Understanding SAT, FMLLR

282 views
Skip to first unread message

subash khanal

unread,
Jun 28, 2019, 1:58:09 PM6/28/19
to kaldi...@googlegroups.com
Hello All,

This is my issue regarding understanding SAT and FMLLR. From following discussions in this topic so far I know "the practice is to estimate fMLLR matrices for the individual training speakers and train on the adapted features; this is called Speaker Adapted Training (SAT)" but still I am not able to understand how the fMLLR matrices for the individual training speakers are useful while performing test on testing speakers. Don't we need to adapt the SAT models (I do not understand how these speaker dependent models are used for testing, there is not one SI model as a reference to adapt) with a portion of test speakers data? If so which script does this and with how much of data? Putting it in more general way I want to know what is happening, when I run
steps/train_sat.sh
steps/decode_fmllr.sh
steps/align_fmllr.sh

I know this issue is about my lack of good theoretical  understanding in these topics. Any help in understanding this concept will be helpful.

With regards,
Subash

Daniel Povey

unread,
Jun 28, 2019, 2:01:31 PM6/28/19
to kaldi-help
The idea is the transforms normalize all speakers to look more like each other, so you are training on "normalized" speech signals.

--
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%2B30hV%3DA_5OuGBHt5ZzJqr05zsybbqL9H3LuKw9zYArAKNZ9vA%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.

subash khanal

unread,
Jun 28, 2019, 2:18:14 PM6/28/19
to kaldi...@googlegroups.com
Hi Dan,

Thank you for your response. My database has native and non native speakers of English and aim is to detect pronunciation errors using the models trained on native American speakers, in such case , I would expect to perform some adaptation to non-native test speakers, so how should I approach this in kaldi?

With regards,
Subash

Daniel Povey

unread,
Jun 28, 2019, 11:40:39 PM6/28/19
to kaldi-help
I wouldn't do anything special about that.

Reply all
Reply to author
Forward
0 new messages