1s_c_d_ld_dd

[asr2010]

I was wondering what is the "ld" in the 1s_c_d_ld_dd feature type?

In the case of 1s_c_d_ld_dd, what is the size of the feature vector?

[Nickolay V. Shmyrev]

ld means long deltas

Just delta

d = (x[-1] + x[1])  / 2

Long delta

ld = (x[-2] + x[2])  / 2

The feature vector size is 52 or 13 * 4

[asr2010]

according to your equation, you are doing an average and long average.
d and ld should be:
d = x-x
ld=x-x

Am I right?

I have another question: from the documentation of sphinx 4 at

http://cmusphinx.sourceforge.net/sphinx4/javadoc/edu/cmu/sphinx/frontend/feat
ure/DeltasFeatureExtractor.html

I found out that d=x-x ..is this correct?

I think there will be a problem If sphinx4 considers d as x-x whilst
sphinxtrain consider it as x-x,

[asr2010]

I have just checked the source code..
d = x - x
ld = x - x

[Nickolay V. Shmyrev]

You are right

[asr2010]

Do you have any reference where they have combined delta and long delta? I
couldn't find something like that

[Nickolay V. Shmyrev]

Do you have any reference where they have combined delta and long delta? I
couldn't find something like that

Sorry, I'm not sure what are you asking about. Who are "they"

[asr2010]

Do you know any research paper or a book where the authors have introduced the
combination of delta and long delta? because in my opinion the combination
will give more weight to delta and the effect of double delta will decrease.

[Nickolay V. Shmyrev]

Long delta features originate from sphinx2. You can read about them from
sphinx2 overview

http://acl.ldc.upenn.edu/H/H93/H93-1016.pdf

However there is no reasoning behind them, only a pure belief for specifically
for semi-continuous system. Nobody uses them either.

[Nickolay V. Shmyrev]

Well, theoretically they increase feature window incorporating long-distance
features which is known to be a good thing from RASTA research. But it can be
done one or other way around.