tuned S1-mini

Raw dictation is honest to a fault. It has every "gonna", every "the the", every "wait, where was I", and it arrives as one unbroken block. I clean my transcripts in more or less the same way every time, so I trained a small model to do it the way I would.

S1-ASR-Cleanup is a fine-tune of S1-mini by Superwhisper (superwhisper/s1-mini), a 0.6B Qwen3-based model that Superwhisper built for exactly this job. S1-mini was already a good starting point. What I wanted was for it to match my own house style, so I fine-tuned it on pairs of my raw transcripts and the versions I'd actually send.

What it does differently

• Paragraphs. Long dictation becomes short paragraphs, one thought each, with greetings and sign-offs on their own lines. Stock S1-mini mostly returns one block.

• Dates and money, consistently. "The twelfth of March" becomes March 12, "two thousand two hundred shekels" becomes 2,200 NIS, and "forty bucks" becomes $40.

• Informal contractions expanded. "Gonna", "gotta" and "shoulda" become "going to", "have to" and "should have", while ordinary contractions like "don't" stay.

• Spelled-out names dropped. "Ronen, that's R O N E N" becomes just "Ronen".

• Stray capitals fixed. The capital a recogniser leaves after a pause ("we could Use those") is lowercased, while real names stay capitalised.

• Self-corrections resolved to what I finally meant, and asides to people in the room left out. Romanised Hebrew words and tech names are kept as they are.

An example

Here's a held-out transcript, shortened, and what the model makes of it:

Hi Sharon just checking about the annual tax form deadline because Mas Hachnasah moved it to the twelfth of March instead of March the nineteenth so could you send me the updated PDF ... thanks.

Hi Sharon,

Just checking about the annual tax form deadline because Mas Hachnasah moved it to March 12 instead of March 19.

...

Thanks.

The model card has a fuller set of side-by-side examples against stock S1-mini, one per behaviour, and I didn't cherry-pick them beyond that. They also show where it still falls short: it sometimes leaves a run-on clause alone ("send me the updated PDF ping Guy on WhatsApp") that a person would have split.

Getting it

It's one standalone model with the LoRA merged in. There's a transformers build, a Q8_0 GGUF for llama.cpp, and a 397 MB Q4_K_M GGUF for phones. Use S1-mini's prompt format as given on the card, with thinking disabled and greedy decoding. It is licensed under Apache 2.0 with S1-mini's naming term, so anything that uses it has to credit S1-mini by Superwhisper. The training code is in ASR-Post-Processor-Model.

This is the two-step approach done properly, a speech model followed by a small text model. It's also the baseline for Proofed, my experiment in getting the speech model to do this cleanup itself, in one pass. It sits alongside my earlier paragraph-break model, which only adds breaks and never changes a word.