good-tape/whisker-co-is-stage-c-long-nano
nano-cohere-transcribe export of the Icelandic Stage C long-form checkpoint (checkpoint-1000).
Full fine-tune from Stage B (checkpoint-1512) on continuous 10–30 s windows (Spjallrómur conversation + joined Alþingi segments) while keeping Málrómur / Samrómur read speech in the mix. Punctuation/capitalisation LoRA is not included — add that in a later step if needed.
Use with nano-cohere-transcribe:
from nano_cohere_transcribe import from_pretrained
model = from_pretrained("good-tape/whisker-co-is-stage-c-long-nano", device="cuda")
print(model.transcribe(waveform, language="is"))
For Hugging Face transformers loading, export a full repo separately (same weights as this nano bundle).
Gate metrics (beam 5)
| Benchmark | WER |
|---|---|
| Málrómur official validation | 7.9% |
FLEURS is_is validation | 13.6% |
FLEURS is_is test | 13.1% |
Stage B reference (same decode): Málrómur ~6.9% (512 in-train), FLEURS val 14.2%, test 16.0%.
Derived from CohereLabs/cohere-transcribe-03-2026; Apache 2.0.