VoxForge
VoxForge
Hi,
I recently installed sphinx3 on my pc and was able to setup my own recognition system.
I could also adapt the RM1 database in a supervised way to my voice and found the recognition to improve.
When I started off with supervised adaptation, I came across this function in sphinxtrain( bw ) which carries out 1 iteration of the forward-backward algoritm.
When i do not specify any transcript file(i.e. unsupervised), the function just dies when i call it in command line. It seems to be working fine when it(transcript file) is specified.
So, does sphinx3 offer unsupervised adaptation features? If yes, how do I go about doing it? bw and mllr dont seem to be working..atleast i dont seem to be able to figure out how to make them work.
pl. post on suggestions and methods to do unsupervised speaker adaptation in sphinx3.
Cheers
--- (Edited on 5/27/2009 8:11 am [GMT-0500] by dsubbu) ---
Could you please be more patient and don't crosspost the same message?
If you want offline unsupervised adaptation there is no way except to recognize and adapt with the recognition result then. Mllr_solve and bw require the list of transcriptions.
--- (Edited on 5/28/2009 4:37 pm [GMT-0500] by nsh) ---
hi,
im so sorry for doing that.. yes, i was being impatient because i wanted to clear it before the weekend and look for other engines which might support it.
So, if i have to recognise and adapt, the alignment of transcriptions, measuring word confidences etc must only be done by me right?
Thanks a lot for the reply and can anyone suggest an open souce engine which can actually implement such an option?
sorry again for posting the same msg twice.
Thanks.
Cheers
--- (Edited on 5/28/2009 11:00 pm [GMT-0500] by dsubbu) ---
> an open souce engine which can actually implement such an option?
I'm not aware about this.
--- (Edited on 5/29/2009 1:53 am [GMT-0500] by nsh) ---
hi dsubbu
U are done Speaker adaptation.
Plz u can give any information or document about speaker adaptation how to do the speaker adaptation.
Thanks
--- (Edited on 6/21/2009 8:08 am [GMT-0500] by Visitor) ---
hi dsubbu
U are done Speaker adaptation.
Plz u can give any information or document about speaker adaptation how to do the speaker adaptation.
Thanks
--- (Edited on 6/21/2009 8:08 am [GMT-0500] by Visitor) ---
Hi Raj,
Im just getting started on this stuff myself. But the Sphinx3 documentation (the Hieroglyphs) book has a chapter (chapter 10) on speaker adaptation which gave me improvements in recognition
If you are using HTK, there is a good tutorial on the voxforge site itself
http://www.voxforge.org/home/dev/acousticmodels/linux/adapt/htkjuliu
Cheer
--- (Edited on 6/22/2009 12:38 am [GMT-0500] by Visitor) ---
Hi dsubbu
Thanks for given reply.
I am using sphinx3 and my database telugu.
Training and decoding process is over.
Now I want do speaker adaptation.
In sphinx have doc is
http://www.speech.cs.cmu.edu/cmusphinx/moinmoin/AcousticModelAdaptation
and Hieroglypgs book also I seen.
But I am not understand what is directory structure.
In sphinx doc all files put in one folder
like
adapt/.wav files
/.dict
/.fileids
/.transaction
/bw
/mllr_solve
/ .txt
my qustion is
1.Again we need SphinxTrain and sphinxbase to convert wav to mfc yes or no.
2.What is the use of sphinxbase_fe command
sphinx_fe :it convert wav files to mfcs but mfcs also find with SphinxTrain also then the documentation why tell sphinx_fe.
what is difference between SphinxTrain and adaptation.
If u know plz help to me,because this is My final year project it is.
Atleast u have tell what directory structure ur following
ie what order ur files are puting
Thanks u so much.
--- (Edited on 6/23/2009 7:32 am [GMT-0500] by Visitor) ---
Hi Raj,
As I told you already, I am new to this stuff too but this is the method I followed.
Collect adaptation data first
Use bw, mllr_solve and mllr_transform on your mean, variance and transition matrix files to obtain the transformed model. Sphinx gives a few of these sets depending on the number of senones.
I chose the set that gave me the best results before I did the adaptation
I followed section 10.3 for this and did not pursue it further once I found that I could see tangible improvements.
I hope this is of some use.Also, I am not using Sphinx anymore now.
Alternatively, you could also look at MAP adaptation in sec 10.4
--- (Edited on 6/23/2009 11:35 pm [GMT-0500] by Visitor) ---
hi dsubbu
Thanks a lot given reply to me.sorry for u,again i am asking question.
After adaptation I get mixw_counts,tmat_counts,gauden_cunts ,adapt.matrix and out.mean.
Its fine.
Now I wish to decode with adapted model
here I don't know which command to run on sphinx3 command line
I think sphinx3_decode and what command line arguments I want to give ,where I can adapt.matrix is specified. or how to specify configuration file
Plz tell me because my project duration time very less.
--- (Edited on 6/27/2009 11:00 am [GMT-0500] by Visitor) ---