whisper-large-spjallromur-v2

whisper-large-spjallromur-v2

Icelandic ASR model. Fine-tuned from language-and-voice-lab/whisper-large-icelandic-62640-steps-967h (pretrained on ~967h of Icelandic, including Samrómur Milljón) on palli23/spjallromur-v2, a conversational Icelandic speech corpus for the journal article "Spjallrómur: A Dual-Channel Corpus of Conversational Icelandic with Mixing Pipeline and ASR Baselines".

Training

• Train split packed into <30s segments (0.3s silence between utterances); dev/test evaluated per-utterance, no packing.
• Training labels: annotation tags (<...>, [...]) stripped before use.
• 8 epochs, fp16.

Evaluation

Repetition-controlled generation (repetition_penalty=1.2, no_repeat_ngram_size=3, max_new_tokens=225), on palli23/spjallromur-v2 test split (3,450 utt., individually, no concatenation). Text lowercased and punctuation-stripped identically for predictions and references before scoring. Bootstrap std, n=1000 resamples, seed 42.

Raw predictions were also scanned for literal bracketed tokens (e.g. <unk>) as a training-contamination check: 1/3,450 utterances contained one, and it was Whisper's own generation control tokens (<|pl|>, <|notimestamps|>) leaking into decoded text once, not a transcription artifact.