Menu
▾
▴
kaldi-developers — Kaldi Developers
You can subscribe to this list here.
| 2011 |
Jan
|
Feb
|
Mar
|
Apr
|
May
|
Jun
(4) |
Jul
|
Aug
|
Sep
(1) |
Oct
(4) |
Nov
(1) |
Dec
(14) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2012 |
Jan
(1) |
Feb
(8) |
Mar
|
Apr
(1) |
May
(3) |
Jun
(13) |
Jul
(7) |
Aug
(11) |
Sep
(6) |
Oct
(14) |
Nov
(16) |
Dec
(1) |
| 2013 |
Jan
(3) |
Feb
(8) |
Mar
(17) |
Apr
(21) |
May
(27) |
Jun
(11) |
Jul
(11) |
Aug
(21) |
Sep
(39) |
Oct
(17) |
Nov
(39) |
Dec
(28) |
| 2014 |
Jan
(36) |
Feb
(30) |
Mar
(35) |
Apr
(17) |
May
(22) |
Jun
(28) |
Jul
(23) |
Aug
(41) |
Sep
(17) |
Oct
(10) |
Nov
(22) |
Dec
(56) |
| 2015 |
Jan
(30) |
Feb
(32) |
Mar
(37) |
Apr
(28) |
May
(79) |
Jun
(18) |
Jul
(35) |
Aug
|
Sep
(1) |
Oct
|
Nov
|
Dec
|
|
From: Daniel P. <dp...@gm...> - 2013-12-03 16:20:34
|
Hi, It looks like those files were removed from that server. It was probably accidental, but anyway, in order to avoid this type of thing happening in future I replaced all of those files with a location that I maintain, at www.danielpovey.com. Dan On Tue, Dec 3, 2013 at 8:12 AM, Vassil Panayotov <vas...@gm... > wrote: > Hi, > the primary server will be probably fixed, but I uploaded the file to > Kaldi's file area just in case. If you run "make sph2pipe" in tools it > should fall back to SourceForge if the file is not found on the primary > host. > > (If someone knows a reason (e.g. licensing) why this is not desirable, > please feel free to revert the change) > > Vassil > > > On Mon, Dec 2, 2013 at 1:52 PM, Eamonn Kenny <Eam...@cs...>wrote: > >> Dear Kaldi Developers, >> >> I've tried to build Kaldi trunk for the past week and find that >> it always fails on the download of sph2pipe. Will the Czech >> site where the tarball resides be fixed soon? >> >> Also, I'll discovered on Ubuntu 13.10 (Saucy) that it fails to >> build openfst with dlopen and dlerror issues. I've seen this >> issue mentioned on a number of sites, but wonder is it now >> resolved or fixable? I've found that Debian Wheezy (7.2) is >> much more reliable. >> >> -- >> Best Regards, >> Eamonn Kenny >> >> + + + >> Ema...@cs... CNGL/Speech Communication Lab, >> Tel: 00+353-1-8961797 Dept. of Computer Science, >> F.34, O'Reilly Institute, >> http://www.cs.tcd.ie/Eamonn.Kenny Trinity College Dublin, >> http://eamonnmkenny.wordpress.com Dublin 2, Ireland. >> + + + >> >> >> >> ------------------------------------------------------------------------------ >> Rapidly troubleshoot problems before they affect your business. Most IT >> organizations don't have a clear picture of how application performance >> affects their revenue. With AppDynamics, you get 100% visibility into your >> Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics >> Pro! >> >> http://pubads.g.doubleclick.net/gampad/clk?id=84349351&iu=/4140/ostg.clktrk >> _______________________________________________ >> Kaldi-developers mailing list >> Kal...@li... >> https://lists.sourceforge.net/lists/listinfo/kaldi-developers >> > > > > ------------------------------------------------------------------------------ > Rapidly troubleshoot problems before they affect your business. Most IT > organizations don't have a clear picture of how application performance > affects their revenue. With AppDynamics, you get 100% visibility into your > Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics > Pro! > http://pubads.g.doubleclick.net/gampad/clk?id=84349351&iu=/4140/ostg.clktrk > _______________________________________________ > Kaldi-developers mailing list > Kal...@li... > https://lists.sourceforge.net/lists/listinfo/kaldi-developers > > |
|
From: Vassil P. <vas...@gm...> - 2013-12-03 13:12:08
|
Hi, the primary server will be probably fixed, but I uploaded the file to Kaldi's file area just in case. If you run "make sph2pipe" in tools it should fall back to SourceForge if the file is not found on the primary host. (If someone knows a reason (e.g. licensing) why this is not desirable, please feel free to revert the change) Vassil On Mon, Dec 2, 2013 at 1:52 PM, Eamonn Kenny <Eam...@cs...> wrote: > Dear Kaldi Developers, > > I've tried to build Kaldi trunk for the past week and find that > it always fails on the download of sph2pipe. Will the Czech > site where the tarball resides be fixed soon? > > Also, I'll discovered on Ubuntu 13.10 (Saucy) that it fails to > build openfst with dlopen and dlerror issues. I've seen this > issue mentioned on a number of sites, but wonder is it now > resolved or fixable? I've found that Debian Wheezy (7.2) is > much more reliable. > > -- > Best Regards, > Eamonn Kenny > > + + + > Ema...@cs... CNGL/Speech Communication Lab, > Tel: 00+353-1-8961797 Dept. of Computer Science, > F.34, O'Reilly Institute, > http://www.cs.tcd.ie/Eamonn.Kenny Trinity College Dublin, > http://eamonnmkenny.wordpress.com Dublin 2, Ireland. > + + + > > > > ------------------------------------------------------------------------------ > Rapidly troubleshoot problems before they affect your business. Most IT > organizations don't have a clear picture of how application performance > affects their revenue. With AppDynamics, you get 100% visibility into your > Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics > Pro! > http://pubads.g.doubleclick.net/gampad/clk?id=84349351&iu=/4140/ostg.clktrk > _______________________________________________ > Kaldi-developers mailing list > Kal...@li... > https://lists.sourceforge.net/lists/listinfo/kaldi-developers > |
|
From: Serg Re <re...@gm...> - 2013-12-03 09:25:20
|
Hi. After update building was stopped in fgmmbin with linking error. It seems that in src/fgmmbin/Makefile in ADDLIBS missing library ../thread/kaldi-thread.a. In my case adding this library is solve problem. (BTW in src/gmmbin/Makefile in ADDLIBS library ../gmm/kaldi-gmm.a is referred twice) Best regards, Sergey |
|
From: Eamonn K. <Eam...@cs...> - 2013-12-02 12:09:04
|
Dear Kaldi Developers,
I've tried to build Kaldi trunk for the past week and find that
it always fails on the download of sph2pipe. Will the Czech
site where the tarball resides be fixed soon?
Also, I'll discovered on Ubuntu 13.10 (Saucy) that it fails to
build openfst with dlopen and dlerror issues. I've seen this
issue mentioned on a number of sites, but wonder is it now
resolved or fixable? I've found that Debian Wheezy (7.2) is
much more reliable.
--
Best Regards,
Eamonn Kenny
+ + +
Ema...@cs... CNGL/Speech Communication Lab,
Tel: 00+353-1-8961797 Dept. of Computer Science,
F.34, O'Reilly Institute,
http://www.cs.tcd.ie/Eamonn.Kenny Trinity College Dublin,
http://eamonnmkenny.wordpress.com Dublin 2, Ireland.
