Running Deepspeech 0.7.4 on Google Commands Dataset

Yes. If you don’t pass --test_files it would have skipped the test phase :slight_smile:

Great. It worked. It exported. Looks like it is giving blank inferences for my files in the test set though.

This is a peculiar dataset since the utterances are only 1 second long, I am curious if using a scorer is beneficial for this type of data or it might be better without.

Is having blank inferences usually an issue due to not training enough? Obviously 2 epochs of training won’t be enough for a good model but I’d hope to at least get some character predictions. I just want to make sure it isn’t something else that is going astray.

Yeah, with this few training it’s not really surprising you get blank inference

1 Like

Just to mention one thing which I noticed in many of the scripts you shared here: you can’t comment out a single line in a command of the form

command \
    param1 \
    param2 \
    param3

In Bash a comment goes from the # character until the end of the line. Using \ means you’re escaping the newline, so the entire thing is seen by bash as a single line. If you put a comment somewhere, then everything after that point gets commented out.

Thanks! That makes sense and helps a lot. Appreciate it.