Understanding CTC params

DeepSpeech
#1

I’m trying to debug a regression (in our code) in terms of the behaviors of the log-probabilities on characters between 0.6 an 0.7. It’s very likely something I’ve done wrong but it would be super helpful if someone was able to give me a quick run down on these params from CTC Decode

image
image504×804 45.8 KB

To ask in text, form, in case someone is using a screen reader. What are these parameters? Is there any documentation on what they represent ?

  log_prob_b_prev = -NUM_FLT_INF;
  log_prob_nb_prev = -NUM_FLT_INF;
  log_prob_b_cur = -NUM_FLT_INF;
  log_prob_nb_cur = -NUM_FLT_INF;
  log_prob_c = -NUM_FLT_INF;

Which one is closest to the actual logit value of the final layer in the NN?

#2

I believe log_prob_c is closest to the actual logit value.
I traced down to “ctc_beam_search.cpp” and noticed that “log_prob_c” is derived directly from the probability distribution vector of a time slice (frame). And for computing efficiency the probability is applied the log function.

But I wonder what are log_prob_b_prev and log_prob_nb_prev are ? what does the nb abbr. stand for ? Many thanks

#3

b = blank, nb = non blank

1 Like
#4

thanks a million :slight_smile:
will come back when i got some good results
peace