+ + +
|
|
From: Daniel P. <dp...@gm...> - 2013-11-28 16:30:19
|
If you don't have the RM data there is no point trying to exactly replicate that structure. There is a "data preparation" page on kaldi.sf.net that explains how you need to prepare the data in the general case. Dan On Wed, Nov 27, 2013 at 12:36 PM, Tuan Dinh <tua...@is...> wrote: > Dear all, > > I'm studying deep neural network in speech recognition and Kaldi is a very > interesting and appropriate framework for my research. Now I have some > problems with data preparation. For some reasons, I can't obtain LDC93S3A > so I want to use another speech data instead. According to the tutorial, I > have to prepare a folder /LDC/LDC93S3A/rm_comp which contains 3 sub folders: > rm1_audio1, rm1_audio2 and rm2_audio. So what do the sub-folers contain? > Where do you put audio files and text files in LDC data folder? I want to > know more detail about the structure of the data folder. > I very appreciate your helps. > > Best regards, > Dinh Anh Tuan. > > ------------------------------------------------------------------------------ > Rapidly troubleshoot problems before they affect your business. Most IT > organizations don't have a clear picture of how application performance > affects their revenue. With AppDynamics, you get 100% visibility into your > Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics > Pro! > http://pubads.g.doubleclick.net/gampad/clk?id=84349351&iu=/4140/ostg.clktrk > _______________________________________________ > Kaldi-developers mailing list > Kal...@li... > https://lists.sourceforge.net/lists/listinfo/kaldi-developers > |
|
From: Daniel P. <dp...@gm...> - 2013-11-28 16:27:43
|
There is no such thing as speaker adaptive training using MAP. I think there may be a script that decodes with MAP somewhere but I doubt it's what you really need, unless you have >1 hour of data per speaker Dan On Thu, Nov 28, 2013 at 10:20 AM, Arif Khan <ife...@gm...> wrote: > Dear Kaldi Authors, > > Do we have adaptation script that uses MAP. Like Speaker adaptive training > using MAP and than decode with MAP. > > Best regards, > Arif > > > On Fri, Nov 15, 2013 at 5:03 PM, Daniel Povey <dp...@gm...> wrote: >> >> What you're talking about here is basis-fMLLR, not regular fMLLR. >> basis-fMLLR is a kind of fMLLR suitable for adapting on short data, >> e.g. 10 to 30 seconds of data per speaker. >> You probably just want regular fMLLR in this instance. Do e.g. >> train_sat.sh and decode_fmllr.sh, that is the normal recipe. >> basis_fmllr is something for a special case >> Dan >> >> > Dear Kaldi Authors, >> > >> > I want to use the fmllr adaptation, but in the wsj/s5 recipe >> > [ger_fmllr_basis.sh] script, >> > >> > it is written that: >> > --------------------------------------------------------- >> > per_utt=true >> > If true, then treat each utterance as a separate speaker for purposes of >> > basis training... this is recommended if the number of actual speakers >> > in >> > your >> > training set is less than (feature-dim) * (feature-dim+1). >> > --------------------------------------------------------- >> > >> > If I don't have the number of speaker >= ~1500 in my training set, than >> > fmllr adaptation per speaker is not good? >> > >> > Can you please explain in some detail? >> > >> > I have a small set of data about 1000 utterances, that comes from 5 >> > speaker >> > with 3 to 4 second per utterance. Like between 100 - 200 utterances per >> > speaker. What kind of adaptation technique should I use to get good >> > results >> > per speaker? >> > >> > Best regards, >> > Arif >> > >> > >> > ------------------------------------------------------------------------------ >> > DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps >> > OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access >> > Free app hosting. Or install the open source package on any LAMP server. >> > Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native! >> > >> > http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk >> > _______________________________________________ >> > Kaldi-developers mailing list >> > Kal...@li... >> > https://lists.sourceforge.net/lists/listinfo/kaldi-developers >> > > > |
|
From: Daniel P. <dp...@gm...> - 2013-11-28 16:24:50
|
It looks like there is a workaround for this issue suggested here: http://www.openfst.org/twiki/bin/view/Forum/FstForum Dan On Thu, Nov 28, 2013 at 10:26 AM, Jan Trmal <af...@ce...> wrote: > ---------- Forwarded message ---------- > From: "Jan Trmal" <af...@ce...> > Date: Nov 28, 2013 10:25 AM > Subject: Re: [Kaldi-developers] Problem making openfst > To: "Saman Azadi" <aza...@gm...> > Cc: > > Hi, perhaps you should make distclean in the openfst dir and run the > configure and make by hand... This seems more like openfst problem than > kaldi install script problem > Y. > > On Nov 28, 2013 4:54 AM, "Saman Azadi" <aza...@gm...> wrote: >> >> Hi, >> >> I'm trying to compile kaldi on ubuntu 13.10 but while making openfst it >> says: >> >> ... >> /bin/bash ../../libtool --tag=CXX --mode=link g++ -g -O2 -o >> fstarcsort fstarcsort.o ../script/libfstscript.la ../lib/libfst.la -lm -ldl >> libtool: link: g++ -g -O2 -o .libs/fstarcsort fstarcsort.o >> ../script/.libs/libfstscript.so ../lib/.libs/libfst.so -lm -ldl -Wl,-rpath >> -Wl,/home/saman/Thesis/kaldi-trunk/tools/openfst-1.3.2/lib >> ../script/.libs/libfstscript.so: undefined reference to `dlopen' >> ../script/.libs/libfstscript.so: undefined reference to `dlerror' >> collect2: error: ld returned 1 exit status >> make[3]: *** [fstarcsort] Error 1 >> make[3]: Leaving directory >> `/home/saman/Thesis/kaldi-trunk/tools/openfst-1.3.2/src/bin' >> make[2]: *** [install-recursive] Error 1 >> make[2]: Leaving directory >> `/home/saman/Thesis/kaldi-trunk/tools/openfst-1.3.2/src' >> make[1]: *** [install-recursive] Error 1 >> make[1]: Leaving directory >> `/home/saman/Thesis/kaldi-trunk/tools/openfst-1.3.2' >> make: *** [openfst_compiled] Error 2 >> >> Seems there is a linker problem with libdl or a misorder GCC (or g++) >> option. How can I fix it? >> >> Thanks in advance for any help you are able to provide. >> >> >> ------------------------------------------------------------------------------ >> Rapidly troubleshoot problems before they affect your business. Most IT >> organizations don't have a clear picture of how application performance >> affects their revenue. With AppDynamics, you get 100% visibility into your >> Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics >> Pro! >> >> http://pubads.g.doubleclick.net/gampad/clk?id=84349351&iu=/4140/ostg.clktrk >> _______________________________________________ >> Kaldi-developers mailing list >> Kal...@li... >> https://lists.sourceforge.net/lists/listinfo/kaldi-developers >> > > ------------------------------------------------------------------------------ > Rapidly troubleshoot problems before they affect your business. Most IT > organizations don't have a clear picture of how application performance > affects their revenue. With AppDynamics, you get 100% visibility into your > Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics > Pro! > http://pubads.g.doubleclick.net/gampad/clk?id=84349351&iu=/4140/ostg.clktrk > _______________________________________________ > Kaldi-developers mailing list > Kal...@li... > https://lists.sourceforge.net/lists/listinfo/kaldi-developers > |
|
From: Daniel P. <dp...@gm...> - 2013-11-28 16:20:38
|
