g2p-seq2seq icon indicating copy to clipboard operation
g2p-seq2seq copied to clipboard

Published 20 hours ago • verified publisher icon cmusphinx

Confidence values for predictions?

Open olix20 opened this issue 6 years ago • 1 comments

At test time, some of the words could be quite new and radically different from the set of words used for training the model (think foreign/medical words). Also there could be samples in the training set which are mislabelled. Is there a way to identify such cases?

olix20 avatar Jun 11 '18 14:06 olix20

It's maybe a bit irrelevant to this topic but I was wondering if you could report confidences in the beam search result; that's useful there too.

Danial-Alh avatar Jul 06 '18 10:07 Danial-Alh