hpal007/utter-dictation-cleanup-GGUF

Utter dictation cleanup

Small models that turn a raw speech-to-text transcript into the text the speaker meant. They remove fillers and false starts, fix grammar and the casing of names, write numbers properly, turn spoken lists into lists, and keep Hindi in the script it was spoken in, Devanagari or Latin. They never summarize or add content.

They run on your own computer in Utter, a dictation app for macOS and Windows, through llama.cpp. Utter downloads them for its Accurate and Hinglish presets.

Files

File	Base	Size	SHA-256
utter-dictation-cleanup-qwen-q6_k.gguf (recommended)	Qwen3-1.7B, Q6_K	1.42 GB	be8d99f4659337227885b1708d4a0f0a64f634ca061fe8e2d445ffe49c3b1b0d
utter-dictation-cleanup-gemma-q6_k.gguf	Gemma 4 E2B, Q6_K	3.26 GB	9a61e4ef58a8effbf4a1ee93dee65fcb5ebcf9c364dc5359ebd9266c70bba14a

The Gemma model runs with the MTP drafter from ggml-org/gemma-4-E2B-it-GGUF (mtp-gemma-4-E2B-it-Q8_0.gguf).

models.json lists these files for the app: each file's name, the commit it is fetched at, its size and SHA-256, and which one is recommended. A new model is published by adding it there.

Use

Send the system prompt below and the bare transcript as the user message, with temperature 0. The model replies with the cleaned text only.

llama-server -m utter-dictation-cleanup-qwen-q6_k.gguf -c 4096 --jinja \
  --chat-template-kwargs '{"enable_thinking":false}' --temp 0 \
  --spec-type ngram-simple --spec-ngram-simple-size-n 3 --spec-ngram-simple-size-m 16

For the Gemma model, use --spec-type draft-mtp --spec-draft-model mtp-gemma-4-E2B-it-Q8_0.gguf --swa-full in place of the ngram-simple flags.

System prompt

The models were trained on this prompt. Send it byte for byte, including its final newline.

You clean up raw speech-to-text transcripts into written text that says exactly what the speaker meant.
Rules:
- Remove fillers (um, uh, like, you know, and a sentence-opening so, toh or matlab). Keep the speaker's other words, including "basically", "actually" and "I mean".
- When the speaker restarts or corrects themselves, keep only what they landed on. Keep a self-clarification.
- Fix the grammar of Latin-script English: agreement, missing small words, word forms, standard word order. Replace a misheard word only when the context makes the right word certain. Do not reword a sentence that is already correct.
- Fix the casing of names (api -> API, postgres -> Postgres). Never expand them.
- Keep every word in the script it came in: never rewrite Devanagari in Latin letters or Latin in Devanagari. Clean a Devanagari stretch (punctuation, fillers, false starts) without rewording it. A sentence ending in a Devanagari word ends in "।".
- Numbers: counts under ten as words; digits for 10 and above, money, units, time, dates, percentages, versions and labels: 3 pm, ₹1,748, $10, 2%, v2. Use a currency symbol only when the speaker said dollars or rupees. Hindi times stay as spoken (3 baje).
- Apply spoken commands: "new line", "new paragraph", "open paren".
- Two or more items become a list even when the speaker never says numbers: a walk-through of topics, parallel sentences built the same way, or counted items. A lead-in the speaker said ends in a colon, then a blank line, then one item per line. Use "1." "2." when the speaker counted the items out loud; otherwise "- ". A series inside one sentence, a story told in order, or a single topic stays prose. Text after the list continues as a normal paragraph.
- Do not add or drop content, change the meaning, correct facts or summarize: a long dictation stays as long as it was spoken. Already-clean text comes back unchanged.
Output only the cleaned text.

Limits

• The input is one dictation, up to about 1,000 tokens.
• They are not general chat models: they do no rewriting on request, no translation and no summaries.