Thanks for pointing this out. I should have recompiled everything before committing that last change. I've fixed it now. Dan On Thu, Nov 28, 2013 at 9:31 AM, Christopher Kermorvant <chr...@a2...> wrote: > Hi, > > I am looking for the exact error. > > So far I did not reproduce the compilation error but I have an error > during the tes phase : > > make -C transform test > make[1]: Entering directory > `/home/kermorvant/Tools/kaldi-code/src/transform' > g++ -msse -msse2 -Wall -I.. -fPIC -DKALDI_DOUBLEPRECISION=0 > -DHAVE_POSIX_MEMALIGN -Wno-sign-compare -Wno-unused-local-typedefs > -Winit-self -DHAVE_EXECINFO_H=1 -rdynamic -DHAVE_CXXABI_H -DHAVE_ATLAS > -I/home/kermorvant/Tools/kaldi-code/tools/ATLAS/include > -I/home/kermorvant/Tools/kaldi-code/tools/openfst/include -g > -DHAVE_CUDA -I/usr/local/cuda/include -c -o > regtree-fmllr-diag-gmm-test.o regtree-fmllr-diag-gmm-test.cc > g++ -rdynamic > -Wl,-rpath=/home/kermorvant/Tools/kaldi-code/tools/openfst/lib > -L/usr/local/cuda/lib64 -Wl,-rpath,/usr/local/cuda/lib64 > regtree-fmllr-diag-gmm-test.o kaldi-transform.a ../gmm/kaldi-gmm.a > ../tree/kaldi-tree.a ../util/kaldi-util.a ../matrix/kaldi-matrix.a > ../base/kaldi-base.a > -L/home/kermorvant/Tools/kaldi-code/tools/openfst/lib -lfst -L/usr/lib > -llapack -lcblas -latlas -lf77blas -lm -lpthread -ldl -lcublas -lcudart > -o regtree-fmllr-diag-gmm-test > ../thread/kaldi-thread.h:61: error: undefined reference to 'vtable for > kaldi::MultiThreadable' > mle-diag-gmm.cc:490: error: undefined reference to > 'kaldi::MultiThreadable::~MultiThreadable()' > mle-diag-gmm.cc:495: error: undefined reference to > 'kaldi::MultiThreadable::~MultiThreadable()' > mle-diag-gmm.cc:514: error: undefined reference to > 'kaldi::MultiThreadable::~MultiThreadable()' > mle-diag-gmm.cc:514: error: undefined reference to > 'kaldi::MultiThreadable::~MultiThreadable()' > ../gmm/kaldi-gmm.a(mle-diag-gmm.o):mle-diag-gmm.cc:function typeinfo for > kaldi::AccumulateMultiThreadedClass: error: undefined reference to > 'typeinfo for kaldi::MultiThreadable' > collect2: ld returned 1 exit status > make[1]: *** [regtree-fmllr-diag-gmm-test] Error 1 > make[1]: Leaving directory `/home/kermorvant/Tools/kaldi-code/src/transform' > > > -- > Chris > > Le 11/28/2013 07:16 AM, jen...@a2... a écrit : >> Kaldi - Build # 270 - Failure: >> >> Check console output at http://jenkins.a2ialab.com/jenkins/job/Kaldi/270/ to view the results. >> >> >> >> ------------------------------------------------------------------------------ >> Rapidly troubleshoot problems before they affect your business. Most IT >> organizations don't have a clear picture of how application performance >> affects their revenue. With AppDynamics, you get 100% visibility into your >> Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics Pro! >> http://pubads.g.doubleclick.net/gampad/clk?id=84349351&iu=/4140/ostg.clktrk >> >> >> _______________________________________________ >> Kaldi-developers mailing list >> Kal...@li... >> https://lists.sourceforge.net/lists/listinfo/kaldi-developers > > > -- > Christopher Kermorvant > Responsable R&D - A2iA - France > 39 rue de la Bienfaisance - 75008 Paris > +33 (0) 144 420 086 / +33 (0) 689 513 601 (mobile) > > > ------------------------------------------------------------------------------ > Rapidly troubleshoot problems before they affect your business. Most IT > organizations don't have a clear picture of how application performance > affects their revenue. With AppDynamics, you get 100% visibility into your > Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics Pro! > http://pubads.g.doubleclick.net/gampad/clk?id=84349351&iu=/4140/ostg.clktrk > _______________________________________________ > Kaldi-developers mailing list > Kal...@li... > https://lists.sourceforge.net/lists/listinfo/kaldi-developers |
|
From: Jan T. <af...@ce...> - 2013-11-28 15:26:37
|
---------- Forwarded message ---------- From: "Jan Trmal" <af...@ce...> Date: Nov 28, 2013 10:25 AM Subject: Re: [Kaldi-developers] Problem making openfst To: "Saman Azadi" <aza...@gm...> Cc: Hi, perhaps you should make distclean in the openfst dir and run the configure and make by hand... This seems more like openfst problem than kaldi install script problem Y. On Nov 28, 2013 4:54 AM, "Saman Azadi" <aza...@gm...> wrote: > Hi, > > I'm trying to compile kaldi on ubuntu 13.10 but while making openfst it > says: > > ... > /bin/bash ../../libtool --tag=CXX --mode=link g++ -g -O2 -o > fstarcsort fstarcsort.o ../script/libfstscript.la ../lib/libfst.la -lm > -ldl > libtool: link: g++ -g -O2 -o .libs/fstarcsort fstarcsort.o > ../script/.libs/libfstscript.so ../lib/.libs/libfst.so -lm -ldl -Wl,-rpath > -Wl,/home/saman/Thesis/kaldi-trunk/tools/openfst-1.3.2/lib > ../script/.libs/libfstscript.so: undefined reference to `dlopen' > ../script/.libs/libfstscript.so: undefined reference to `dlerror' > collect2: error: ld returned 1 exit status > make[3]: *** [fstarcsort] Error 1 > make[3]: Leaving directory > `/home/saman/Thesis/kaldi-trunk/tools/openfst-1.3.2/src/bin' > make[2]: *** [install-recursive] Error 1 > make[2]: Leaving directory > `/home/saman/Thesis/kaldi-trunk/tools/openfst-1.3.2/src' > make[1]: *** [install-recursive] Error 1 > make[1]: Leaving directory > `/home/saman/Thesis/kaldi-trunk/tools/openfst-1.3.2' > make: *** [openfst_compiled] Error 2 > > Seems there is a linker problem with libdl or a misorder GCC (or g++) > option. How can I fix it? > > Thanks in advance for any help you are able to provide. > > > ------------------------------------------------------------------------------ > Rapidly troubleshoot problems before they affect your business. Most IT > organizations don't have a clear picture of how application performance > affects their revenue. With AppDynamics, you get 100% visibility into your > Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics > Pro! > http://pubads.g.doubleclick.net/gampad/clk?id=84349351&iu=/4140/ostg.clktrk > _______________________________________________ > Kaldi-developers mailing list > Kal...@li... > https://lists.sourceforge.net/lists/listinfo/kaldi-developers > > |
|
From: Arif K. <ife...@gm...> - 2013-11-28 15:20:45
|
Dear Kaldi Authors, Do we have adaptation script that uses MAP. Like Speaker adaptive training using MAP and than decode with MAP. Best regards, Arif On Fri, Nov 15, 2013 at 5:03 PM, Daniel Povey <dp...@gm...> wrote: > What you're talking about here is basis-fMLLR, not regular fMLLR. > basis-fMLLR is a kind of fMLLR suitable for adapting on short data, > e.g. 10 to 30 seconds of data per speaker. > You probably just want regular fMLLR in this instance. Do e.g. > train_sat.sh and decode_fmllr.sh, that is the normal recipe. > basis_fmllr is something for a special case > Dan > > > Dear Kaldi Authors, > > > > I want to use the fmllr adaptation, but in the wsj/s5 recipe > > [ger_fmllr_basis.sh] script, > > > > it is written that: > > --------------------------------------------------------- > > per_utt=true > > If true, then treat each utterance as a separate speaker for purposes of > > basis training... this is recommended if the number of actual speakers in > > your > > training set is less than (feature-dim) * (feature-dim+1). > > --------------------------------------------------------- > > > > If I don't have the number of speaker >= ~1500 in my training set, than > > fmllr adaptation per speaker is not good? > > > > Can you please explain in some detail? > > > > I have a small set of data about 1000 utterances, that comes from 5 > speaker > > with 3 to 4 second per utterance. Like between 100 - 200 utterances per > > speaker. What kind of adaptation technique should I use to get good > results > > per speaker? > > > > Best regards, > > Arif > > > > > ------------------------------------------------------------------------------ > > DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps > > OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access > > Free app hosting. Or install the open source package on any LAMP server. > > Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native! > > > http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk > > _______________________________________________ > > Kaldi-developers mailing list > > Kal...@li... > > https://lists.sourceforge.net/lists/listinfo/kaldi-developers > > > |
|
From: Christopher K. <chr...@a2...> - 2013-11-28 14:32:02
|
