some demo about Beam Search with LM about py-ctc-decode HOT 8 CLOSED

I extended the example in the Readme a bit.
https://github.com/ynop/py-ctc-decode#beam-search-with-lm

Hope it helps.
Otherwise just ask.

from py-ctc-decode.

Thanks for your reply!
my test logits' shape is (1, 131, 1224)(batch_size,time_step,V), the true result is a sentence with 20 words. But what I got from decode_batch has only one word.

from py-ctc-decode.

• Are you sure that the given sentence is predicted with the language model you use? Words not in the language model won't be predicted.
• Have you used ' ' as space and '_' as blank in the vocabulary?
• Have you used log probabilities?

Otherwise maybe try best-path to check if it works there.

from py-ctc-decode.

Thanks for your reply!

1. yes, I use the same corpus to train the language model and the acoustic model

2. '_' is in my vocabulary, but ' ' is not

3. log probabilities? Are you mean when compute the ctc-loss, transpose the logits with tf.log first? if so ,yes,I used it.

I have try the best-path method. what I got is a bunch of words without blank in them. Then, I add the blank in BestPathDecoder.decode[line15] : pred = ' '.join(pred).replace('_', '') ,I got the reasonable result.
Should I retrain my acoustic model with a new vocabulary( the version which adds ' ')

from py-ctc-decode.

The corpus I use is like this -->'zhe4 feng1 xin4 xie3 yu2 gong1 yuan2 yi1 liu4 wu3 si4 nian2 shi4 '
I would split the sentence and transform the words into ids, so, theoretically，space is not a part of the input data. In this case, can I use your【Beam Search with LM】 in some way?

from py-ctc-decode.

Hmm, the space is needed for since it is the point which triggers the language model.
If your input are word ids you would have to adapt the algorithm.

from py-ctc-decode.

I seem to understand. I will try these two methods below

1. retrain the acoustic model with new data (add a space id in the middle of each word id )
2. when use the 【Beam Search with LM】,add space to each symbol I got.
   Thank you for your patience. I'll reply this when I get results

from py-ctc-decode.

I have try the second method. Add a space before symbol when calculate value. It works.
Thanks a lot

from py-ctc-decode.