RyeAI/danish-asr-leaderboard

Open Danish ASR Leaderboard — Results

Benchmark results backing the Open Danish ASR Leaderboard — an open, reproducible comparison of Danish speech-to-text models.

Every model is transcribed and scored identically on the same five independent public Danish test sets, so the numbers compare directly: Word Error Rate (WER) and Character Error Rate (CER) — lower is better — plus speed. Open-weight Danish speech recognition models you can run yourself and hosted transcription APIs are both included.

Dansk tale-til-tekst benchmark

Åben benchmark for dansk tale til tekst. Sammenlign hvor præcist forskellige modeller til dansk talegenkendelse omsætter tale til tekst.

Alle modeller måles ens på de samme fem offentlige danske testsæt, så tallene kan sammenlignes direkte: ordfejlrate (WER) og tegnfejlrate (CER) — lavere er bedre — samt hastighed. Både åbne modeller, du selv kan køre lokalt, og kommercielle tale-til-tekst-API'er indgår.

Den aktuelle rangliste findes på leaderboardet.

Using this dataset

The results config (shown by default) has one row per evaluated model. Scores are WER / CER (%) — lower is better. Each model also has its own config exposing the raw, un-normalised transcriptions (reference vs hypothesis per utterance) for GPU-free re-scoring and error analysis.

Test sets

Column prefix	Dataset	Split	Domain
coral_conversation	CoRal-project/coral-v3 — conversation	test	Spontaneous conversation
coral_read_aloud	CoRal-project/coral-v3 — read_aloud	test	Read-aloud speech
ftspeech	alexandrainst/ftspeech	test_balanced	Parliamentary / broadcast
cv17_da	Common Voice 27.0 — da, 2,756 clips (the CV 25.0 test split)	test	Crowd-sourced read speech
fleurs_da	google/fleurs — da_dk	test	Read speech

Schema

results config — one row per model:

Column	Type	Description
model	string	Markdown link: [org/name](https://huggingface.co/org/name) for HF models, plain name for hosted APIs
params_b	float	Parameter count in billions from safetensors metadata; NaN for API models
access	string	open = open weights; proprietary = hosted or closed model
mean_wer	float	Macro-averaged WER (%) across the five core test sets
mean_cer	float	Macro-averaged CER (%) across the five core test sets
coral_conversation_wer	float|null	WER on CoRal v3 conversation
coral_read_aloud_wer	float|null	WER on CoRal v3 read-aloud
ftspeech_wer	float|null	WER on FTSpeech
cv17_da_wer	float|null	WER on Common Voice (Danish), 2,756-clip test set
fleurs_da_wer	float|null	WER on FLEURS (Danish)
coral_conversation_cer	float|null	CER on CoRal v3 conversation
coral_read_aloud_cer	float|null	CER on CoRal v3 read-aloud
ftspeech_cer	float|null	CER on FTSpeech
cv17_da_cer	float|null	CER on Common Voice (Danish), 2,756-clip test set
fleurs_da_cer	float|null	CER on FLEURS (Danish)
speed_x	float|null	Audio seconds / wall-clock second (higher = faster). Measured on one NVIDIA A100 80 GB at batch size 16; network-bound for API models. NaN if not measured.
submitted	string	ISO 8601 date the result was submitted (YYYY-MM-DD)

Per-model configs (outputs/<model-slug>) — one row per utterance: dataset, id, reference, hypothesis (raw, un-normalised).

Text normalisation

Applied identically to hypothesis and reference before scoring, so WER/CER reflect recognition errors rather than formatting:

1. Unicode NFKC — folds compatibility characters; optional acute accents (dér / der) are ignored while other Danish letters are preserved.
2. Exact word equivalents — selected historical and joined spellings (saa / så; idag, igår, ialt, iøvrigt, igang, afsted, tilstede and their spaced forms) are unified. Attested abbreviations (bl.a., ca, dvs., evt., f.eks., fx, inkl., ekskl., m.fl., m.m., mv., osv., pct.) and units (km, cm, kg, kr, numeric m, mm, m², km/h, km/t, m/s) match their spoken forms. Numeric mio. and mia. match the corresponding number of millions or billions.
3. Danish number canonicalisation — thousands separators are stripped (1.234 → 1234), while decimals retain a marker (3,14 / 3.14 → 3komma14). Trailing decimal zeroes are ignored (5.0 = 5.00); frequency units are separated from attached numerals and compound suffixes (2.400 GHz = 2.4 GHz), and % / procents become procent. Clock times (kl. 06:30 = klokken 6.30) and ratios (3:2 = 3 til 2) are canonicalised; numeric slashes become boundaries.
4. Lowercase and punctuation removal — apostrophes inside a word (det's) are preserved; other punctuation and symbols are removed.
5. Spelled-out numerals → digits — text2num folds spelling and spacing variants (otte og tredive and otteogtredive → 38; hundrede → 100). Bare en / et remain indefinite articles.
6. Numerals → words — standalone integers become Danish cardinal words via num2words (4 → fire, 24 → fireogtyve); embedded digits such as 1960'erne are left alone.
7. Hesitation filler removal and whitespace collapse.

A dash between two digits becomes a space in step 4 rather than being deleted, so 1-3 scores as two numbers. Deleting it glued the range into 13 → tretten, and the time range 10 00-11 00 into ti elleve nul — references no correct transcription could match, which also handed free matches to models that emit digits while penalising ones that spell the range out.

Known limitation. A bare three-digit separator (2.400) is treated as thousands; GHz context is needed for the decimal reading. Numeric slashes only separate numbers: 4/5 matches 4 5, but is not assumed to mean a fraction or a date. Ordinals written with a full stop are also left alone (3. plads → tre plads); converting them to tredje was measured and rejected as net-harmful, because most N. in these references are sentence-final cardinals rather than true ordinals. Bare mm also remains literal because it can be a hesitation sound; dotted m.m. is handled separately as med mere.

Danish hesitation fillers (øh, hmm, …) are removed, from reference and hypothesis alike, using the same pattern as alexandrainst/coral. Scoring them verbatim measured transcription convention rather than accuracy — a model trained to omit disfluencies was penalised for a style choice. Pass --no-filler-words to scripts/rescore.py to recover verbatim scoring from the saved raw outputs.

General Danish orthographic replacements (aa↔å, oe↔ø, ae↔æ) are not applied — the digraphs occur legitimately as letter sequences. The exact historical words in step 2 are the only exceptions. scripts/rescore.py can re-derive WER/CER from saved raw outputs without re-running inference.

Adding a model

There are two paths, depending on whether you have run the evaluation yourself:

• Request a model (we run it): open a GitHub issue with the model id, backend, and where to find it — we'll run it through the harness and add it.
• Submit a score (you ran it): run the harness from the GitHub repo and open a pull request with results/<model-slug>.json plus the raw outputs/<model-slug>/ transcriptions. On merge, CI publishes both here and updates the leaderboard automatically.

Whichever path a model arrives by, we re-evaluate it independently on our own hardware before publishing — to confirm the scores reproduce and catch configuration differences. Do not modify the normalisation or metrics; run the harness as-is so results stay comparable.

License

MIT — see LICENSE.