skolfus/kviskr-tegnsetting-no

Kviskr tegnsetting (Norwegian punctuation and casing)

Adds punctuation and capitalization to raw Norwegian speech-to-text output, without ever changing a word. Built for Kviskr, a macOS dictation app, where it runs on-device after NB-Whisper.

• Base: ltg/norbert3-xs, 15M parameters
• Size: 18.8 MB (INT8 ONNX), 20.1 MB including tokenizer
• Languages: Norwegian Bokmål and Nynorsk
• Replaces: skolfus/kviskr-punctuation-no (XLM-R base, 275 MiB)

How it works

The model predicts two labels for each whitespace-separated word: the punctuation mark that follows it, and its casing. The output is rebuilt from the original words and the original whitespace, so the text is preserved character for character.

Input	input_ids [B, T] int64, attention_mask [B, T] int64
Output	tegn [B, T, 7], stor [B, T, 5] (logits)
Tokenizer	ltg/norbert3-xs (WordPiece, 50,000)
Punctuation classes	none · , · . · ? · ! · : · ; (! disabled in calibration)
Casing classes	lower · Upper · UPPER · SEGMENT (TV-aksjonen) · BOTH (Nord-Norge)

Usage

1. Split the input on whitespace, but keep the whitespace itself. Strip trailing .,!?;: from each word and remember what you removed.
2. Feed the words lowercased, with is_split_into_words=True.
3. Read the logits on the first subword of each word.
4. Rebuild the text from the original words and whitespace. If a word already had a trailing mark, keep it and ignore the prediction for that word.
5. Change casing only when the change is reversible: every changed character must be the original's own upper or lower form, and the reverse operation must give back the original. Never touch a word that already has an internal capital (iPhone, NB-Whisper, USA).

Keep a guard that compares the word sequence before and after and falls back to the input if anything differs. With the rules above, it should never trigger.

Word preservation, measured: 0 violations across 3,833 texts (375 ASR clips with Ordrett input, 375 with Standard input, 3,000 text chunks and 83 dictations), under four definitions: strict, symmetric, whitespace-exact, and an independent character-level check.

Evaluation (13 Sep 2026)

All candidates receive the same ASR text; the difference is the punctuation layer alone. 95% paired bootstrap.

Benchmark	n	Previous model	This model	Paired difference
NB benchmark, NB-Whisper large verbatim	327 clips	58.2	63.7	+5.4 [1.2, 9.5]
Spontaneous speech (Storting)	60	47.3	58.1	+10.8 [3.1, 18.6]
Read news prose (FLEURS)	150	72.7	71.7	−0.9 [−6.0, 4.3]
Held-out Norwegian text (NCC), leak-free subset	2,484 chunks	61.4	75.5	+14.1 [12.8, 15.5]
Dictation text	83	77.6	91.1	+13.5 [4.7, 23.3]

Casing F1 on the NB benchmark: 77.5 → 81.1. Full-stop F1 on the text benchmark: 88.5 → 96.1. WER change: +0.0 [0.0, 0.1].

504 of the 3,000 text chunks share at least one 10-gram with the training data, so the leak-free subset above is the number to cite.

Training

Fine-tuned from ltg/norbert3-xs on 28.2M words of Norwegian text with human punctuation from NCC: Storting documents (13.8M), books (8.9M), Målfrid (3.9M) and Wikipedia (1.5M). 10,842 steps, batch 64, AdamW, cosine schedule. Decision thresholds calibrated per class on a held-out validation set.

Norwegian speech corpora cannot serve as ground truth for punctuation: NPSC has 0.0 commas per 1,000 words, NST 17.3, and NB Tale lacks punctuation in 48 of 49 clips in the National Library's evaluation set.

Limitations

• The two heads are independent and can disagree: 6.4% of predicted commas on ASR text are followed by a capital letter (reference: 1.6%).
• Slightly more commas than humans on real ASR text: 49.4 per 1,000 words against 43.1 in the reference.
• Internal capitals cannot be recovered from lowercase input: iphone becomes Iphone.
• One- and two-word dictations get the wrong end mark (hei → Hei?). Training chunks were 8–70 words, and 15% of real dictations are three words or fewer.

License

Apache 2.0, the same as ltg/norbert3-xs and the NCC sources it was trained on.