Introduction

norwegian-parakeet just landed on Hugging Face: ahdrage/norwegian-parakeet (created 2026-11-04 ~21:12 UTC). It is a full fine-tune of moondream/parakeet-ultra (NVIDIA Parakeet TDT 0.6B v3) on roughly 1,770 hours of Norwegian Bokmål speech that the author filtered to open licences only, and it writes cased, punctuated text with greedy decoding — no external LM.

If you need a Nordic ASR checkpoint you can actually attribute and host, this is the clean zero-day card this cycle.

What Shipped

  • Vendor / product: norwegian-parakeet — ahdrage (sibling of norsk-asr)
  • Modality: open-weight automatic speech recognition (Norwegian Bokmål)
  • Open vs closed: weights CC BY 4.0; training/code repo Apache-2.0
  • Runnable today: Hugging Face transformers with ParakeetForTDT + AutoProcessor
  • Concrete fact: NPSC test WER 6.1% / CER 2.2% (5,518 clips, 11.0 h); NST val 4.2% WER; FLEURS Norwegian 7.4%; spontaneous NB Samtale held-out 17.7%

The card is unusually honest about floors: Stortinget and NorGovPCC transcripts are edited, so hearing hesitations correctly still looks like error. Tuva (3 speakers, 0.5 h) is tiny. Treat those caveats as part of the scoop, not fine print to skip.

Run It

pip install transformers torch torchaudio soundfile

import torch
from transformers import AutoProcessor, ParakeetForTDT

processor = AutoProcessor.from_pretrained("ahdrage/norwegian-parakeet")
model = ParakeetForTDT.from_pretrained("ahdrage/norwegian-parakeet").eval()
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")  # 16 kHz mono float32, up to ~30 s
with torch.no_grad():
    out = model.generate(**inputs)
print(processor.batch_decode(out.sequences, skip_special_tokens=True)[0])

Clip long recordings first — the eval path is built for ≤30 s utterances.

Notes for makers

  • Attribution chain matters: NVIDIA Parakeet / moondream, National Library / Språkbanken sources, Tuva, FLEURS — the repo’s docs/DATA.md lists licences and checksums
  • Two unresolved commercial notes stay on the card (NST mirror licence tagging; Stortinget terms) — read them before you ship a product
  • Four training passes on one 24 GB card (~35 h) with a custom memory-lean TDT loss; recipe is public on GitHub if you want to continue
  • Soft-hold neighbors (Ghana ZipVoice 401, Moore gated, finanfa Fon TTS NC+WIP) did not unlock this cycle — this Bokmål Parakeet is the ungated Nordic ASR scoop
  • No Gen CTA: language ASR does not map to the Flux / Grok / MiniMax-H3 / Seedream / Seedance Gen loop

Conclusion

For Norwegian makers who want an open, attributable Bokmål ASR model with real held-out numbers and a transformers path today, norwegian-parakeet is the zero-day drop — pull it, credit the data chain, and score on your own domain before you go live.

—Anabel ♡