svale-600M

Danish speech recognition. nvidia/parakeet-tdt-0.6b-v3 fine-tuned on 2,850 h of public Danish speech (CoRal-v3, FTSpeech, Common Voice, FLEURS, YODAS; train splits only). Lowercase, no punctuation. GPU.

WER, Danish ASR leaderboard normaliser:

CoRal conversation	CoRal read-aloud	FTSpeech	Common Voice	FLEURS	mean
21.58	15.53	7.33	8.79	10.34	12.71

Use

from huggingface_hub import hf_hub_download
import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.restore_from(hf_hub_download("3dio-ai/svale-600M", "svale-600M.nemo"))
print(model.transcribe(["audio.wav"]))

NeMo 2.1+, 16 kHz mono.

Licence

NVIDIA Open Model License. CoRal OpenRAIL-D use restrictions apply: no speech synthesis, no biometric identification.