From: apurv t. <apu...@gm...> - 2011-05-11 10:33:09
|
Hi Everyone, @James Baker: Thanks for the encouragement. I will keep those suggestions in mind while working on this project. @Nickolay : 19:00 UTC works fine for me as well. @Michal: Hi ! :-) Congrats for the project. We should get in touch soon. I am glad to be working on this project and with this team ! Regards, - Apurv Tiwari On Wed, May 11, 2011 at 3:14 AM, Nickolay V. Shmyrev <nsh...@ne...>wrote: > В Втр, 10/05/2011 в 13:29 -0400, James Baker пишет: > > Hello Michal and Apurv, > > > > I don't know if you have seen each other's messages or posting. If > > not, let this email be your introduction to each other. > > > > You are working on two closely related, but complementary projects. > > You are both working on problems related to training acoustic models > > using long audio files. You both must work on implementation issues > > that occur with any work on long audio files, including reducing > > memory requirements. > > > > Beyond that, Michal's main objective is to speed up processing by > > porting acoustic model training algorithms to a GPGPU environment > > using CUDA. This speed will, for example, help us to be able to > > process the 15,000 hours of audio data on librivox.org. > > > > Apurv's objective is to make the acoustic model training robust > > against script errors, disfluencies and other disturbances that affect > > any acoustic model training, but that are more severe with long audio > > files. This robustness is needed to help us handle many real-world > > sources of data. > > > > For Michal, we have suggested a memory saving trick that works for > > hidden Markov process processing in general. It reduces the memory > > requirement to be O(sqrt(T)) and appears to be appropriate to > > implementation on parallel processors. > > > > For Apruv, we have suggested a memory saving trick that, for speech > > alignment computations, reduces the average memory requirement to > > O(1), but that does not seem to be reasonable for a parallel > > processing implementation. > > > > You should get to know each other and share your work wherever that > > makes you more efficient. > > > > Nickolay and I, and other members of the Sphinx team, are available to > > answer questions and to make suggestions to help you be more > > productive. This is not a course homework assignment in which you > > must do all the work by yourself. This is a real-world team > > development effort where we are all working together to produce the > > best end result. > > Very true. > > Btw Apurv and Michal, I'm planning a quick voice chat between us on > Skype somewhere over the weekend. An evening around 7PM UTC works for > me. Please provide the information when you can be available. > > |