Reading models

[jvel07]

Hi, this is one of the best tools for reading ark and scp files! However, wanted to know how can one read models generated by kaldi, e.g. fullUBM models or ivector extractor models, which are usually generated by kaldi with 'ubm' extensions. For example: final.ubm, or final.dubm.

[kamo-naoyuki]

Sorry, basically, this tool is aimed at just loading matrix binary files.

Actually, ivector-model is not implemented in python, then, what do you mean loading model file? If you mean that just parsing the parameters in the model file into numpy object or float array, then it's not impossible, but, even if so, we may not support it.

[jvel07]

Thanks for your reply, @kamo-naoyuki . Yes, I meant loading them into a numpy, I intend to use the models generated for a specific experiment in computational paralinguistics. Now I understand kaldiio is not supporting this. So, in the case it's not impossible, are you aware of the way one can read this? or maybe point out someone or some blog that can help, perhaps? Would be really helpful since I am searching for this for several days already. Sorry for the inconvenience.

[kamo-naoyuki]

How about pykaldi?https://pykaldi.github.io/api/kaldi.ivector.html#kaldi.ivector.IvectorExtractor

Direct way is also not hard. If you don't need the python API to read the model, how about directly using kaldi library in c++ and write parameters to your desired format? Kaldi's code is really simple, so it's not hard to understand to read the source code.

https://github.com/kaldi-asr/kaldi/blob/a2573871ba185b8fd83ec5e66270a9a2301e4300/src/ivector/ivector-extractor.cc#L710-L730

It's also not hard to translate this code to python, if you are familiar to both language.

[kamo-naoyuki]

Sorry, I mistakenly pointed, here is ivector-extractor read: https://github.com/kaldi-asr/kaldi/blob/a2573871ba185b8fd83ec5e66270a9a2301e4300/src/ivector/ivector-extractor.cc#L828-L849

[jvel07]

Thank you so much, @kamo-naoyuki. I will try this out and see. :)