Hi, I am looking for the exact error. So far I did not reproduce the compilation error but I have an error during the tes phase : make -C transform test make[1]: Entering directory `/home/kermorvant/Tools/kaldi-code/src/transform' g++ -msse -msse2 -Wall -I.. -fPIC -DKALDI_DOUBLEPRECISION=0 -DHAVE_POSIX_MEMALIGN -Wno-sign-compare -Wno-unused-local-typedefs -Winit-self -DHAVE_EXECINFO_H=1 -rdynamic -DHAVE_CXXABI_H -DHAVE_ATLAS -I/home/kermorvant/Tools/kaldi-code/tools/ATLAS/include -I/home/kermorvant/Tools/kaldi-code/tools/openfst/include -g -DHAVE_CUDA -I/usr/local/cuda/include -c -o regtree-fmllr-diag-gmm-test.o regtree-fmllr-diag-gmm-test.cc g++ -rdynamic -Wl,-rpath=/home/kermorvant/Tools/kaldi-code/tools/openfst/lib -L/usr/local/cuda/lib64 -Wl,-rpath,/usr/local/cuda/lib64 regtree-fmllr-diag-gmm-test.o kaldi-transform.a ../gmm/kaldi-gmm.a ../tree/kaldi-tree.a ../util/kaldi-util.a ../matrix/kaldi-matrix.a ../base/kaldi-base.a -L/home/kermorvant/Tools/kaldi-code/tools/openfst/lib -lfst -L/usr/lib -llapack -lcblas -latlas -lf77blas -lm -lpthread -ldl -lcublas -lcudart -o regtree-fmllr-diag-gmm-test ../thread/kaldi-thread.h:61: error: undefined reference to 'vtable for kaldi::MultiThreadable' mle-diag-gmm.cc:490: error: undefined reference to 'kaldi::MultiThreadable::~MultiThreadable()' mle-diag-gmm.cc:495: error: undefined reference to 'kaldi::MultiThreadable::~MultiThreadable()' mle-diag-gmm.cc:514: error: undefined reference to 'kaldi::MultiThreadable::~MultiThreadable()' mle-diag-gmm.cc:514: error: undefined reference to 'kaldi::MultiThreadable::~MultiThreadable()' ../gmm/kaldi-gmm.a(mle-diag-gmm.o):mle-diag-gmm.cc:function typeinfo for kaldi::AccumulateMultiThreadedClass: error: undefined reference to 'typeinfo for kaldi::MultiThreadable' collect2: ld returned 1 exit status make[1]: *** [regtree-fmllr-diag-gmm-test] Error 1 make[1]: Leaving directory `/home/kermorvant/Tools/kaldi-code/src/transform' -- Chris Le 11/28/2013 07:16 AM, jen...@a2... a écrit : > Kaldi - Build # 270 - Failure: > > Check console output at http://jenkins.a2ialab.com/jenkins/job/Kaldi/270/ to view the results. > > > > ------------------------------------------------------------------------------ > Rapidly troubleshoot problems before they affect your business. Most IT > organizations don't have a clear picture of how application performance > affects their revenue. With AppDynamics, you get 100% visibility into your > Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics Pro! > http://pubads.g.doubleclick.net/gampad/clk?id=84349351&iu=/4140/ostg.clktrk > > > _______________________________________________ > Kaldi-developers mailing list > Kal...@li... > https://lists.sourceforge.net/lists/listinfo/kaldi-developers -- Christopher Kermorvant Responsable R&D - A2iA - France 39 rue de la Bienfaisance - 75008 Paris +33 (0) 144 420 086 / +33 (0) 689 513 601 (mobile) |
|
From: Saman A. <aza...@gm...> - 2013-11-28 09:54:16
|
Hi, I'm trying to compile kaldi on ubuntu 13.10 but while making openfst it says: ... /bin/bash ../../libtool --tag=CXX --mode=link g++ -g -O2 -o fstarcsort fstarcsort.o ../script/libfstscript.la ../lib/libfst.la -lm -ldl libtool: link: g++ -g -O2 -o .libs/fstarcsort fstarcsort.o ../script/.libs/libfstscript.so ../lib/.libs/libfst.so -lm -ldl -Wl,-rpath -Wl,/home/saman/Thesis/kaldi-trunk/tools/openfst-1.3.2/lib ../script/.libs/libfstscript.so: undefined reference to `dlopen' ../script/.libs/libfstscript.so: undefined reference to `dlerror' collect2: error: ld returned 1 exit status make[3]: *** [fstarcsort] Error 1 make[3]: Leaving directory `/home/saman/Thesis/kaldi-trunk/tools/openfst-1.3.2/src/bin' make[2]: *** [install-recursive] Error 1 make[2]: Leaving directory `/home/saman/Thesis/kaldi-trunk/tools/openfst-1.3.2/src' make[1]: *** [install-recursive] Error 1 make[1]: Leaving directory `/home/saman/Thesis/kaldi-trunk/tools/openfst-1.3.2' make: *** [openfst_compiled] Error 2 Seems there is a linker problem with libdl or a misorder GCC (or g++) option. How can I fix it? Thanks in advance for any help you are able to provide. |
|
From: <jen...@a2...> - 2013-11-28 06:55:25
|
Kaldi - Build # 270 - Failure: Check console output at http://jenkins.a2ialab.com/jenkins/job/Kaldi/270/ to view the results. |
|
From: Tuan D. <tua...@is...> - 2013-11-27 18:43:58
|
Dear all, I'm studying deep neural network in speech recognition and Kaldi is a very interesting and appropriate framework for my research. Now I have some problems with data preparation. For some reasons, I can't obtain LDC93S3A so I want to use another speech data instead. According to the tutorial, I have to prepare a folder /LDC/LDC93S3A/rm_comp which contains 3 sub folders: rm1_audio1, rm1_audio2 and rm2_audio. So what do the sub-folers contain? Where do you put audio files and text files in LDC data folder? I want to know more detail about the structure of the data folder. I very appreciate your helps. Best regards, Dinh Anh Tuan. |
|
From: Vesely K. <ve...@gm...> - 2013-11-22 18:25:14
|
Yes, last year during the Babel evaluations with Mirko we found out that, variance normalization can improve the results till the end of the recipe. It was in the particular case when we re-segmented using non-kaldi VAD, and when we computed cmvn stats only on speech frames. The difference was not big (especially at 'sat' stage and after), but still an improvement. With regards, Karel. On 11/22/2013 05:54 PM, Daniel Povey wrote: > Karel-- > can you give us some idea under what circumstances it helps to add the > --norm-vars true option? > Dan > > > ---------- Forwarded message ---------- > From: Repository Kaldi code <no...@co...> > Date: Fri, Nov 22, 2013 at 11:50 AM > Subject: [kaldi:code] [r3202] - vesis84: trunk: Adding possibility to > choose between cmn/cmvn by adding option --norm-vars, up to now the > default was cmn. Similarly to $splice_opts in case that further steps > rely on features from previous step, the normalization type is loaded > from $srcdir/norm_vars. The scritps are backward compatible, if no > value given the default is 'false'. > To: Repository Kaldi code <no...@co...> > > > trunk: Adding possibility to choose between cmn/cmvn by adding option > --norm-vars, up to now the default was cmn. Similarly to $splice_opts > in case that further steps rely on features from previous step, the > normalization type is loaded from $srcdir/norm_vars. The scritps are > backward compatible, if no value given the default is 'false'. > > http://sourceforge.net/p/kaldi/code/3202/ > > ________________________________ > > Sent from sourceforge.net because you indicated interest in > https://sourceforge.net/p/kaldi/code/ > > To unsubscribe from further messages, please visit > https://sourceforge.net/auth/subscriptions/ > > ------------------------------------------------------------------------------ > Shape the Mobile Experience: Free Subscription > Software experts and developers: Be at the forefront of tech innovation. > Intel(R) Software Adrenaline delivers strategic insight and game-changing > conversations that shape the rapidly evolving mobile landscape. Sign up now. > http://pubads.g.doubleclick.net/gampad/clk?id=63431311&iu=/4140/ostg.clktrk > _______________________________________________ > Kaldi-developers mailing list > Kal...@li... > https://lists.sourceforge.net/lists/listinfo/kaldi-developers |
|
From: Daniel P. <dp...@gm...> - 2013-11-22 16:54:25
|
Karel-- can you give us some idea under what circumstances it helps to add the --norm-vars true option? Dan ---------- Forwarded message ---------- From: Repository Kaldi code <no...@co...> Date: Fri, Nov 22, 2013 at 11:50 AM Subject: [kaldi:code] [r3202] - vesis84: trunk: Adding possibility to choose between cmn/cmvn by adding option --norm-vars, up to now the default was cmn. Similarly to $splice_opts in case that further steps rely on features from previous step, the normalization type is loaded from $srcdir/norm_vars. The scritps are backward compatible, if no value given the default is 'false'. To: Repository Kaldi code <no...@co...> trunk: Adding possibility to choose between cmn/cmvn by adding option --norm-vars, up to now the default was cmn. Similarly to $splice_opts in case that further steps rely on features from previous step, the normalization type is loaded from $srcdir/norm_vars. The scritps are backward compatible, if no value given the default is 'false'. http://sourceforge.net/p/kaldi/code/3202/ ________________________________ Sent from sourceforge.net because you indicated interest in https://sourceforge.net/p/kaldi/code/ To unsubscribe from further messages, please visit https://sourceforge.net/auth/subscriptions/ |
|
From: Vesely K. <ve...@gm...> - 2013-11-22 12:56:58
|
