Datasets:
WAXAL sna pseudo-labels from the v2 punct model
- Teacher:
sulaimank/w2vbert-shona-sd2@aeb002fb32b2133edace0ced91c76dd343b0cf02 - Source:
google/WaxalNLPsna-unlabeled - Filter: confidence >= 0.9, non-empty. Repetition is not filtered — the speakers repeat themselves and a trigram limit discards correct transcripts.
- Kept: 81996 clips, 473.8 hours, 1.0-36.0s
The teacher is lowercase with punctuation, so these transcripts carry no case and no accented characters. Training on them alongside cased data gives a model contradictory signals about capitalisation.
Confidence is the mean max-softmax over non-blank CTC frames: comparable across clips labelled by this model, and not comparable to any other model's scale.
The corrected Phase 2 test clips are present in the source pool — all 892 matched at cosine >= 0.9991 — and so are present here.
- Downloads last month
- 19