Minutes from RT Teleconference Oct. 30, 2008
Role call
Jon Fiscus - NIST
Travis Rose - NIST
Nick Evans - Eurocom
Meghan Glenn - LDC
Andreas Stolcke - SRI/ICSI
Dr. Kapuska
David van Leeuwe
Florian Metze
Bin Ma
Mike Lincoln
Friedrich Faubel (friedrich.faubel@lsv.uni-saarland.de)
Dietrich Klakow (dietrich.klakow@lsv.uni-saarland.de)
Administrative:
- The RT mailing list will change to rt09@nist.gov. rt07 will be
discontinued.
- Question: Will RT be combined with RT?
- Ans: No. Lack of funding/data. RT will focus on meeting data.
- Question: Will there be any lecture data?
- Question: The Call for Participation omitted the IHM condition,
was this on purpose?
- Ans: No. It should have been included.
Data
Resources
- JF: To get access to new data, need to have license
agreements. Previously signed agreements are sufficient for previous
participants until the RT-09 agreements are ready. New participants
must sign the RT-07 agreements too get training resources and then they
must sign the RT-09 agreements when there available.
- JF: NIST will not be supplying data already in the LDC catalog.
Participants must contact them directly.
- Video data for the RT-06 data is available. Send NIST a disk
in order to get it. The RT-07 video data will be ready in a month.
- Evaluation Data Contributions
- NIST will prepare
a Conference data test set similar to last year.
- Data will be
slightly different than last year. The AMIDA data includes
teleconference phone:
- Mike L: There is
a telecon phone, but the remote person has a head mic and there is
a mic array in the
room as well.
- Andreas S.: The
issue from the last evaluation with telecon phones involved the IHM
condition where the
speaker had no corresponding IHM channel. This raises the issue of how
IHM is treated: as a contrast condition, or as a technology application
of interest.
- JF: One way to
handle it would be to exclude remote participant speech from IHM
scoring.
- JF: Since Thomas
H. wasn't on the phone, JF thought the issue of how to handle the IHM condition should be
tabled.
- Action
Item: Discuss the issue next telecon.
- DvL.: TNO has some Dutch meeting data that might be useful for
diarization tests. Would the
community be interested in working with this data?
- New Training resources: NIST can do modest amount of Quick
Transcriptions for new NIST training resources. This is to give people
experience with our new data which includes 4 NIST Mark III arrays and
higher resolution cameras.
- New idea: Online Diarization
- DVL: In the AMIDA project, they are interested in online
treatment of meeting data. For
instance a remote connection join in and already see results. It's a
different problem from offline speaker diarization. He proposes
doing an online diarization task, without using info from the future.
Perhaps allow a 1-second lag prior to output for example. This
would be closer to real-time applications.
- AS: How would we specify realtime arrival? It's hard to
simulate the condition?
- DVL: Make it a 'realistic' simulation based on look ahead.
- JF: We've administered similar evaluations that have been
successful.
- Action
Item: Propose to specify this condition in the eval plan for all
tasks. Only allow information you
have processed up to N seconds in the future.
- New Idea: Explicit context switching (Dr. Kapuska, please
check/correct this)
- Dr. Kapuska: In meetings, it would be useful to know when a
person is addressing a
person, addressing a computing resource, or talking about someone. Does
the data support this?
- JF: No. That would require a Wizard of Oz setup to test and the
meetings we use are targeted at human-human communication.
- Beam Forming Assessment task:
- JF: Two ways to
evaluate beamforming: (1) evaluate the benefit to an application (STT) (2) measure the signal
directly. This task focuses on method (2).
- JF: The blind source separation community uses a number of
metrics to measure signal quality. Would like to employ them on
this dataset. Has anyone tried to assess quality? (No one has).
- Action
Item: JF will send out a list of known metrics.
- Leveraging multimodal/mic array output: What can be done to share
processed data?
- Xavi's BeamformIt tool: http://www.icsi.berkeley.edu/~xanguera/beamformit/index.html
Evaluation
Schedule
- Eval data released:
Mar 2
- Submissions due: Mar
27
- Results Released:
April 10
- Workshop: Late May
(We need a venue. Any ideas/voluteers?)
Other
items:
- Next telecon:
11/20/2008 (NIST has a confilict the week of Nov. 11).
Adjourn