typurr-edit-tagger — cleanup with NO text generation

Most "AI text cleanup" pays an autoregressive language model to retype your whole sentence around three fixes. This model doesn't generate at all: one encoder pass tags every word — keep / delete, punctuation-after, capitalize — and reconstruction is pure string ops.

• ~150 ms on CPU (fixed-shape ONNX, 96 tokens, runs in tract — no GPU, no llama.cpp)
• keep/delete 99.6%, punctuation 97.3%, capitalization 98.3% (held-out token accuracy)
• handles fillers ("um", "you know"), repeats, overwritten corrections, casing, and sentence punctuation; spoken-number formatting is routed to Typurr's generative tiers instead

in	out
"um so i think we should uh we should ship the roadmap to dana you know before the offsite"	"So I think we should ship the roadmap to Dana before the offsite."

Files: tagger.onnx (fp32, [1,96] fixed shape, inputs input_ids/attention_mask int64, outputs keep/punct/cap logits) + tokenizer.json (roberta BPE, prefix-space). Labels ride each word's first subtoken.

---

Typurr — speak, it types. Nothing leaves your machine.

A Windows dictation app and voice assistant that runs entirely on your own hardware: hold a hotkey, talk, release — finished text lands at your cursor in any app. No account, no telemetry, no audio in anyone's cloud.

• Instant finish — its models clean your speech while you talk; the text is ready the moment you release the key
• Speaks & listens — neural voice read-backs, review-before-send by voice, "typurr do…" compound commands, wake word
• Learns you — your vocabulary, your corrections, your style; all in plain files on your disk
• Gives AI agents a voice — local MCP server: your agents can speak, ask you questions aloud, and type at your cursor

Get it: typurr.com · GitHub · scoop install https://raw.githubusercontent.com/typurrapp/typurr/main/typurr.json