Florian Metze

The C++ implementation has all the code to do ASGD training, but there always was a small loss in final training accuracy, so we never fully cleaned up and released...

We would very much like to do that, and have rudimentary code to do that, but do not have capacity to do this right now. Would you like to give...

Hi, you have abnormal Objs even before (-1e+27, …), which indicates exploding gradients or similar problems. Are you using projection layers? They are more sensitive. If not, you could try...

There is no easy way to log the utterance id at training time, but the utterances are processed in the order in which they are in the feature file, so...

Yes, we are using the following script in some train_ctc_parallel.sh variant that we are using. It is quite a simple check. if $sort_by_len; then td=$(mktemp -d) feat-to-len scp:$data_tr/feats.scp ark,t:- |...

I don't know what nnet-am-info does, but by and large, the size of your model file expressed in "floats" (4 bytes) should be a good approximation to the number of...

Thanks, will see - I think Kaldi has also modified those scripts, and it may be useful to pull in their changes as well. Will have to check ...

We've looked at it, and it does not seem to be that straightforward. Baidu's code is only CTC (no LSTM) so one would need to put a few more things...

it is quite a bit slower, but typically faster than real-time for a single thread. however you can run recognition in a python interpreter, which is kind of cool.

We do have a multi-GPU implementation. Would some of you be available to help test it? We’ll need help in determining the best parameterization (when to average models, how many...