I would like to create MLLR regression classes and I have a mapping from each
senone to regression classes.
The problem is that I don't know what should be the structure of the "nmap"
file when using the mk_mllr_class command.
I need just a sample for line in the "nmap" file.
Another thing that I cannot understand, what is the purpose of the "nclass"
parameters in the same "mk_mllr_class" command? as far as I understood, this
number can be estimated from the mapping file.
Your help is appreciated.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
There was a little issue with documentation, the nmap argument doesn't provide
a file name but a number of input senones.
The updated doc looks this way:
./mk_mllr_class \
-help yes
Description:
Create the senone to mllr class mapping. This program just reads the mapping
from
stdin one map per line and converts it to binary form. No real work on
estimation
of the number of classes id done. The format of the input maps is simple,
just one mapping per line:
0001
Which means that we map senones 0,1,2 to class 0 and senone 3 to class 1.
It seems that each senones should belong to regression class. In case there
are some specific senones that I don't want to adapt ...what shall we we do??
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
It seems that each senones should belong to regression class. In case there
are some specific senones that I don't want to adapt ...what shall we we do??
Write code to implement missing feature
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
if I create a Mapping file manually (help of Phonetic knowledge) and my map
file is like this
1 0 2 1
I think we are providing here only the base classes....(if I am wrong ,kindly
correct me)
if so ,where and how to provide regression tree to generate transformations
for those nodes that
have sufficient adaptation data.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
ok.for regression trees,during the adaptation process, the occupation counts
for each of the base classes should be found...is it possible with BW ? and
how?.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
If we are talking specifically about tree, why not. You just collect stats for
all the tree nodes, there is no problem to do that in bw.
Right now bw can collect statistics for raw adaptation classes that it
supports.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I would like to create MLLR regression classes and I have a mapping from each
senone to regression classes.
The problem is that I don't know what should be the structure of the "nmap"
file when using the mk_mllr_class command.
I need just a sample for line in the "nmap" file.
Another thing that I cannot understand, what is the purpose of the "nclass"
parameters in the same "mk_mllr_class" command? as far as I understood, this
number can be estimated from the mapping file.
Your help is appreciated.
There was a little issue with documentation, the nmap argument doesn't provide
a file name but a number of input senones.
The updated doc looks this way:
./mk_mllr_class \
-help yes
Description:
Create the senone to mllr class mapping. This program just reads the mapping
from
stdin one map per line and converts it to binary form. No real work on
estimation
of the number of classes id done. The format of the input maps is simple,
just one mapping per line:
Which means that we map senones 0,1,2 to class 0 and senone 3 to class 1.
Example:
Thank you for your detailed reply.
basically the total number of lines in the in.cd2mllr.txt files should be the
same as nmap, and nclass is equal to the "last line value +1"
for the example you mentioed: nmap = 4 and nclass = 2.
Am I right?
Well, class id shouldn't be strictly growing, it can be like "1 1 1 0" but you
got idea correctly.
great..
It seems that each senones should belong to regression class. In case there
are some specific senones that I don't want to adapt ...what shall we we do??
Write code to implement missing feature
if I create a Mapping file manually (help of Phonetic knowledge) and my map
file is like this
1 0 2 1
I think we are providing here only the base classes....(if I am wrong ,kindly
correct me)
if so ,where and how to provide regression tree to generate transformations
for those nodes that
have sufficient adaptation data.
SphinxTrain doesn't support regression trees. Only fixed set of classes. You
need to add this support yourself if you need it.
ok.for regression trees,during the adaptation process, the occupation counts
for each of the base classes should be found...is it possible with BW ? and
how?.
If we are talking specifically about tree, why not. You just collect stats for
all the tree nodes, there is no problem to do that in bw.
Right now bw can collect statistics for raw adaptation classes that it
supports.