corticph / prefix-beam-search Goto Github PK

https://github.com/corticph/prefix-beam-search/blob/master/prefix_beam_search.py#L61 should be

Pnb[t][l] += ctc[t][c_ix] * Pb[t - 1][l]

Hi, thanks for this awesome tutorial! I was wondering, how could i integrate the end symbol ('>') in the algorithm if my CTC tables do not contain it? Would it be possible with the Language Model by predicting EOS? Another thing is, how to train the model with CTC and include the end symbol? Would ju just append each sentence with the end mark? Thanks!

Thank you for the implementation.

Is there any way to find the path the beam search uses for each final hypothesis?

For example, if the hypothesis was

"a loud laugh followed at chunkys expense"

-------a-- l-l-l--oo--ud---- etc. ...

A_prev is [' '] so it doesn't loop. Is this a bug?

Edit: I mean, nothing is added to my A_prev in the loop. And in the end A_prev is empty.

in function greedy_decoder, alphabet = list(ascii_lowercase) + [' ', '>']. But in function prefix_beam_search, alphabet = list(ascii_lowercase) + [' ', '>', '%']. i feel confused.

I have my corpus in plain text and language model in .arpa format generated from KenLM.
How can I input those to the algorithm?