Linguistic Resources for Meeting Recognition Meghan Glenn, Stephanie Strassel Linguistic Data Consor

1
Linguistic Resources for Meeting Recognition
Meghan Glenn, Stephanie Strassel Linguistic Data
Consortiummlglenn, strassel_at_ldc.upenn.eduht
tp//www.ldc.upenn.edu/Projects/NISTMeet
2
Scope of Work

• Training data
• (Pre-publication) distribution
• Conference room test data
• Transcription
• Careful
• Quick
• Comparison and analysis
• Infrastructure
• XTrans Toolkit
• Features for meetings

3
RT-05S Training Datadistributed by LDC

• (Pre-publication) distribution via e-corpus to
  RT-05 participants
• All available from www.ldc.upenn.edu/Catalog

4
RT-05S Evaluation Datatranscribed by LDC

• Conference room data
• Ten meeting sessions, 12 minutes each
• Contributed by five sites
• Multiple recording conditions for each session
• Primarily business meeting content
• Transcribers report it was faster, easier and
  more interesting to transcribe than RT-04 meeting
  eval data
• All data carefully transcribed (CTR)
• Half of data quickly transcribed (QTR)
• for contrastive study

5
CTR Process

• Using IHM channels
• One exception participant on speakerphone
• 1st pass manual segmentation
• Turns ? breath groups
• 3-8 seconds per segment, designed for ease of
  transcription only
• 10 ms padding around each segment boundary
• No segmentation or transcription of isolated
  speaker noise
• 2nd pass initial verbatim transcription
• No time limit
• Goal is to get everything right
• 3rd pass verify existing transcription and
  timestamps, add additional markup
• Indicate proper names, filled pauses, noise, etc.
• Revisit difficult sections

6
CTR Quality Control

• Additional QC pass by lead transcriber
• Using mixed IHM recordings and/or SDM
• Merge individual transcripts
• Speaker assignment
• Transcription accuracy, completeness
• Markup consistency
• Spell check
• Syntax (format) check
• Check consistency and accuracy of names,
  acronyms, terminology
• Check silence (untranscribed) regions for missed
  speech using customized tool

7
QTR Process

• 0th pass automatic audio segmentation
• Pause detection algorithm
• No manual correction
• 1st pass verbatim transcription
• Limited to five times real time
• Goal is to get the words right only
• No special markup, orthography or capitalization
• No extra time spent on difficult sections (e.g.,
  disfluencies)
• QC pass minimal, semi-automated
• Spell check
• Format check
• No check of transcript content, consistency of
  names/terms, etc.

8
CTR vs. QTR
9
Example
10
Unique Challenges

• Many speakers takes longer to transcribe!
• Impact of overlapping speech, even using IHM
  audio
• Varying levels of speaker participation
• Often no speech but other speaker/background
  noise
• Meeting content
• All over the map, from games to technical
  meetings
• Lack of customized transcription tools
• Existing tools optimized for
• 1-channel, multispeaker per channel (BN)
• 2-channel, one speaker per channel (CTS)
• Needed a tool that merges features of each
• Arbitrary number of channels, speakers
• Easily move between mixed and individual signal
  playback
• Access to video would also help disambiguate

11
XTrans

• Multipurpose speech annotation tool
• Multilingual, multi-platform
• written in Python
• AGTK infrastructure
• Customized task modules
• Careful transcription
• Specialized QC functions
• Quick transcription
• Timed mode
• Metadata annotation
• Structural features, speaker diarization
• Correction mode
• e.g., correct automatic transcript or QTR ? CTR
• Comparison and adjudication of multiple
  transcripts
• Allows video input

12
One Channel View
13
One Speaker, MultiChannel
14
MultiSpeaker, MultiChannel
15
PanelView
16
Adjudication Mode