Utter cleanup: Gemma 4 E4B, fine-tuned (ft5)
The cleanup model of Utter, a local voice dictation app. It takes what the speech recognizer (NVIDIA Parakeet) heard and returns the text the speaker meant to type:
- Punctuation and capitals nobody said.
- Spoken symbols and numbers written as people type them ("four x faster" -> "4× faster", "about two forty" -> "~240").
- Fillers and false starts removed.
- Corrections applied: "at 2, actually 3"; "scratch that" (the last sentence); "let's start over" (everything).
- Far-back edits: "actually, change the eggs to three boxes".
- Spoken lists as bullets, numbered for steps.
- Grammar fixed, in the speaker's own words.
Files
| File | What |
|---|---|
utter-cleanup-e4b-ft5.q4_0.gguf |
The model (current), Q4_0, for llama.cpp (b11535 or later); about 3.3 GB of GPU memory when the per-layer embeddings stay in system RAM |
adapter-ft5/ |
Its LoRA adapter (rank 16) on the base checkpoint, for merging or further training |
utter-cleanup-e4b-ft4.q4_0.gguf, adapter/ |
The previous version (ft4) |
How it must be prompted
The rules are in the weights, so the system prompt is short, and it must be exactly this:
Turn this dictation into the text the speaker meant to type: punctuated, formatted and corrected the way a careful writer would type it, in the speaker's own words.
The user message is the transcript after Dictated: , optionally preceded by context lines:
App: email
Before: "text before the cursor"
Lists: off
Dictionary: Utter, Nguyen
Dictated: <what the recognizer heard>
The reply is the cleaned text only. Thinking must be off (chat_template_kwargs: {"enable_thinking": false}).
Training
- Method: LoRA (rank 16, alpha 32, lr 2e-4, 1 epoch) on Google's unquantized QAT checkpoint of Gemma 4 E4B, then merged and quantized to Q4_0, the format the QAT was trained for.
- Data: 89,499 pairs of (recognizer output, intended text):
- Written side: real passages from Common Pile (Stack Exchange, GitHub, arXiv abstracts, Ubuntu IRC) and Wikipedia, plus lists from the same sources and LLM-written everyday lists.
- Spoken side: written by Gemma 4 26B-A4B in eight kinds (plain, spoken punctuation, disfluent, self-correction, retraction command, sentence restatement, grammar mistakes, far-back edit), checked by code, spoken by Kokoro TTS and recognized by Parakeet.
- Answer keys: fixed for typos and grammar, with each fix verified as grammar by ERRANT.
- Held out: no evaluation passage was trained on.
Evaluation
Ready to send: the share of dictations a careful writer would send exactly as typed. Gemma 4 26B-A4B grades each one against the intended text, with a rubric calibrated on human review.
Test set:
- 950 held-out realistic dictations: everyday messages, emails, notes and AI prompts, plus short real forum and chat posts.
- Each is spoken one way: plainly, with fillers, a self-correction, a command, a later edit, spoken punctuation or broken grammar.
| ft4 | ft5 (this) | |
|---|---|---|
| Ready to send, Kokoro voices | 83.8% | 88.8% |
| Ready to send, everyday half only (583) | 88.7% | 96.2% |
| Ready to send, unseen Chatterbox voices | 74.7% | 80.7% |
Strict scores on the earlier sets, each cell no edit needed / characters to fix:
| Test set | Gemma 4 E4B, prompted | Gemma 4 26B-A4B, prompted | ft5 |
|---|---|---|---|
| Main (4309) | 7.2% / 18.0% | 11.3% / 7.7% | 18.8% / 3.5% |
| Same, unseen voices (4309) | 6.8% / 19.0% | 10.6% / 8.8% | 16.5% / 4.8% |
| Grammar mistakes (688) | 5.4% / 6.3% | 6.7% / 5.7% | 14.5% / 3.5% |
| Lists (897) | 2.7% / 11.4% | 3.2% / 11.7% | 47.4% / 3.6% |
| Far-back edits (1268) | 3.2% / 29.2% | 4.8% / 22.3% | 40.9% / 3.5% |
Known gaps:
- A leading filler with a comma ("Uh, so...") can survive.
- Pronoun case ("me and him went") is not always fixed.
- Speech in the evaluation is synthetic (two TTS engines), not recorded dictation.
License
Apache 2.0, as the base model. The training text came from openly licensed sources (public domain, CC0, CC BY-SA, MIT, Apache-2.0).
- Downloads last month
- 24
4-bit