Hi, I was not able to reproduce the error. It could have been introduced by previous merging commit 3196. I downloaded fresh kaldi-trunk, compiled w/o cuda, and run the tests of src/cudamatrix on CentOS 6.4. In my case all the tests have passed well: ... Running cu-block-matrix-test ...... SUCCESS ... Can you please send a more detailed log, so we can see a precise line where the test was failing? Thanks, Karel On 11/22/2013 10:07 AM, Christopher Kermorvant wrote: > Hi, > > I have a broken test after > > *Revision 3197 by vesis84 > <http://jenkins.a2ialab.com/jenkins/user/vesis84/>: * > trunk,cudamatrix : added reset of cuda error state by > cudaGetLastError(), when allocator runs out of gpu memory before releasing > > The test sequence shows : > > Running cu-vector-test ...... SUCCESS > Running cu-matrix-test ...... SUCCESS > Running cu-math-test ...... SUCCESS > Running cu-test ...... SUCCESS > Running cu-sp-matrix-test ...... SUCCESS > Running cu-packed-matrix-test ...... SUCCESS > Running cu-tp-matrix-test ...... SUCCESS > Running cu-block-matrix-test ...... FAIL > Running cu-matrix-speed-test ...... SUCCESS > Running cu-vector-speed-test ...... SUCCESS > Running cu-sp-matrix-speed-test ...... SUCCESS > Running cu-array-test ...... SUCCESS > make[1]: *** [test] Error 1 > make[1]: Leaving directory `/var/lib/jenkins/sharedspace/kaldi-cpu/src/cudamatrix' > make: *** [cudamatrix/test] Error 2 > Build step 'Execute shell' marked build as failure > Email was triggered for: Failure > Sending email for trigger: Failure > Sending email to: ck...@a2... > Finished: FAILURE > > the compilation configuration is > > ./configure --use-cuda=no > --atlas-root="/share/apps/ubuntu/64/atlas-3.10.1/" --mathlib="ATLAS" ; > make depend > make -j20 > make test > > > Curiously, I have another build agent configured with > > ./configure; > make depend; > make -j2 > make test > > which is not broken. > > Do you reproduce this ? > |
|
From: Christopher K. <ck...@a2...> - 2013-11-22 09:07:54
|
Hi, I have a broken test after *Revision 3197 by vesis84 <http://jenkins.a2ialab.com/jenkins/user/vesis84/>: * trunk,cudamatrix : added reset of cuda error state by cudaGetLastError(), when allocator runs out of gpu memory before releasing The test sequence shows : Running cu-vector-test ...... SUCCESS Running cu-matrix-test ...... SUCCESS Running cu-math-test ...... SUCCESS Running cu-test ...... SUCCESS Running cu-sp-matrix-test ...... SUCCESS Running cu-packed-matrix-test ...... SUCCESS Running cu-tp-matrix-test ...... SUCCESS Running cu-block-matrix-test ...... FAIL Running cu-matrix-speed-test ...... SUCCESS Running cu-vector-speed-test ...... SUCCESS Running cu-sp-matrix-speed-test ...... SUCCESS Running cu-array-test ...... SUCCESS make[1]: *** [test] Error 1 make[1]: Leaving directory `/var/lib/jenkins/sharedspace/kaldi-cpu/src/cudamatrix' make: *** [cudamatrix/test] Error 2 Build step 'Execute shell' marked build as failure Email was triggered for: Failure Sending email for trigger: Failure Sending email to: ck...@a2... Finished: FAILURE the compilation configuration is ./configure --use-cuda=no --atlas-root="/share/apps/ubuntu/64/atlas-3.10.1/" --mathlib="ATLAS" ; make depend make -j20 make test Curiously, I have another build agent configured with ./configure; make depend; make -j2 make test which is not broken. Do you reproduce this ? -- Christopher Kermorvant |
|
From: Pirros T. <pi...@gm...> - 2013-11-20 16:23:50
|
|
From: Sakriani S. <ss...@is...> - 2013-11-19 04:50:39
|
OK, I'll try to check again. Thanks a lot!! -Sakti- On 2013/11/19 12:06, Daniel Povey wrote: > There does seem to be something called batch-size, but it defaults to > about 0.3 seconds (27 frames) which already seems pretty fast to me. > What may be going on is that you have a lot of language model context > and it takes a while for the decoder to be sure which words will > eventually "win" (you'll see in the code, the concept of an immortal > token, which is a token that all other tokens eventually trace back > to). > Dan > > > On Mon, Nov 18, 2013 at 9:57 PM, Sakriani Sakti <ss...@is...> wrote: >> Hi Daniel, >> >> When I try online decoding with: online-wav-gmm-decode-faster >> and give a very long wav file of TED talk. It seems it output the text >> gradually. >> But, I am not sure whether the delay is due to partial output mechanism. >> >> I checked the code inside, it has something like: >> if (dstate & (decoder.kEndFeats | decoder.kEndUtt)): >> decoder.FinishTraceBack(&out_fst) >> else: >> decoder.PartialTraceback(&out_fst) >> I am not so familiar with the code, but I guess the "else" part handle >> partial >> outputs. Is it correct? >> >> But, in online-audio-server-gmm-decode-faster, only >> decoder.FinishTraceBack(&out_fst) is used. >> >> If there is mechanism where we can explicitly set the output time interval >> of progressive output, >> that would be really helpful for us. At this moment, our ASR with Kaldi give >> much better >> performance compare with Julius, but unfortunately we are still not able to >> replace it >> due to this issues. >> >> -Sakti- >> >> >> >> >> On 2013/11/19 1:58, Daniel Povey wrote: >>> If you use the OnlineFasterDecoder, I've noticed that the code seems >>> to be designed to allow that, but I'm not sure what the interface is. >>> >>> Dan >>> >>> On Mon, Nov 18, 2013 at 4:13 AM, Sakriani Sakti <ss...@is...> >>> wrote: >>>> Hi Daniel, >>>> >>>> OK, thanks. We have reduce the rt-max. >>>> >>>> One more thing, I miss to ask one important point. >>>> In conventional speech translation systems, we usually use full sentences >>>> as the fundamental unit of processing. Normally, MT module is started >>>> after the ASR module finishes recognition of one sentence and the TTS >>>> module >>>> is started after MT module finishes translation, causing a delay between >>>> the >>>> start >>>> of the speaker’s utterance to the end of synthesis. >>>> >>>> As for simultaneous speech translation, the utterance (such as talk, >>>> lectures) >>>> is very long. Here, we need a mechanism that enable progressive output of >>>> the >>>> partial results. Julius decoder have this features, but we couldn't find >>>> in >>>> Kaldi decoding. >>>> >>>> So is there any way, to provide the text outputs before the end of speech >>>> utterances? >>>> >>>> -Sakti- >>>> >>>> >>>> >>>> On 2013/11/18 13:35, Daniel Povey wrote: >>>> >>>> I think you should probably reduce rt-max to 1.0 if you want it to >>>> decode in real-time. >>>> Dan >>>> >>>> >>>> On Sun, Nov 17, 2013 at 11:34 PM, Sakriani Sakti <ss...@is...> >>>> wrote: >>>> >>>> Hi Daniel and Vassil, >>>> >>>> Thank you for the response. >>>> >>>> We first learn from voxforge setup. As we need to put our ASR in the >>>> server >>>> and receive the input directly from the microphone, we currently use: >>>> >>>> onlinebin/online-audio-server-decode-faster >>>> >>>> with following parameters: >>>> beam=15.0 >>>> rt-min=0.5 >>>> rt-max=3.0 >>>> >>>> Our old system based on Julius decoder, but compare with Julius, >>>> the response with Kaldi online decoding is really slow. Please let >>>> us know how to make sure it stays within real time. >>>> >>>> Thank you. >>>> -Sakti- >>>> >>>> >>>> >>>> On 2013/11/17 1:04, Vassil Panayotov wrote: >>>> >>>> Yes it has some basic/experimental logic to automatically tweak the >>>> beam according to the current decoding real time >>>> factor(--rt-min/--rt-max parameters), but I don't think anyone has >>>> tested how well it works in practice(my impressions is that it's far >>>> from perfect). >>>> If this doesn't work for the original poster for some reason, the only >>>> other options that wouldn't require changing the code would be to >>>> tweak the beam or the maximum active states. >>>> >>>> Vassil >>>> >>>> On Sat, Nov 16, 2013 at 5:46 PM, Daniel Povey <dp...@gm...> wrote: >>>> >>>> When you say the online decoding, you mean e.g. the voxforge setup? >>>> the stuff in onlinebin/ ? I thought the online decoder had a >>>> mechanism to make sure it stays within real time. >>>> Dan >>>> >>>> >>>> On Fri, Nov 15, 2013 at 12:18 AM, Sakriani Sakti <ss...@is...