Request
Add NB-Whisper, the Norwegian Whisper fine-tunes from the National Library of Norway, as whisper variants.
- Upstream repos:
NbAiLab/nb-whisper-{tiny,base,small,medium,large} (Apache-2.0)
- Family key:
whisper (existing)
- Variant names:
nb-whisper-tiny … nb-whisper-large
- Architecture pattern: encoder-decoder, unchanged from stock Whisper
This comes from a discussion on Handy, where @cjpais asked for the models to be quantized through transcribe.cpp and tracked here.
Upfront
I don't understand all of the technical detail here, and I used AI (Claude Code) for the conversion work and the measurements below. I did run the models myself, on my own machine, and the improvement is real for me and not only in the numbers. I wanted to be honest about that split rather than present this as more than it is.
My testing
I dropped the Norwegian model into Handy as a custom model and used it for normal dictation. It is a clear improvement.
The biggest single thing: Handy used to often detect my Norwegian as Danish or Swedish and transcribe the whole thing in the wrong language. That stopped. Compound words hold together better, and I am correcting the output far less than before. Before this I was running a much larger model and still fixing text afterwards.
Conversion results
Four of five sizes converted and quantized through scripts/convert-whisper.py + transcribe-quantize.
WER on 8 clips from the Fleurs nb_no test split (the benchmark NB-Whisper's own paper uses), 175 reference words, greedy decoding, -l no, CPU backend. This is a small sample, so treat it as a check that the conversion is faithful rather than a real benchmark:
| Model (Q5_K_M) |
Download |
Measured WER |
Paper |
| NB-Whisper Tiny |
44 MB |
12.57% |
15.2% |
| NB-Whisper Base |
64 MB |
9.14% |
11.5% |
| NB-Whisper Small |
194 MB |
8.00% |
8.3% |
| NB-Whisper Medium |
583 MB |
7.43% |
7.2% |
| Stock Whisper Large v3 |
1.16 GB |
11.43% |
10.4% |
| Stock Whisper Small |
194 MB |
30.29% |
29.6% |
Per-quant for small: F32 8.00, F16 8.00, Q8_0 8.00, Q6_K 7.43, Q4_K_M 7.43. No quant level degrades.
Other checks:
ctest -R whisper green with TRANSCRIBE_WHISPER_GGUF pointed at the converted model.- GGUF metadata correct:
general.name "NB-Whisper Small", author/organization NbAiLab, pinned general.repo_url, license apache-2.0.
- All five small quant file sizes come out byte-identical to the shipped stock whisper-small entries, as expected for identical architecture.
Pinned revisions: tiny 8b38492d, base 2ab372b6, small e9bb5cb8, medium 0ed074d5, large 8c6249fd. SHA-256s for the 24 produced files available if useful, though they are reproducible from the commands above.
Patch needed
convert-whisper.py rejects unknown slugs via VARIANT_DISPLAY_NAMES. Five lines, following the existing breeze-asr-25 precedent for a community fine-tune. grep -rn breeze over the repo returns only that one line, so nothing else needs registering:
"breeze-asr-25": "Breeze-ASR-25",
+ "nb-whisper-tiny": "NB-Whisper Tiny",
+ "nb-whisper-base": "NB-Whisper Base",
+ "nb-whisper-small": "NB-Whisper Small",
+ "nb-whisper-medium": "NB-Whisper Medium",
+ "nb-whisper-large": "NB-Whisper Large",
Happy to open this as a PR if you want it that way.
nb-whisper-large is blocked on sharded safetensors
NbAiLab/nb-whisper-large ships model-00001-of-00002.safetensors (4.99 GB) + model-00002-of-00002.safetensors (1.18 GB) + model.safetensors.index.json, with no single model.safetensors. convert-whisper.py hardcodes model_dir / "model.safetensors" (line 486) and fails the existence check, so large cannot be converted as-is.
The pattern that would fix it already exists in this repo: convert-voxtral.py (~line 84) reads the index's weight_map when present and falls back to the single file otherwise. convert-granite, convert-medasr, convert-moss and convert-qwen3_asr do the same. I have not attempted that change, since it is converter code rather than a table entry and seems like your call.
