How to pass new features to gmm-init-mono

[hyungwon yang]

Hi all.

I made a new feature set and this feature set is structured just like mfcc features.

e.g., [data ID]  [

          234.3424 12.3434 45.234324 ...

          11.2323 13.2323 16.2222 ... ]

I wonder how to put his feature set to gmm-init-mono function in 'train_mono.sh'

In train_mono.sh, gmm-init-mono receives $feats variable and it looked as follows.

feats="ark,s,cs:apply-cmvn $cmvn_opts --utt2spk=ark:$sdata/JOB/utt2spk scp:$sdata/JOB/c\

mvn.scp scp:$sdata/JOB/feats.scp ark:- | add-deltas ark:- ark:- |"

I don't want to use apply-cmvn function (because I didn't apply cmvn on my feature set) but just put this features directly to gmm-init-mono function.

I tried copy-feats in order to give it to the function but then I cannot separate those jobs into 4 or 8. 

I improvised as follows.

feats="ark:copy-feats ark:$data_dir/new_feats.ark ark:- |"

Could anyone tell me which function I have to use?

I want to use number of jobs option otherwise it takes too much time to be trained.

Thanks in advance.

[Danijel Korzinek]

To keep things simple, maybe just create a CMVN file set that has all zeros for mean and 1 for variance? Thus your features will be unaffected by CMVN.

On the other hand, you should consider some form of data normalization/standardization if you want to use the data with any kind of ML model.

[Daniel Povey]

The easiest way would be to format your data in the standard data-directory format (with feats.scp and so on), compute CMVN stats using steps/compute_cmvn_stats.sh, and then call the script with the option

--cmvn-opts "--norm-means=false" to disable cepstral mean normalization.  (Variance normalization is off by default).

Dan

--
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[hyungwon yang]

Hello Dan,

Wow. Why didn't I think of that!

Thank you so much. I will try it right away and I believe it will solve the problem.

2016년 11월 16일 수요일 오전 12시 48분 7초 UTC+9, hyungwon yang 님의 말:

[hyungwon yang]

Hello Danijel,

Your comment is also always helpful.

Thank you!

I finally solved the problem.

2016년 11월 16일 수요일 오전 12시 48분 7초 UTC+9, hyungwon yang 님의 말:

Hi all.