good-tape/whisker-co-is-pnc-nano
nano-cohere-transcribe export of Icelandic Stage A on 15–30 s windows, then a full punctuation polish.
Content init is windows Stage A checkpoint-2500 (Málrómur val WER 4.92%). The PnC step continues from that checkpoint on silver orthography from listener-verified Samrómur 21.05 and Samrómur Children so <|pnc|> decode emits sentence .?!. Encoder is updated at 0.25× the decoder LR.
Do not load this repo with AutoModelForSpeechSeq2Seq. Use nano-cohere-transcribe:
from nano_cohere_transcribe import from_pretrained
model = from_pretrained("good-tape/whisker-co-is-pnc-nano", device="cuda")
print(model.transcribe(waveform, language="is"))
preprocessor.featurizer.window and preprocessor.featurizer.fb are copied from CohereLabs/cohere-transcribe-03-2026. tokenizer.model is restored from that same base.
Training
| Field | Value |
|---|---|
| Content init | Stage A windows checkpoint-2500 (Málrómur val WER 4.92%) |
| PnC step | Full FT, encoder 0.25×, 2000 steps |
| Train data | Samrómur-PnC (21.05 64% / Children 36%) |
| Selected step | 250 (lowest in-loop FLEURS is_is content WER, 13.3%) |
FLEURS Icelandic validation is 36 clips, so that 13.3% is a noisy gate.
Derived from CohereLabs/cohere-transcribe-03-2026; Apache 2.0.