ParentFull threadhooloovoo_zoo·I think they’re held out but all data comes from the same speaker and since they have like 10k samples times say 10 words per sample, practically every word will have been in the training data.View on HN