Angel-Freak/edda-v0.2-ggml

Edda v0.2 for whisper.cpp (GGML, q5_0)

Edda v0.2 by Danish Foundation Models, a Danish fine-tune of Whisper large-v3-turbo, converted to GGML for whisper.cpp and quantized to q5_0. Nothing else was changed: same weights, same licence (Apache 2.0). All credit for the model goes to its authors; see the original model card for training data, evaluation and limitations.

Published for Fennec, a Danish dictation app for Linux, so it can download the model directly instead of converting it on each machine.

File	Size	SHA-256
edda-v0.2-q5_0.bin	574 MB	9bdfd79be6fad58b0a6f273526fe7e3fdc856bea7b31c19d032c8352d0bf0197

How it was made

From danish-foundation-models/edda-v0.2 at revision 98ec5eee028b111a19733890fe32ea65267b662b, with whisper.cpp 95ea8f9:

python models/convert-h5-to-ggml.py <edda-v0.2 checkout> <openai/whisper checkout> out/
whisper-quantize out/ggml-model.bin edda-v0.2-q5_0.bin q5_0

Use

whisper-cli -m edda-v0.2-q5_0.bin -l da -f audio.wav

As the original card notes, Edda v0.2 was trained with the same speaker's previous utterance as a prompt and handles it well (--prompt); Edda v0.1 did not. Measured in Fennec on held-out FLEURS Danish clips, prompting with the previous sentence kept its punctuation (transcripts with no punctuation at all fell from 24/60 to 1/60) at the same or lower WER.