recently I am using Kaldi for ASR with our own feature type, which has almost 600 dimension.
definitely it is much larger than e.g. #39 of MFCC+d+dd.
if I understood correctly, I need to adjust the 'beam' beam width to larger than e.g. 13 to get
through the mono and triphone training.
otherwise, I got warnings: No alignment found!
during decoding, I always got warnings: "no final-state reached"
hence, the WER also reaches almost 100%.
I would like to ask:
whether there are some rules of thumb to adjust these values:
--beam
--latbeam
--acwt
I tested with WSJ corpus, and tried lots of combinations of the above parameters e.g. as
--beam 60 (retry with 240)
--latbeam 36
--acwt 0.008333 (1/120 with lmwt 120)
feat dim = 657
num_gauss(15000) and num_leaves(2500) are kept the same as the default Kaldi script.
however, still got the warning: "no final-state reached", resulting in experiments fail !
please help to suggestion which way should I adjust these parameters...
thanks in advance ~
Fei
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I think that is too high dimension for common gmm/sgmm system. So even if
you find beam width that will cause the decoding succeed, i don't think
the performance of such system will be any good.
You should either use lda ( or other dimensionality reduction method) or
(better) train a nnet system.
recently I am using Kaldi for ASR with our own feature type, which has
almost 600 dimension.
definitely it is much larger than e.g. #39 of MFCC+d+dd.
if I understood correctly, I need to adjust the 'beam' beam width to
larger than e.g. 13 to get
through the mono and triphone training.
otherwise, I got warnings: No alignment found!
during decoding, I always got warnings: "no final-state reached"
hence, the WER also reaches almost 100%.
I would like to ask:
whether there are some rules of thumb to adjust these values:
--beam
--latbeam
--acwt
I tested with WSJ corpus, and tried lots of combinations of the above
parameters e.g. as
--beam 60 (retry with 240)
--latbeam 36
--acwt 0.008333 (1/120 with lmwt 120)
feat dim = 657
num_gauss(15000) and num_leaves(2500) are kept the same as the default
Kaldi script.
however, still got the warning: "no final-state reached", resulting in
experiments fail !
please help to suggestion which way should I adjust these parameters...
thanks in advance ~
thanks for the info.
HLDA and DNN work fine for such large dim. features.
however, just would like to compare if we feed these features directly to GMM/SGMM,
what the WER will be.
Fei
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
do you know the relation of the beam width between HtK and Kaldi,
e.g. from HTK, beam is typically set to "250" until "2000"
from Kaldi scripts, it is chosen as beam '13' until retry_beam '13*4'
then how to correspond each other of these settings?
some scaling factor from Kaldi to HTK for this beam width?
thanks in advance !
Fei
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
In Kaldi we scale the acoustics down and in HTK they scale the
language-model scale up, so there is a difference of a factor of 10 that
comes from that.
The (beam,retry_beam) seems to be (10,40) in most of the Kaldi scripts.
Dan
do you know the relation of the beam width between HtK and Kaldi,
e.g. from HTK, beam is typically set to "250" until "2000"
from Kaldi scripts, it is chosen as beam '13' until retry_beam '13*4'
then how to correspond each other of these settings?
some scaling factor from Kaldi to HTK for this beam width?
do you know the relation of the beam width between HtK and Kaldi,
e.g. from HTK, beam is typically set to "250" until "2000"
from Kaldi scripts, it is chosen as beam '13' until retry_beam '13*4'
then how to correspond each other of these settings?
some scaling factor from Kaldi to HTK for this beam width?
Hi Kaldis,
recently I am using Kaldi for ASR with our own feature type, which has almost 600 dimension.
definitely it is much larger than e.g. #39 of MFCC+d+dd.
if I understood correctly, I need to adjust the 'beam' beam width to larger than e.g. 13 to get
through the mono and triphone training.
otherwise, I got warnings: No alignment found!
during decoding, I always got warnings: "no final-state reached"
hence, the WER also reaches almost 100%.
I would like to ask:
whether there are some rules of thumb to adjust these values:
--beam
--latbeam
--acwt
I tested with WSJ corpus, and tried lots of combinations of the above parameters e.g. as
--beam 60 (retry with 240)
--latbeam 36
--acwt 0.008333 (1/120 with lmwt 120)
feat dim = 657
num_gauss(15000) and num_leaves(2500) are kept the same as the default Kaldi script.
however, still got the warning: "no final-state reached", resulting in experiments fail !
please help to suggestion which way should I adjust these parameters...
thanks in advance ~
Fei
I think that is too high dimension for common gmm/sgmm system. So even if
you find beam width that will cause the decoding succeed, i don't think
the performance of such system will be any good.
You should either use lda ( or other dimensionality reduction method) or
(better) train a nnet system.
Y.
Y.
On Nov 2, 2014 10:41 PM, "Fei Xiong" xffmqjx@users.sf.net wrote:
Hi Jan,
thanks for the info.
HLDA and DNN work fine for such large dim. features.
however, just would like to compare if we feed these features directly to GMM/SGMM,
what the WER will be.
Fei
Hi Jim,
do you know the relation of the beam width between HtK and Kaldi,
e.g. from HTK, beam is typically set to "250" until "2000"
from Kaldi scripts, it is chosen as beam '13' until retry_beam '13*4'
then how to correspond each other of these settings?
some scaling factor from Kaldi to HTK for this beam width?
thanks in advance !
Fei
In Kaldi we scale the acoustics down and in HTK they scale the
language-model scale up, so there is a difference of a factor of 10 that
comes from that.
The (beam,retry_beam) seems to be (10,40) in most of the Kaldi scripts.
Dan
On Thu, Nov 6, 2014 at 10:06 AM, Fei Xiong xffmqjx@users.sf.net wrote:
Kaldi scales down the acoustic cost:
cost = am_cost * am_scale + lm_cost (am_scale < 1)
HTK scales up the lm_cost:
cost = am_cost + lm_cost * lm_scale (lm_scale > 1)
On 6 November 2014 16:06, Fei Xiong xffmqjx@users.sf.net wrote:
Hi Dan, hi Paul,
thanks for the explanations.
after fine-tuning the 'acwt' but keeping the 'beam' untouched, now some reasonable decoding results come out !
Fei
Last edit: Amiel Tadesse 2015-03-23