Fbank features to Mfcc features

[srik...@zapr.in]

Hi,

   I am extracting Mel-filter bank filter from a wav file. I do some processing on the Fbank features to clean up the background music.

Now i have to convert the Fbank features to Mfcc before feeding it to the decoder.

Is there a script/code that takes Fbank features and generate Mfcc features....    

[Daniel Povey]

Not really, but it's not very difficult, basically applying a cosine transform, you can see the MFCC code for how to do it.

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
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+...@googlegroups.com.
To post to this group, send email to kaldi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/940ee1e3-9fc3-41a5-b83e-230413dcde95%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Srikanth Konjeti]

Thanks Dan... just checking before I do that

On Feb 2, 2019 11:29 PM, "Daniel Povey" <dpo...@gmail.com> wrote:

Not really, but it's not very difficult, basically applying a cosine transform, you can see the MFCC code for how to do it.

On Sat, Feb 2, 2019 at 9:00 AM <srik...@zapr.in> wrote:

Hi,

   I am extracting Mel-filter bank filter from a wav file. I do some processing on the Fbank features to clean up the background music.

Now i have to convert the Fbank features to Mfcc before feeding it to the decoder.

Is there a script/code that takes Fbank features and generate Mfcc features....    

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
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.

To post to this group, send email to kaldi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/940ee1e3-9fc3-41a5-b83e-230413dcde95%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

Go to http://kaldi-asr.org/forums.html find out how to join
---
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.

To post to this group, send email to kaldi...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAEWAuyR%3Daf-Lo4gUnwnn3cYaQ71TVpovbP9YL_r0YcMt92NYDw%40mail.gmail.com.

[feifei xiong]

sorry to re-active this post, since I have a similar question: 

I'd like to convert FBANK back to MFCC, and I did 'dct' from Matlab. However, when I compared the FBANK+dct with MFCC from Kaldi itself, the features are not exactly the same. How can I get the FBANK features exactly back to Kaldi-MFCC (which is used by my trained model)? 

Thanks!!!

On Saturday, February 2, 2019 at 5:59:55 PM UTC, Dan Povey wrote:

Not really, but it's not very difficult, basically applying a cosine transform, you can see the MFCC code for how to do it.

On Sat, Feb 2, 2019 at 9:00 AM <srik...@zapr.in> wrote:

Hi,

   I am extracting Mel-filter bank filter from a wav file. I do some processing on the Fbank features to clean up the background music.

Now i have to convert the Fbank features to Mfcc before feeding it to the decoder.

Is there a script/code that takes Fbank features and generate Mfcc features....    

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
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...@googlegroups.com.

[Daniel Povey]

There are some 'liftering coefficients' used to scale the MFCC coeffs
to get them in the same range.
Also I think there is some kind of transpose when you do the inverse
DCT (you probably had that already though.)

> To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/4151b0bd-dac3-4e7e-a8ec-01f72c8cfd97%40googlegroups.com.

[feifei xiong]

Hi Dan,

Many thanks! 

I forgot to consider the 'liftering coef' (default 22). Now I think they are equivalent: Kaldi-FBANK (dim x frames) + dct (Matlab version type II = inverse dct labeled in Kaldi with correct normalized factor) + liftering ( L=22, @(N,L)(1+0.5*L*sin(pi*[0:N-1]/L) ) ) = Kaldi-MFCC.

> To unsubscribe from this group and stop receiving emails from it, send an email to kaldi...@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/4151b0bd-dac3-4e7e-a8ec-01f72c8cfd97%40googlegroups.com.