Hello Everyone,
@kdavis @reuben
I wanted to know that every deep speech version model be it 0.4.1,0.5.1 or 0.6 have WER with the inclusion of the language model or without it?
All reported WER values are with the language model.
So if i get WER 0.31 and CER 0.08 with loss of 4 on test dataset without LM then is it considerably good?
Cause with 3 lacs dataset i was getting the same on test dataset and even after around 6 lacs dataset i am still getting the same WER.
Should i include my own language model on which i trained and then see the WER? Is that the right way?
@reuben Okay after reading through the forum regarding language model i came to know that by default it takes default binary and trie file. So basically i was testing my trained model which was done on 20k specific word on the general trie and lm. So i need to include that specific words in my language model and create trie and binary for proper evaluation. Am i right?