Reading target (alignment) files

[giampierosalvi]

Hi, I was trying to use kaldi_io to import alignment files, but I could not find out how to do it, and if it's possible.

I ran the timit recipe and ended up with a number of ali.<n>.gz files for example under the exp/mono_ali/ directory. I can convert those files from transition model IDs to PDF IDs with the command (for example): ali-to-pdf exp/mono_ali/final.mdl "ark:gunzip -c exp/mono_ali/ali.1.gz|" ark,t:mono_ali.1.pdf.txt The resulting file contains a line for each utterance, with utterance ID (for example faem0_si1392) and then a list of integer identifiers of the states in the model for each frame in that utterance.

• Is there a way to import this file into python using kaldi_io?
• Is there a way to pipe the ali-to-pdf command when opening the ali.1.gz file, so that I don't need to run it separately?

Thank you! Giampiero

[mnabihali]

Dear Giampiero Salvi; Did you find a solution to this problem, because I face the same problem now.

[KarelVesely84]

Hi, the solution is: kaldi_io.read_ali_ark('ali-to-pdf exp/mono_ali/final.mdl "ark:gunzip -c exp/mono_ali/ali.1.gz|" ark:- |')

This will convert the alignment to PDF-id format and dump it in binary format to be read by your python on-the-fly...

If you already dumped the 'ASCII-text' alingments, this is your command: kaldi_io.read_ali_ark('mono_ali.1.pdf.txt')

Let me know, if you experience any problems... Karel

[KarelVesely84]

Those will return generators that return (key, value) tuples...

[mnabihali]

Dear Karel;

Thanks for your reply and your help. I will try and if I have any issues I will back to you if you don't mind.

Thanks a lot

On Tue, Feb 25, 2020 at 7:35 PM Karel Vesely [email protected] wrote:

Those will return generators that return (key, value) tuples...

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/vesis84/kaldi-io-for-python/issues/36?email_source=notifications&email_token=ANWBGBHFWZQ3PLVDQIK3XSLREVQG3A5CNFSM4I4WCOUKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEM47ROY#issuecomment-591001787, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANWBGBBEVJBVN4PNFL4ZRETREVQG3ANCNFSM4I4WCOUA .

-- Mohamed Nabih Ali *Assistant **Lecturer * *Faculty of Computers and IT * Egyptian E-Learning University Ain Shams Center *Mail: [email protected] [email protected] * Mobile: +201285659213

Work: 02-33318417

[csukuangfj]

@mnabihali

Maybe you are interested in kaldi pybind for the next generation kaldi.

We have wrapped kaldi io to Python. For example, to read alignment information, you can use kaldi.read_vec_int.

You can find an example usage here: https://github.com/kaldi-asr/kaldi/blob/pybind11/src/pybind/tests/test_io_util.py#L21

Note that it is on the pybind11 branch instead of the master branch.

[mnabihali]

Thanks all, but another problem I want to read the kaldi featuers (feat.scp) and apply cmvn, utt2speak and delta features how can I do this with kaldi_io

[KarelVesely84]

Hi, you can do this by reading from a pipeline with "apply-cmvn" and "add-deltas" binaries in it. Look into GMM training scripts to see how the 'pipeline-string' looks like. Then you use this as filename when reading with 'kaldi-io-for-python'. Make sure that the 'pipeline-string' has '|' at the end.

[mnabihali]

Thanks Karel But when I run the full timit kaldi recipe I generate the fmllr features but its cmvn.ark files is not generated do you know why this happens?

On Thu, Mar 19, 2020, 12:10 Karel Vesely [email protected] wrote:

Hi, you can do this by reading from a pipeline with "apply-cmvn" and "add-deltas" binaries in it. Look into GMM training scripts to see how the 'pipeline-string' looks like. Then you use this as filename when reading with 'kaldi-io-for-python'. Make sure that the 'pipeline-string' has '|' at the end.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/vesis84/kaldi-io-for-python/issues/36#issuecomment-601121429, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANWBGBCPG5LUFAJ4ZS4Z7XTRIH4TLANCNFSM4I4WCOUA .

[KarelVesely84]

Hi, this happens because 'fmllr' features are already normalized from earlier stage of processing. So we don't do CMVN on top of them... (actually the 'power' of cmvn is already contained in 'fmllr', both are affine transforms of feature spaces: cmvn has a diagonal matrix, fmllr has a full matrix. however the way these two are estimated is very different...) K.