> >>>> wrote: >>>> >>>> Dear Kaldi-developers, >>>> >>>> First of all, we would like to join Kaldi-users mailing list. Please >>>> kindly >>>> include me into Kaldi-users mailing list. >>>> >>>> We are currently build simultaneous speech translation, where we >>>> build the ASR system using Kaldi. To construct simultaneous speech >>>> translation, we really need real-time ASR decoding. Currently, the >>>> online decoding in Kaldi seems respond very slow and we couldn't find >>>> the way to multi-thread decoding as in offline case. Is there any way, >>>> to >>>> make the online decoding a little faster? >>>> >>>> Another things, the ctm format file mainly output the recognized word >>>> with time information. Is there any way to get more detail information, >>>> such as phone with time information or HMM-state with time information >>>> as in HTK decoding output? >>>> >>>> It would be great if you could help us for this. >>>> Thank you. >>>> Sincerely yours, >>>> Sakriani Sakti >>>> >>>> >>>> >>>> >>>> ------------------------------------------------------------------------------ >>>> DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps >>>> OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access >>>> Free app hosting. Or install the open source package on any LAMP server. >>>> Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native! >>>> >>>> >>>> http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk >>>> _______________________________________________ >>>> Kaldi-developers mailing list >>>> Kal...@li... >>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers >>>> >>>> >>>> ------------------------------------------------------------------------------ >>>> DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps >>>> OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access >>>> Free app hosting. Or install the open source package on any LAMP server. >>>> Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native! >>>> >>>> >>>> http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk >>>> _______________________________________________ >>>> Kaldi-developers mailing list >>>> Kal...@li... >>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers >>>> >>>> |
|
From: Daniel P. <dp...@gm...> - 2013-11-19 03:06:38
|
There does seem to be something called batch-size, but it defaults to about 0.3 seconds (27 frames) which already seems pretty fast to me. What may be going on is that you have a lot of language model context and it takes a while for the decoder to be sure which words will eventually "win" (you'll see in the code, the concept of an immortal token, which is a token that all other tokens eventually trace back to). Dan On Mon, Nov 18, 2013 at 9:57 PM, Sakriani Sakti <ss...@is...> wrote: > Hi Daniel, > > When I try online decoding with: online-wav-gmm-decode-faster > and give a very long wav file of TED talk. It seems it output the text > gradually. > But, I am not sure whether the delay is due to partial output mechanism. > > I checked the code inside, it has something like: > if (dstate & (decoder.kEndFeats | decoder.kEndUtt)): > decoder.FinishTraceBack(&out_fst) > else: > decoder.PartialTraceback(&out_fst) > I am not so familiar with the code, but I guess the "else" part handle > partial > outputs. Is it correct? > > But, in online-audio-server-gmm-decode-faster, only > decoder.FinishTraceBack(&out_fst) is used. > > If there is mechanism where we can explicitly set the output time interval > of progressive output, > that would be really helpful for us. At this moment, our ASR with Kaldi give > much better > performance compare with Julius, but unfortunately we are still not able to > replace it > due to this issues. > > -Sakti- > > > > > On 2013/11/19 1:58, Daniel Povey wrote: >> >> If you use the OnlineFasterDecoder, I've noticed that the code seems >> to be designed to allow that, but I'm not sure what the interface is. >> >> Dan >> >> On Mon, Nov 18, 2013 at 4:13 AM, Sakriani Sakti <ss...@is...> >> wrote: >>> >>> Hi Daniel, >>> >>> OK, thanks. We have reduce the rt-max. >>> >>> One more thing, I miss to ask one important point. >>> In conventional speech translation systems, we usually use full sentences >>> as the fundamental unit of processing. Normally, MT module is started >>> after the ASR module finishes recognition of one sentence and the TTS >>> module >>> is started after MT module finishes translation, causing a delay between >>> the >>> start >>> of the speaker’s utterance to the end of synthesis. >>> >>> As for simultaneous speech translation, the utterance (such as talk, >>> lectures) >>> is very long. Here, we need a mechanism that enable progressive output of >>> the >>> partial results. Julius decoder have this features, but we couldn't find >>> in >>> Kaldi decoding. >>> >>> So is there any way, to provide the text outputs before the end of speech >>> utterances? >>> >>> -Sakti- >>> >>> >>> >>> On 2013/11/18 13:35, Daniel Povey wrote: >>> >>> I think you should probably reduce rt-max to 1.0 if you want it to >>> decode in real-time. >>> Dan >>> >>> >>> On Sun, Nov 17, 2013 at 11:34 PM, Sakriani Sakti <ss...@is...> >>> wrote: >>> >>> Hi Daniel and Vassil, >>> >>> Thank you for the response. >>> >>> We first learn from voxforge setup. As we need to put our ASR in the >>> server >>> and receive the input directly from the microphone, we currently use: >>> >>> onlinebin/online-audio-server-decode-faster >>> >>> with following parameters: >>> beam=15.0 >>> rt-min=0.5 >>> rt-max=3.0 >>> >>> Our old system based on Julius decoder, but compare with Julius, >>> the response with Kaldi online decoding is really slow. Please let >>> us know how to make sure it stays within real time. >>> >>> Thank you. >>> -Sakti- >>> >>> >>> >>> On 2013/11/17 1:04, Vassil Panayotov wrote: >>> >>> Yes it has some basic/experimental logic to automatically tweak the >>> beam according to the current decoding real time >>> factor(--rt-min/--rt-max parameters), but I don't think anyone has >>> tested how well it works in practice(my impressions is that it's far >>> from perfect). >>> If this doesn't work for the original poster for some reason, the only >>> other options that wouldn't require changing the code would be to >>> tweak the beam or the maximum active states. >>> >>> Vassil >>> >>> On Sat, Nov 16, 2013 at 5:46 PM, Daniel Povey <dp...@gm...> wrote: >>> >>> When you say the online decoding, you mean e.g. the voxforge setup? >>> the stuff in onlinebin/ ? I thought the online decoder had a >>> mechanism to make sure it stays within real time. >>> Dan >>> >>> >>> On Fri, Nov 15, 2013 at 12:18 AM, Sakriani Sakti <ss...@is...> >>> wrote: >>> >>> Dear Kaldi-developers, >>> >>> First of all, we would like to join Kaldi-users mailing list. Please >>> kindly >>> include me into Kaldi-users mailing list. >>> >>> We are currently build simultaneous speech translation, where we >>> build the ASR system using Kaldi. To construct simultaneous speech >>> translation, we really need real-time ASR decoding. Currently, the >>> online decoding in Kaldi seems respond very slow and we couldn't find >>> the way to multi-thread decoding as in offline case. Is there any way, >>> to >>> make the online decoding a little faster? >>> >>> Another things, the ctm format file mainly output the recognized word >>> with time information. Is there any way to get more detail information, >>> such as phone with time information or HMM-state with time information >>> as in HTK decoding output? >>> >>> It would be great if you could help us for this. >>> Thank you. >>> Sincerely yours, >>> Sakriani Sakti >>> >>> >>> >>> >>> ------------------------------------------------------------------------------ >>> DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps >>> OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access >>> Free app hosting. Or install the open source package on any LAMP server. >>> Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native! >>> >>> >>> http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk >>> _______________________________________________ >>> Kaldi-developers mailing list >>> Kal...@li... >>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers >>> >>> >>> ------------------------------------------------------------------------------ >>> DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps >>> OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access >>> Free app hosting. Or install the open source package on any LAMP server. >>> Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native! >>> >>> >>> http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk >>> _______________________________________________ >>> Kaldi-developers mailing list >>> Kal...@li... >>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers >>> >>> > |
|
From: Sakriani S. <ss...@is...> - 2013-11-19 02:57:16
|
Hi Daniel,
When I try online decoding with: online-wav-gmm-decode-faster
and give a very long wav file of TED talk. It seems it output the text
gradually.
But, I am not sure whether the delay is due to partial output mechanism.
