FMLLR features and decoding

[subash khanal]

Hi all,

I am trying to use TIMIT as a base recipe for my database. I wanted to compute FMLLR features and used

steps/nnet/make_fmllr_feats.sh data-fmllr/train data/train exp/tri3a exp/make_fmllr_feats2/log plp2/processed/
steps/nnet/make_fmllr_feats.sh data-fmllr/train data/train exp/tri3a exp/make_fmllr_feats2/log plp2/processed/

It threw me error as " Invalid feature type [UNKNOWN]" 

The graph in tri3a is after LDA+MLLT+SAT training.

I thought the steps/decode_fmllr.sh also creates fmllr transforms and used it. While decoding, I am required the conf file phones.60-48-39.map as for timit, I don't have that so instead I created three columns of all the phonemes in the my corpus (total 41 (39 + SIL + OOV) and named the file as phones.60-48-39.map, just to avoid changing its name in locations in scripts. While running the decode_fmllr I now get error " Bad line: T T T at local/timit_norm_trans.pl ..."I believe it has to do with my way of creating the phone's map. Moreover, I believe the decoding steps will make use of such maps so what should one do in using the custom phones.x-x-x.map for our corpus and still use TIMIT recipe framework?

 

 Either way, I am still unable to have FMLLR features for my dataset.

Can you help me in obtaining FMLLR features from a set of features I already have (Say MFCC+DELTA+DELTA)?

Any help would be highly appreciated.

[Daniel Povey]

I have a policy of never answering questions about TIMIT because the real answer for any such question is, "Don't use TIMIT, and don't base your recipes on TIMIT".

--
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/CA%2B30hVk1J%3DG1pur7Rujd%3DNPSX9Bp5QBgo-ChjPQ7wqebangMNA%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.

[subash khanal]

Hello,

Apologies for my approach above. All I am stuck at is while using steps/nnet/make_fmllr_feats.sh . My PER of the system is very poor(60%) and I want to see if fmllr features would do any better. The make_fmllr_feats.sh is not able to figure out the feature type,(feature type is [UNKNOWN]) the tri3 GMM graph and the data files are in right place. This is a general question about getting fmllr features from the existing features(maybe MFCC+DELTA+DELTA or something else). I would appreciate any help.

Thank you.

Subash

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAEWAuyT_zNy2B6AuAOgRgAAcYDe2OU8EwEv2LMgg8Q5hQv3NUQ%40mail.gmail.com.

[Disenmas]

It's so cute..LOL

在 2019年6月30日星期日 UTC+8上午2:59:25，Dan Povey写道：

I have a policy of never answering questions about TIMIT because the real answer for any such question is, "Don't use TIMIT, and don't base your recipes on TIMIT".

On Sat, Jun 29, 2019 at 11:24 AM subash khanal <subash....@gmail.com> wrote:

Hi all,

I am trying to use TIMIT as a base recipe for my database. I wanted to compute FMLLR features and used

steps/nnet/make_fmllr_feats.sh data-fmllr/train data/train exp/tri3a exp/make_fmllr_feats2/log plp2/processed/
steps/nnet/make_fmllr_feats.sh data-fmllr/train data/train exp/tri3a exp/make_fmllr_feats2/log plp2/processed/

It threw me error as " Invalid feature type [UNKNOWN]" 

The graph in tri3a is after LDA+MLLT+SAT training.

I thought the steps/decode_fmllr.sh also creates fmllr transforms and used it. While decoding, I am required the conf file phones.60-48-39.map as for timit, I don't have that so instead I created three columns of all the phonemes in the my corpus (total 41 (39 + SIL + OOV) and named the file as phones.60-48-39.map, just to avoid changing its name in locations in scripts. While running the decode_fmllr I now get error " Bad line: T T T at local/timit_norm_trans.pl ..."I believe it has to do with my way of creating the phone's map. Moreover, I believe the decoding steps will make use of such maps so what should one do in using the custom phones.x-x-x.map for our corpus and still use TIMIT recipe framework?

 

 Either way, I am still unable to have FMLLR features for my dataset.

Can you help me in obtaining FMLLR features from a set of features I already have (Say MFCC+DELTA+DELTA)?

Any help would be highly appreciated.

--
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.

[subash khanal]

Sorry for the issue above, the problem was my decode.sh used the scoring function based on Timit recipe (that had mapping of phonemes and all tasks specific to Timit) I changed the scoring function to wsj's which doesn't require such phone map. And I am able to get the decoding results with WER. Thanks.

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/1e072263-6ca1-4151-84f4-417df01b32df%40googlegroups.com.