Minutes for 12/21/06 RT teleconference

- Role call
  - NIST: Fiscus, Bowers, Rose, Garofolo
  - ICSI: Janin,
  - IBM: Makis, Etienne
  - LIMSI: Lamel, Barras, Eric ?.
  - LDC: Glenn
  - UKA: Sebastien ?.
  - SRI: Stolke
  - CMU: Burger


- Old Issues:
  - Coffee break discussion
    - No one had reviewed the segments
    - Makis: What's the point and how does it relate to the lecture
      domain?
    - Bowers: From a CLEAR point, it's important for meeting dynamics
      and side information
    - JG: The danger is it will increase error rates, change the task,
      and there's not enough data to make the measurement informative
    - Lamel: There's also not enough training resources

    Decision:  Since there are arguments on both side, we decided to
       make it a completely separate meeting sub domain and make
       participation voluntary.

  - SASTT task definition
    - No negative comments about the task definition so we will move
      forward with the task

   ACTION: NIST will publish a schedule for software release.

    - Stolcke: Can a "baseline" system be constructed from last year's
      submissions?
    - Fiscus: Yes, NIST will probably do that during testing but there
      may be some problems.

  - Evaluation data
    - AMI data: when will it arrive?
    - Fiscus: It arrived and Adam didn't look at it.  :)    

  - Pending email discussion of close talking mic condition:
    "Can manual segmentation be a primary condition for the IHM tests?"
    - The discussion focused on trying to get to the real issue and
      that is the IHM task was expected to be "easier" since the head
      mics removed crosstalk but they don't in reality.
    - Adding the reference segmentation condition solves crosstalk
      problems during periods without foreground speech, but not
      during competing background/foreground speech
      - Stolcke: But this takes care of most of the problem because
        the background speech has much less energy and therefore the
        foreground speech dominates
    - General consensus that a better solution is to have a single
      site provide a "standard" segmentation output or even better,
      provide a crosstalk-removed version of the test signals.

    Action: Can anyone provide either?  The producer must be willing to
      process some old test sets to provide development resources.

  - Pending email discussion to keep MDM condition the primary condition
    JF: Aside from the change in primary condition, the most important
        thing is to be able to compare across sub-domains.  For Confmtg
        data MDM=SDM, but for lecture data it does not.
    JF/JG: Changing the primary condition will make historical
        comparisons difficult
    Janin: agree
    Lamel: coherence is important
    Stolcke: Fine
    Sebastien: won't comment since McDonough wasn't there
    IBM: as email states they want it to stay the same

    Decision :  Since most people agreed to keep it MDM, the primary
       condition will remain MDM.

- New Issues:
  - Inconsistent conference dates on web site.  RT is May 10-11, 2007.
    CLEAR is May 8-9, 2005

    Noted.
    - Forgot to mention the meeting will be in Baltimore.  More details
      to follow.      

  - Schedule next telecon

    Jan 16, 2007, 11:00 EST

Uncompleted action items to keep in mind:
  - NIST publishes the Forced Aligned reference files from RT-05 ands
    RT-06
  - CHIL will report on the soon-to-be-released DEV data and it will be
    documented on a web site
  - Adam J. to release "quick-and-dirty" scripts for processing AMI
    transcripts

    Done: See http://icsi.berkeley.edu/~janin/rt07s

  - Adam J. to circulate suggested AMI Dev/Train divide