Minor observations
general.languages carries all 99 Whisper languages, inherited from the tokenizer, though these are Norwegian models. The HF card metadata says no/nb/nn/en, which is probably what a catalog entry should use.scripts/lib/hf_source.py::download_snapshot passes no allow_patterns, so it pulled 6.1 GB for nb-whisper-small (that repo ships safetensors + pytorch + TF + Flax + ct2/ + onnx/). Converting from a hand-picked 281 MB subset worked identically. Left out of the patch since it is shared by about a dozen converters.
What I can offer
I am happy to help maintain this where I realistically can, which is testing Norwegian and Nynorsk output and reporting problems. I am probably not the right person to fix converter internals, so I would rather say that now than promise more than I can deliver.
Request
Add NB-Whisper, the Norwegian Whisper fine-tunes from the National Library of Norway, as whisper variants.
NbAiLab/nb-whisper-{tiny,base,small,medium,large}(Apache-2.0)whisper(existing)nb-whisper-tiny…nb-whisper-largeThis comes from a discussion on Handy, where @cjpais asked for the models to be quantized through transcribe.cpp and tracked here.
Upfront
I don't understand all of the technical detail here, and I used AI (Claude Code) for the conversion work and the measurements below. I did run the models myself, on my own machine, and the improvement is real for me and not only in the numbers. I wanted to be honest about that split rather than present this as more than it is.
My testing
I dropped the Norwegian model into Handy as a custom model and used it for normal dictation. It is a clear improvement.
The biggest single thing: Handy used to often detect my Norwegian as Danish or Swedish and transcribe the whole thing in the wrong language. That stopped. Compound words hold together better, and I am correcting the output far less than before. Before this I was running a much larger model and still fixing text afterwards.
Conversion results
Four of five sizes converted and quantized through
scripts/convert-whisper.py+transcribe-quantize.WER on 8 clips from the Fleurs
nb_notest split (the benchmark NB-Whisper's own paper uses), 175 reference words, greedy decoding,-l no, CPU backend. This is a small sample, so treat it as a check that the conversion is faithful rather than a real benchmark:Per-quant for small: F32 8.00, F16 8.00, Q8_0 8.00, Q6_K 7.43, Q4_K_M 7.43. No quant level degrades.
Other checks:
ctest -R whispergreen withTRANSCRIBE_WHISPER_GGUFpointed at the converted model.general.name"NB-Whisper Small", author/organization NbAiLab, pinnedgeneral.repo_url, license apache-2.0.Pinned revisions: tiny
8b38492d, base2ab372b6, smalle9bb5cb8, medium0ed074d5, large8c6249fd. SHA-256s for the 24 produced files available if useful, though they are reproducible from the commands above.Patch needed
convert-whisper.pyrejects unknown slugs viaVARIANT_DISPLAY_NAMES. Five lines, following the existingbreeze-asr-25precedent for a community fine-tune.grep -rn breezeover the repo returns only that one line, so nothing else needs registering:Happy to open this as a PR if you want it that way.
nb-whisper-large is blocked on sharded safetensors
NbAiLab/nb-whisper-largeshipsmodel-00001-of-00002.safetensors(4.99 GB) +model-00002-of-00002.safetensors(1.18 GB) +model.safetensors.index.json, with no singlemodel.safetensors.convert-whisper.pyhardcodesmodel_dir / "model.safetensors"(line 486) and fails the existence check, so large cannot be converted as-is.The pattern that would fix it already exists in this repo:
convert-voxtral.py(~line 84) reads the index'sweight_mapwhen present and falls back to the single file otherwise. convert-granite, convert-medasr, convert-moss and convert-qwen3_asr do the same. I have not attempted that change, since it is converter code rather than a table entry and seems like your call.Minor observations
general.languagescarries all 99 Whisper languages, inherited from the tokenizer, though these are Norwegian models. The HF card metadata saysno/nb/nn/en, which is probably what a catalog entry should use.scripts/lib/hf_source.py::download_snapshotpasses noallow_patterns, so it pulled 6.1 GB for nb-whisper-small (that repo ships safetensors + pytorch + TF + Flax + ct2/ + onnx/). Converting from a hand-picked 281 MB subset worked identically. Left out of the patch since it is shared by about a dozen converters.What I can offer
I am happy to help maintain this where I realistically can, which is testing Norwegian and Nynorsk output and reporting problems. I am probably not the right person to fix converter internals, so I would rather say that now than promise more than I can deliver.