Open HariKrishna-Vydana opened 6 years ago
Greedy decoding? You can simply search for the sequence of peaks in the NN output. Or you could create a fake ARPA file format LM that has unity transition probabilities for all words, like a grammar?
In:
https://github.com/srvk/eesen/blob/tf_clean/tf/ctc-am/test.py
You have --compute_ter that can give you the token error rate (without language model)
is there a way to decode without considering the influence of language model