SpeakoFlow Mini: a local dictation cleanup model that knows when to change nothing

SpeakoFlow Mini is out. It is an 833 MB model that takes a dictated transcript, applies the correction the speaker actually made, and leaves everything else exactly as it was said. It runs offline on a CPU, it is Apache-2.0, and you can pull it from Hugging Face today.

The reason it exists is narrower than "clean up dictation". A current recogniser already punctuates, capitalises, drops filler words and formats numbers. Say "the deposit is three hundred dollars, um, and the meeting is at nine thirty" into SpeakoFlow with cleanup switched off and you get "The deposit is $300 and the meeting is at 9:30" back, with no language model in the loop. What is left over is small, and it is all judgement.

What does SpeakoFlow Mini actually do?

It handles the residue a recogniser cannot reason about. You corrected yourself mid-sentence. You said an email address out loud. You asked for a new paragraph. You said a word that came through the recogniser perfectly and was still the wrong word. Those need a decision about what you meant, and a pattern cannot make one.

The fourth row is the whole point. Nothing was wrong with it, so nothing changed, down to the character. Fifteen categories are trained and scored, from retractions and spoken symbols through to empty input, truncated input and recogniser hallucination loops. Every one of them is a decision, and in several of them the correct decision is to do nothing.

Why is knowing when not to edit the hard part?

Hand a capable language model a sentence that is already correct and it will find something to improve. A comma becomes a full stop. A sentence becomes a heading. A paragraph becomes bullets. A word becomes a smoother synonym.

Each of those edits is defensible on its own, and each one is wrong, because you said the other thing and now you have to undo the improvement. Undoing an unwanted edit costs more than making the edit by hand would have. A cleanup feature that does this is worse than no cleanup feature.

So the evaluation scores two things separately. Restraint is the share of already-correct transcripts returned untouched. Edit accuracy is the share of transcripts needing a change that came back exactly right, on an exact string match with no partial credit. Overall is the unweighted mean of the two.

The mean is unweighted for one reason: the evaluation set is close to evenly split between transcripts that need an edit and transcripts that do not, so a plain accuracy figure would score a program that changes nothing at 45.3%, which reads like it half works. The mean scores that same do-nothing strategy at exactly 50.0%. Anything under 50 is worse than leaving the transcript alone.

What are the numbers?

Held-out evaluation, temperature 0, every system given the same short system prompt with reasoning off. Restraint and edit accuracy are the two scored axes; content damage is the share of transcripts that lost a word the speaker said, reported separately because no penalty weight for it would be anything but arbitrary.

Read the untuned base row and the Gemma row together. Both look respectable on restraint, at 89.7% and 79.4%, and both score below doing nothing. They alter text they should have left alone and then fail to make the change that was actually needed. Restraint on its own is not competence, which is exactly why the overall score is never printed here without both halves beside it.

The strongest result on the page is not the comparison against anything hosted. It is the controlled one. Against the identical untuned Qwen3.5-0.8B at the same file size, quantisation, prompt and hardware, edit accuracy went from 4.9% to 48.8%, a gain of 23.4 points overall with a 95% confidence interval of +16.3 to +30.3. Restraint went from 89.7% to 92.6%, a difference of a single case that this set cannot resolve.

The usual expectation is that teaching a model to edit costs restraint. Here it cost nothing measurable.

How do you run it?

It is a GGUF, so anything that loads GGUF loads it. Send the transcript as the user message with nothing added: no instructions, no delimiters, no examples. The system prompt was fine-tuned into the model and is printed in full on the model card.

1. llama.cpp, straight from the Hub. llama-cli -hf SpeakoFlow/speakoflow-mini:Q8_0, or llama-server for an OpenAI-compatible endpoint on port 8080.
2. Ollama. ollama run hf.co/SpeakoFlow/speakoflow-mini
3. LM Studio. Search speakoflow in the app, or lms get SpeakoFlow/speakoflow-mini@Q8_0.
4. Two settings are not optional. Do not cap max_tokens, because a truncated long transcript counts as lost content. Disable thinking, because any reasoning text in the output fails every already-correct case for reasons that have nothing to do with cleanup. At its defaults this model emits zero reasoning characters.

Five quantisations are published and every one was run through the same evaluation, so the ladder states its own cost rather than leaving you to guess.

Take Q8_0. The spread down the ladder is smaller than this evaluation can resolve, so Q4_K_M being 1.3 points behind means very little. What is a count rather than an estimate is that 7.3% of its outputs differ from Q8_0, so it demonstrably changes what the model says without demonstrably changing how often it is right.

How does it compare to a hosted frontier model?

Overall, it is a tie, and I want to be precise about that. The gap over GPT-5.6 Luna is 5.8 points with a 95% confidence interval of -1.5 to +12.9. That interval contains zero. Gaps under about six points are not resolvable by this evaluation at all.

What is resolved is length. On passages of 500 words and up the gap is +20.9 points, interval +7.0 to +35.6. Long passages are what people actually dictate, so that is the result I care about, but it is one band and not the headline.

So the honest claim is narrow and worth stating exactly: under one fixed short prompt with no reasoning budget, a 0.8B model running locally matches a hosted frontier model on this task. Not beats. Matches, and the confidence interval says the same thing.

The mechanism behind that is the same one behind the fine-tuning gain. Mini has the convention in its weights. Anything hosted has to be told the convention in a prompt on every single call and then has to follow it. Training on a written specification beats prompting for it, and that is a smaller and more defensible finding than beating a frontier model.

The constraint is also the product, not a handicap picked to flatter the result. Cleanup runs while somebody waits for text to appear in the box they are typing into. A longer prompt costs tokens on every call, a reasoning budget costs seconds, and both cost a network round trip.

What does it get wrong?

10.7% of transcripts come back missing a word the speaker said. Losing content is the failure mode I care most about and the one I would most like to see somebody beat, which is why the figure is on the model card rather than in a footnote.

It works on text, so if the recogniser heard the wrong word and nothing in the sentence gives that away, this model cannot recover it. It is English-only: the specification, the training and every number are English, and the only non-English behaviour taught or scored is leaving it alone under an absolute rule never to translate. The base tokenizer covers languages this work does not, so it will produce output for non-English input that nothing here evaluates.

Deterministic fixes are out of scope on purpose. Filler removal, punctuation, capitalisation and number formatting are not trained into it, because the rules layer already does them and a rule is faster and auditable. That layer is published separately under MIT, in Python and TypeScript, with a specification, pattern tables and a conformance suite. The two-stage design has its own write-up, including why the ordering matters.

And the evaluation shares an author with the model, which is a real conflict. The mitigations are structural rather than reassuring: the scoring is deterministic with no model in the loop, the judge experiment that would have flattered this model was discarded and reported, and the sensitivity analysis that puts a competitor ahead is published because it is true.

Frequently asked questions