I checked the code inside, it has something like:
if (dstate & (decoder.kEndFeats | decoder.kEndUtt)):
decoder.FinishTraceBack(&out_fst)
else:
decoder.PartialTraceback(&out_fst)
I am not so familiar with the code, but I guess the "else" part handle
partial
outputs. Is it correct?
But, in online-audio-server-gmm-decode-faster, only
decoder.FinishTraceBack(&out_fst) is used.
If there is mechanism where we can explicitly set the output time
interval of progressive output,
that would be really helpful for us. At this moment, our ASR with Kaldi
give much better
performance compare with Julius, but unfortunately we are still not able
to replace it
due to this issues.
-Sakti-
On 2013/11/19 1:58, Daniel Povey wrote:
> If you use the OnlineFasterDecoder, I've noticed that the code seems
> to be designed to allow that, but I'm not sure what the interface is.
>
> Dan
>
> On Mon, Nov 18, 2013 at 4:13 AM, Sakriani Sakti <ss...@is...> wrote:
>> Hi Daniel,
>>
>> OK, thanks. We have reduce the rt-max.
>>
>> One more thing, I miss to ask one important point.
>> In conventional speech translation systems, we usually use full sentences
>> as the fundamental unit of processing. Normally, MT module is started
>> after the ASR module finishes recognition of one sentence and the TTS module
>> is started after MT module finishes translation, causing a delay between the
>> start
>> of the speaker’s utterance to the end of synthesis.
>>
>> As for simultaneous speech translation, the utterance (such as talk,
>> lectures)
>> is very long. Here, we need a mechanism that enable progressive output of
>> the
>> partial results. Julius decoder have this features, but we couldn't find in
>> Kaldi decoding.
>>
>> So is there any way, to provide the text outputs before the end of speech
>> utterances?
>>
>> -Sakti-
>>
>>
>>
>> On 2013/11/18 13:35, Daniel Povey wrote:
>>
>> I think you should probably reduce rt-max to 1.0 if you want it to
>> decode in real-time.
>> Dan
>>
>>
>> On Sun, Nov 17, 2013 at 11:34 PM, Sakriani Sakti <ss...@is...> wrote:
>>
>> Hi Daniel and Vassil,
>>
>> Thank you for the response.
>>
>> We first learn from voxforge setup. As we need to put our ASR in the server
>> and receive the input directly from the microphone, we currently use:
>>
>> onlinebin/online-audio-server-decode-faster
>>
>> with following parameters:
>> beam=15.0
>> rt-min=0.5
>> rt-max=3.0
>>
>> Our old system based on Julius decoder, but compare with Julius,
>> the response with Kaldi online decoding is really slow. Please let
>> us know how to make sure it stays within real time.
>>
>> Thank you.
>> -Sakti-
>>
>>
>>
>> On 2013/11/17 1:04, Vassil Panayotov wrote:
>>
>> Yes it has some basic/experimental logic to automatically tweak the
>> beam according to the current decoding real time
>> factor(--rt-min/--rt-max parameters), but I don't think anyone has
>> tested how well it works in practice(my impressions is that it's far
>> from perfect).
>> If this doesn't work for the original poster for some reason, the only
>> other options that wouldn't require changing the code would be to
>> tweak the beam or the maximum active states.
>>
>> Vassil
>>
>> On Sat, Nov 16, 2013 at 5:46 PM, Daniel Povey <dp...@gm...> wrote:
>>
>> When you say the online decoding, you mean e.g. the voxforge setup?
>> the stuff in onlinebin/ ? I thought the online decoder had a
>> mechanism to make sure it stays within real time.
>> Dan
>>
>>
>> On Fri, Nov 15, 2013 at 12:18 AM, Sakriani Sakti <ss...@is...>
>> wrote:
>>
>> Dear Kaldi-developers,
>>
>> First of all, we would like to join Kaldi-users mailing list. Please
>> kindly
>> include me into Kaldi-users mailing list.
>>
>> We are currently build simultaneous speech translation, where we
>> build the ASR system using Kaldi. To construct simultaneous speech
>> translation, we really need real-time ASR decoding. Currently, the
>> online decoding in Kaldi seems respond very slow and we couldn't find
>> the way to multi-thread decoding as in offline case. Is there any way,
>> to
>> make the online decoding a little faster?
>>
>> Another things, the ctm format file mainly output the recognized word
>> with time information. Is there any way to get more detail information,
>> such as phone with time information or HMM-state with time information
>> as in HTK decoding output?
>>
>> It would be great if you could help us for this.
>> Thank you.
>> Sincerely yours,
>> Sakriani Sakti
>>
>>
>>
>> ------------------------------------------------------------------------------
>> DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps
>> OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access
>> Free app hosting. Or install the open source package on any LAMP server.
>> Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native!
>>
>> http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk
>> _______________________________________________
>> Kaldi-developers mailing list
>> Kal...@li...
>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>
>> ------------------------------------------------------------------------------
>> DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps
>> OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access
>> Free app hosting. Or install the open source package on any LAMP server.
>> Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native!
>>
>> http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk
>> _______________________________________________
>> Kaldi-developers mailing list
>> Kal...@li...
>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>
>>
|
|
From: Daniel P. <dp...@gm...> - 2013-11-18 16:58:51
|
If you use the OnlineFasterDecoder, I've noticed that the code seems to be designed to allow that, but I'm not sure what the interface is. Dan On Mon, Nov 18, 2013 at 4:13 AM, Sakriani Sakti <ss...@is...> wrote: > Hi Daniel, > > OK, thanks. We have reduce the rt-max. > > One more thing, I miss to ask one important point. > In conventional speech translation systems, we usually use full sentences > as the fundamental unit of processing. Normally, MT module is started > after the ASR module finishes recognition of one sentence and the TTS module > is started after MT module finishes translation, causing a delay between the > start > of the speaker’s utterance to the end of synthesis. > > As for simultaneous speech translation, the utterance (such as talk, > lectures) > is very long. Here, we need a mechanism that enable progressive output of > the > partial results. Julius decoder have this features, but we couldn't find in > Kaldi decoding. > > So is there any way, to provide the text outputs before the end of speech > utterances? > > -Sakti- > > > > On 2013/11/18 13:35, Daniel Povey wrote: > > I think you should probably reduce rt-max to 1.0 if you want it to > decode in real-time. > Dan > > > On Sun, Nov 17, 2013 at 11:34 PM, Sakriani Sakti <ss...@is...> wrote: > > Hi Daniel and Vassil, > > Thank you for the response. > > We first learn from voxforge setup. As we need to put our ASR in the server > and receive the input directly from the microphone, we currently use: > > onlinebin/online-audio-server-decode-faster > > with following parameters: > beam=15.0 > rt-min=0.5 > rt-max=3.0 > > Our old system based on Julius decoder, but compare with Julius, > the response with Kaldi online decoding is really slow. Please let > us know how to make sure it stays within real time. > > Thank you. > -Sakti- > > > > On 2013/11/17 1:04, Vassil Panayotov wrote: > > Yes it has some basic/experimental logic to automatically tweak the > beam according to the current decoding real time > factor(--rt-min/--rt-max parameters), but I don't think anyone has > tested how well it works in practice(my impressions is that it's far > from perfect). > If this doesn't work for the original poster for some reason, the only > other options that wouldn't require changing the code would be to > tweak the beam or the maximum active states. > > Vassil > > On Sat, Nov 16, 2013 at 5:46 PM, Daniel Povey <dp...@gm...> wrote: > > When you say the online decoding, you mean e.g. the voxforge setup? > the stuff in onlinebin/ ? I thought the online decoder had a > mechanism to make sure it stays within real time. > Dan > > > On Fri, Nov 15, 2013 at 12:18 AM, Sakriani Sakti <ss...@is...> > wrote: > > Dear Kaldi-developers, > > First of all, we would like to join Kaldi-users mailing list. Please > kindly > include me into Kaldi-users mailing list. > > We are currently build simultaneous speech translation, where we > build the ASR system using Kaldi. To construct simultaneous speech > translation, we really need real-time ASR decoding. Currently, the > online decoding in Kaldi seems respond very slow and we couldn't find > the way to multi-thread decoding as in offline case. Is there any way, > to > make the online decoding a little faster? > > Another things, the ctm format file mainly output the recognized word > with time information. Is there any way to get more detail information, > such as phone with time information or HMM-state with time information > as in HTK decoding output? > > It would be great if you could help us for this. > Thank you. > Sincerely yours, > Sakriani Sakti > > > > ------------------------------------------------------------------------------ > DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps > OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access > Free app hosting. Or install the open source package on any LAMP server. > Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native! > > http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk > _______________________________________________ > Kaldi-developers mailing list > Kal...@li... > https://lists.sourceforge.net/lists/listinfo/kaldi-developers > > ------------------------------------------------------------------------------ > DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps > OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access > Free app hosting. Or install the open source package on any LAMP server. > Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native! > > http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk > _______________________________________________ > Kaldi-developers mailing list > Kal...@li... > https://lists.sourceforge.net/lists/listinfo/kaldi-developers > > |
