Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: distilbert/distilroberta-base
|
| 4 |
+
pipeline_tag: token-classification
|
| 5 |
+
tags: [dictation, disfluency-removal, punctuation, onnx, zero-decode, typurr]
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
# typurr-edit-tagger — cleanup with NO text generation
|
| 9 |
+
|
| 10 |
+
Most "AI text cleanup" pays an autoregressive language model to retype your
|
| 11 |
+
whole sentence around three fixes. This model doesn't generate at all: **one
|
| 12 |
+
encoder pass** tags every word — *keep / delete*, *punctuation-after*,
|
| 13 |
+
*capitalize* — and reconstruction is pure string ops.
|
| 14 |
+
|
| 15 |
+
- **~150 ms on CPU** (fixed-shape ONNX, 96 tokens, runs in
|
| 16 |
+
[tract](https://github.com/sonos/tract) — no GPU, no llama.cpp)
|
| 17 |
+
- keep/delete **99.6%**, punctuation **97.3%**, capitalization **98.3%**
|
| 18 |
+
(held-out token accuracy)
|
| 19 |
+
- handles fillers ("um", "you know"), repeats, overwritten corrections,
|
| 20 |
+
casing, and sentence punctuation; spoken-number formatting is routed to
|
| 21 |
+
Typurr's generative tiers instead
|
| 22 |
+
|
| 23 |
+
| in | out |
|
| 24 |
+
|---|---|
|
| 25 |
+
| "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." |
|
| 26 |
+
|
| 27 |
+
Files: `tagger.onnx` (fp32, [1,96] fixed shape, inputs `input_ids`/`attention_mask`
|
| 28 |
+
int64, outputs `keep`/`punct`/`cap` logits) + `tokenizer.json` (roberta BPE,
|
| 29 |
+
prefix-space). Labels ride each word's first subtoken.
|
| 30 |
+
|
| 31 |
+
---
|
| 32 |
+
|
| 33 |
+
## Typurr — speak, it types. Nothing leaves your machine.
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
A Windows dictation app and voice assistant that runs **entirely on your own
|
| 38 |
+
hardware**: hold a hotkey, talk, release — finished text lands at your cursor
|
| 39 |
+
in any app. No account, no telemetry, no audio in anyone's cloud.
|
| 40 |
+
|
| 41 |
+
- **Instant finish** — its models clean your speech *while you talk*; the text
|
| 42 |
+
is ready the moment you release the key
|
| 43 |
+
- **Speaks & listens** — neural voice read-backs, review-before-send by voice,
|
| 44 |
+
"typurr do…" compound commands, wake word
|
| 45 |
+
- **Learns you** — your vocabulary, your corrections, your style; all in plain
|
| 46 |
+
files on your disk
|
| 47 |
+
- **Gives AI agents a voice** — local MCP server: your agents can speak, ask
|
| 48 |
+
you questions aloud, and type at your cursor
|
| 49 |
+
|
| 50 |
+
**Get it:** [typurr.com](https://typurr.com) ·
|
| 51 |
+
[GitHub](https://github.com/typurrapp/typurr) ·
|
| 52 |
+
`scoop install https://raw.githubusercontent.com/typurrapp/typurr/main/typurr.json`
|