wav2vec2-large-spjallromur-v2

wav2vec2-large-spjallromur-v2

Icelandic ASR model. Fine-tuned from language-and-voice-lab/wav2vec2-large-xlsr-53-icelandic-ep30-967h (pretrained on ~967h of Icelandic, including Samrómur Milljón) on palli23/spjallromur-v2, a conversational Icelandic speech corpus for the journal paper "Spjallrómur: A Dual-Channel Corpus of Conversational Icelandic with Mixing Pipeline and ASR Baselines".

Training

• Base checkpoint's CTC vocabulary kept as-is (38 tokens: lowercase Icelandic letters, word-delimiter, [UNK], [PAD] — no capitals, no punctuation).
• Training labels normalized to match that vocabulary exactly: annotation tags (<...>, [...]) stripped, text lowercased, punctuation removed. An earlier training pass kept capitalization/punctuation in the labels while feeding a vocab that cannot represent either, which taught the model to emit [UNK] as a routine output class rather than a rare fallback (confirmed via post-hoc scanning of raw predictions: ~7% of test utterances contained a literal [UNK] token). This checkpoint was retrained with labels normalized to the vocabulary's actual character set, eliminating that failure mode entirely (0/3,450 test utterances contain [UNK] in this checkpoint's output).
• Feature encoder frozen for epoch 1, unfrozen after.
• Train split packed into <30s segments (0.3s silence between utterances); dev/test evaluated per-utterance, no packing.
• 8 epochs, batch size 16, gradient accumulation 2, lr 1e-5, fp16.

Evaluation

Greedy CTC decoding (no repetition control — not applicable to CTC), on palli23/spjallromur-v2 test split (3,450 utt., individually, no concatenation). Text normalized (lowercase, punctuation/tag-stripped) identically for predictions and references before scoring. Bootstrap std, n=1000 resamples, seed 42.