The model that cleans your speech while you're still talking.
Typurr feeds it clauses from a streaming recognizer as you speak; it strips
fillers, resolves self-corrections, punctuates, capitalizes, and formats
spoken numbers — so the finished text exists before you release the key.
in (spoken)
out
"um so basically can you send me the uh the report"
"Can you send me the report"
"it costs ten dollars, uh, per seat and dana no wait marcus should get the invoice"
"It costs $10 per seat and Marcus should get the invoice."
Numbers
metric (held-out dictation)
base Qwen3.5-0.8B
this model
exact match
0.0%
90.7%
filler leaks
63%
0%
content lost
2.7%
0%
~150 ms per clause on a laptop GPU (q8_0, llama.cpp). Trained on 58.8k
synthetic (mangled-speech → clean-text) pairs from Typurr's data engine.
Coupling note: tuned to Typurr's exact system prompt
(prompts.rs::SKETCH_RULES); with other prompts it still works, worse.
Inside Typurr it runs behind a fidelity audit — outputs that invent names or
numbers are caught and redone by a larger model.
Typurr — speak, it types. Nothing leaves your machine.
Typurr demo
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
