How to append or add the manual feature for building Speech recognition system

[Prashant Upadhyaya]

Hi All,

I have a question. Suppose we have extracted some feature using wavelet transform for input Speech signal in a similar manner as done for MFCC feature. 

The question is how can we include 39 feature extracted using dyadic wavelet for building Kaldi model for speech recognition. I was able to do this in HTK. But no idea about Kaldi.

I would we thankful if some suggestion can be provided to me. 

[Danijel Korzinek]

They way you would normally do it is to convert your feature set from whatever format you want, into Kaldi's ARK format.

Kaldi has two modes of reading ARK files: binary and text. To convert from one into another, simply use the copy-matrix program, eg:

copy-matrix ark:binary.ark ark,t:text.ark

(where binary.ark and text.ark are the corresponding files)

You can read more about the command line I/O in the documentation here: http://kaldi-asr.org/doc/io_tut.html

You should take a feature file generated by kaldi, convert it to text and see what it looks like. Then try to recreate the same format using your own features and either convert it into binary or read it as text in Kaldi.

To create a sample feature file, first create a simple SCP file with one wave file in it, eg:

test test.wav

Then process it with something like compute-mfcc:

compute-mfcc-feats wav.scp ark,t:features.ark

The features.ark file should be in text format using the command above.

Feel free to ask if you have any further questions...

[Prashant Upadhyaya]

Thank you Sir,

Once I will try this if I have some difficulties I will contact you. Thank you Sir.