Find topic
WS06 topics
Members' area
Tools
Help!
-- KarenLivescu - 15 Dec 2005
|
WS06
>
ProjectMeetingsJun4
Second official planning meeting, Sunday, June 4, 9am-5pm, Newark International Airport Holiday Inn, NJ
- Local information
- The hotel
- There should be a 24-hour free shuttle from the airport to the hotel.
- Present:
- Team members: Karen, Ozgur, Mark, Simon, Nash, Chris, Partha, Lisa, Ari, Bronwyn, Steve
- Satellite/advisory members: Jeff Bilmes (thanks!)
- Presentation slides:
- Karen: Overview; updates on manual transcriptions, AVSR work at MIT, distributed tools, SVitchboard baseline work; discussion, timeline, to-do slides. PPT and PDF.
- Nash: Analysis of manual transcriptions PPT
- Simon: Update on neural network AF classifiers, gmtkTie PPT
- Ozgur: Visual AF classifiers, SVM AF classifiers PPT
- Mark: AVICAR work PPT
Summary of activity since the first planning meeting
- NN AF classifiers trained (Joe, Mathew, Simon)
- SVM AF classifier work (Ozgur)
- NN AF classifiers for video (Ozgur)
- Manual feature transcriptions ongoing (Karen, Xuemin, Lisa)
- Transcriber agreement measures (Nash, Lisa)
- AVSR work on MIT webcam digits (Kate, Karen)
- AVSR work on AVICAR (Mark)
- Distributed tools for GMTK training/decoding (Karen, Partha, Chris, Simon)
- SVitchboard phone baseline updated to be “WS06-ready” (Karen)
- gmtkTie work (Simon)
Main results of meeting
- Decided on rough timeline for workshop
- Assigned people to sub-projects (Hybrid models; tandem & fully-generative models; pronunciation modeling; data analysis; audio-visual speech recognition)
- Decided on rough topics for undergrad projects and assigned mentors (Steve/Simon, Ari/Karen, Bronwyn/Mark)
- Divided up main tasks to be done before workshop:
- Finish manual transcriptions + basic analysis (Karen Nash Lisa)
- Polish distributed scripts (Chris Simon Partha Karen) and finish testing (Partha Simon Mark Arthur Chris Lisa Nash)
- Complete SVitchboard phone-based baselines (Karen Chris Ozgur); i.e.,
- Karen packages up and posts her current phone-based baselines
- Chris modifies Karen’s baselines to use training & pronunciation lattices
- Chris and Ozgur build triphone models
- Complete feature-based generative baseline, + a version with factored obs model (Karen)
- Run trained NN AF classifiers on SVitchboard (Simon)
- Train NN AF classifiers on SVitchboard, i.e.
- Generate phone alignments for SVitchboard using an existing HMM system (Ozgur)
- Run NN AF classifier training (Simon)
- Simon, Karen, Mark provide info & papers for Steve, Ari, & Bronwyn's projects
- Steve, Ari, Bronwyn read up on project background and tools/languages needed (at least Perl).
- See slides above for some more details
-- KarenLivescu - 30 May 2006
|
WS06_planning_meeting2_Karen.pdf 
WS06_planning_meeting2_Karen.ppt