|
From: Sakriani S. <ss...@is...> - 2013-11-18 09:13:37
|
Hi Daniel, OK, thanks. We have reduce the rt-max. One more thing, I miss to ask one important point. In conventional speech translation systems, we usually use full sentences as the fundamental unit of processing. Normally, MT module is started after the ASR module finishes recognition of one sentence and the TTS module is started after MT module finishes translation, causing a delay between the start of the speaker's utterance to the end of synthesis. As for simultaneous speech translation, the utterance (such as talk, lectures) is very long. Here, we need a mechanism that enable progressive output of the partial results. Julius decoder have this features, but we couldn't find in Kaldi decoding. So is there any way, to provide the text outputs before the end of speech utterances? -Sakti- On 2013/11/18 13:35, Daniel Povey wrote: > I think you should probably reduce rt-max to 1.0 if you want it to > decode in real-time. > Dan > > > On Sun, Nov 17, 2013 at 11:34 PM, Sakriani Sakti<ss...@is...> wrote: >> Hi Daniel and Vassil, >> >> Thank you for the response. >> >> We first learn from voxforge setup. As we need to put our ASR in the server >> and receive the input directly from the microphone, we currently use: >> >> onlinebin/online-audio-server-decode-faster >> >> with following parameters: >> beam=15.0 >> rt-min=0.5 >> rt-max=3.0 >> >> Our old system based on Julius decoder, but compare with Julius, >> the response with Kaldi online decoding is really slow. Please let >> us know how to make sure it stays within real time. >> >> Thank you. >> -Sakti- >> >> >> >> On 2013/11/17 1:04, Vassil Panayotov wrote: >>> Yes it has some basic/experimental logic to automatically tweak the >>> beam according to the current decoding real time >>> factor(--rt-min/--rt-max parameters), but I don't think anyone has >>> tested how well it works in practice(my impressions is that it's far >>> from perfect). >>> If this doesn't work for the original poster for some reason, the only >>> other options that wouldn't require changing the code would be to >>> tweak the beam or the maximum active states. >>> >>> Vassil >>> >>> On Sat, Nov 16, 2013 at 5:46 PM, Daniel Povey<dp...@gm...> wrote: >>>> When you say the online decoding, you mean e.g. the voxforge setup? >>>> the stuff in onlinebin/ ? I thought the online decoder had a >>>> mechanism to make sure it stays within real time. >>>> Dan >>>> >>>> >>>> On Fri, Nov 15, 2013 at 12:18 AM, Sakriani Sakti<ss...@is...> >>>> wrote: >>>>> Dear Kaldi-developers, >>>>> >>>>> First of all, we would like to join Kaldi-users mailing list. Please >>>>> kindly >>>>> include me into Kaldi-users mailing list. >>>>> >>>>> We are currently build simultaneous speech translation, where we >>>>> build the ASR system using Kaldi. To construct simultaneous speech >>>>> translation, we really need real-time ASR decoding. Currently, the >>>>> online decoding in Kaldi seems respond very slow and we couldn't find >>>>> the way to multi-thread decoding as in offline case. Is there any way, >>>>> to >>>>> make the online decoding a little faster? >>>>> >>>>> Another things, the ctm format file mainly output the recognized word >>>>> with time information. Is there any way to get more detail information, >>>>> such as phone with time information or HMM-state with time information >>>>> as in HTK decoding output? >>>>> >>>>> It would be great if you could help us for this. >>>>> Thank you. >>>>> Sincerely yours, >>>>> Sakriani Sakti >>>>> >>>>> >>>>> >>>>> ------------------------------------------------------------------------------ >>>>> DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps >>>>> OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access >>>>> Free app hosting. Or install the open source package on any LAMP server. >>>>> Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native! >>>>> >>>>> http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk >>>>> _______________________________________________ >>>>> Kaldi-developers mailing list >>>>> Kal...@li... >>>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers >>>> ------------------------------------------------------------------------------ >>>> DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps >>>> OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access >>>> Free app hosting. Or install the open source package on any LAMP server. >>>> Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native! >>>> >>>> http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk >>>> _______________________________________________ >>>> Kaldi-developers mailing list >>>> Kal...@li... >>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers |
|
From: Daniel P. <dp...@gm...> - 2013-11-18 04:35:39
|
I think you should probably reduce rt-max to 1.0 if you want it to decode in real-time. Dan On Sun, Nov 17, 2013 at 11:34 PM, Sakriani Sakti <ss...@is...> wrote: > Hi Daniel and Vassil, > > Thank you for the response. > > We first learn from voxforge setup. As we need to put our ASR in the server > and receive the input directly from the microphone, we currently use: > > onlinebin/online-audio-server-decode-faster > > with following parameters: > beam=15.0 > rt-min=0.5 > rt-max=3.0 > > Our old system based on Julius decoder, but compare with Julius, > the response with Kaldi online decoding is really slow. Please let > us know how to make sure it stays within real time. > > Thank you. > -Sakti- > > > > On 2013/11/17 1:04, Vassil Panayotov wrote: >> >> Yes it has some basic/experimental logic to automatically tweak the >> beam according to the current decoding real time >> factor(--rt-min/--rt-max parameters), but I don't think anyone has >> tested how well it works in practice(my impressions is that it's far >> from perfect). >> If this doesn't work for the original poster for some reason, the only >> other options that wouldn't require changing the code would be to >> tweak the beam or the maximum active states. >> >> Vassil >> >> On Sat, Nov 16, 2013 at 5:46 PM, Daniel Povey <dp...@gm...> wrote: >>> >>> When you say the online decoding, you mean e.g. the voxforge setup? >>> the stuff in onlinebin/ ? I thought the online decoder had a >>> mechanism to make sure it stays within real time. >>> Dan >>> >>> >>> On Fri, Nov 15, 2013 at 12:18 AM, Sakriani Sakti <ss...@is...> >>> wrote: >>>> >>>> Dear Kaldi-developers, >>>> >>>> First of all, we would like to join Kaldi-users mailing list. Please >>>> kindly >>>> include me into Kaldi-users mailing list. >>>> >>>> We are currently build simultaneous speech translation, where we >>>> build the ASR system using Kaldi. To construct simultaneous speech >>>> translation, we really need real-time ASR decoding. Currently, the >>>> online decoding in Kaldi seems respond very slow and we couldn't find >>>> the way to multi-thread decoding as in offline case. Is there any way, >>>> to >>>> make the online decoding a little faster? >>>> >>>> Another things, the ctm format file mainly output the recognized word >>>> with time information. Is there any way to get more detail information, >>>> such as phone with time information or HMM-state with time information >>>> as in HTK decoding output? >>>> >>>> It would be great if you could help us for this. >>>> Thank you. >>>> Sincerely yours, >>>> Sakriani Sakti >>>> >>>> >>>> >>>> ------------------------------------------------------------------------------ >>>> DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps >>>> OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access >>>> Free app hosting. Or install the open source package on any LAMP server. >>>> Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native! >>>> >>>> http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk >>>> _______________________________________________ >>>> Kaldi-developers mailing list >>>> Kal...@li... >>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers >>> >>> >>> ------------------------------------------------------------------------------ >>> DreamFactory - Open Source REST & JSON Services for HTML5 & Native Apps >>> OAuth, Users, Roles, SQL, NoSQL, BLOB Storage and External API Access >>> Free app hosting. Or install the open source package on any LAMP server. >>> Sign up and see examples for AngularJS, jQuery, Sencha Touch and Native! >>> >>> http://pubads.g.doubleclick.net/gampad/clk?id=63469471&iu=/4140/ostg.clktrk >>> _______________________________________________ >>> Kaldi-developers mailing list >>> Kal...@li... >>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers > > |
23 messages has been excluded from this view by a project administrator.
