-
Hello Nikko,
What is the error measure used in the back-prop? since we assume that the output activities are posterior probabilities, literature says the cross entropy measure should be used so the network can map to a Bayesian classifier. Does it use cross entropy by default?
-S.
2009-02-12 23:10:24 UTC in The NICO Toolkit
-
Hello,
I dont know if you are still using this forum, but hope so.
I wanted to ask if you tried your code on a speech pitch extraction, not music chords.
and did you try to make it real time ?
Greets,
Samer.
2008-02-19 00:53:36 UTC in MFFM pitch algorithm
-
As you said, the backpropagation will just tend to converge to give a higher accuracy disregaring what it is classifying, and definitely that depends on your data. the backpropagation cant take the decision of ignoring (frequent occuring) input, i would say, the best is that you try to trim parts of your data (like silence) as much as you can, so you get a balanced number of occurances among...
2007-11-29 14:06:47 UTC in The NICO Toolkit
-
Hi Alfonso.
You dont need to divide your data, it's enough to give the backprop two script files, one including the names of the files for training and one for the vaildation, and using the -S option you just pass the script file with the -V option, you dont need to make new input and output streams as long as the files have the same structure.
following you see a BackProp call...
2007-11-26 12:14:45 UTC in The NICO Toolkit
-
Ok, But I need a fast phone recognizer, which gives priority to NN.
What if I just need to give as an output the probability of more than one phone per frame, (a kind of digitized lattice on the frame level)?
Would also the Lab2Targ support it?.
2007-11-02 13:48:11 UTC in The NICO Toolkit
-
Hi again,
Is it possible in principle to use the tool to estimate phone lattice instead of only a string of phones? does Lab2Targ support this representation?
Thanks,
2007-10-31 15:21:19 UTC in The NICO Toolkit
-
I meant, how does the BackProp internally deal with the boundaries (to understand how it's running)?
While I am training. running the BackProp, I am getting this error "Warning: External data are of large magnitute (>10.0)"
Do you have a guess?.
2007-10-26 10:49:10 UTC in The NICO Toolkit
-
Alright, Tack!
Can you tell me how Nico deals with boundries (in case like example time delayed recurrent training)?
Samer.
2007-10-24 14:59:31 UTC in The NICO Toolkit
-
Hi,
I need to train a network for phonetic transcription, for the moment, i just need to transfer the structure of the aligned utterances to one which is readable by NICO.
I have the HTK features (MFCCs) in a tree directories, as well, with each file there is the .rec file (mlf) in the structure of (beginframe endframe phoneme) accompanied with the feature file in the same directory, and i...
2007-10-23 14:26:13 UTC in The NICO Toolkit
