Twenty-First Century Automatic Speech Recognition: Meeting Rooms and Beyond

1
Twenty-First CenturyAutomatic Speech
RecognitionMeeting Rooms and Beyond

• ASR 2000
• September 20, 2000
• John Garofolo
• John.Garofolo_at_NIST.gov

2
Challenges

• Target for the new millenium in ASR Technology
• Meeting Room Transcription and Annotation Task
• multiple sensors
• stationary, mobile, and arrays of mics in
  conjunction with video input devices
• noise and microphone robustness
• speaker-independent recognition
• speaker identification
• automatic production of usable transcriptions
  with speakers identified and with properly
  formatted, capitalized, and punctuated text.
• Perfect research task to move forward the
  state-of-the-art
• Development infrastructure will require
• new metrics, evaluation tools
• new I/O specifications
• research corpora, new methods of collecting,
  compiling, and annotating data

3
NIST Proposed Initiative

• Collaborate with ASR research community to create
  evaluation infrastructure
• Develop corpus design and transcription and ASR
  system output specifications
• Revise and update NIST SCLITE ASR scoring
  software to extend beyond classical word error
  rate measurements
• Collaborate with NIST Smart Space Lab to collect,
  transcribe, and annotate a pilot meeting room
  transcription corpus
• Sponsor Evaluations and Workshops

4
Meeting Room Pilot Corpus

• Meeting type
• Possible focus group discussions requiring
  information lookup and real consensus building
• Participants
• At least 4 per meeting plus moderator
• Native speakers?
• Multi-microphones
• Head-mounted control
• Microphone array
• Lapel mikes worn by, or desk-top mikes for each
  participant
• Table/wall-mounted stationary mikes

• Video
• Wide-angle view positioned so that it can be
  correlated with mike array for source location.
  Possibly other views to capture faces head-on.
• Annotation
• Transcription (words with capitalization/punctuati
  on)
• Speaker ID
• Background noise conditions
• Some initial exploration of annotating dialogue,
  people movement, gestures, lip movement,
  interaction with devices

5
NIST Smart SpaceTest Bed Laboratory

• 59-mic array, assorted conventional mics
• Cameras/video capture
• Large screen display
• Pervasive devices
• Palm tops
• Tablets
• Wireless LAN
• Data collection servers
• Gigabit Ethernet
• High-bandwidth data flow system
• Well-suited for creating pilot meeting corpus

6
Approach for 2000 - 2001

• NIST will collaborate closely with a few research
  sites who will be the early users of the data to
  create the project specifications.
• Via E-mail list and Web site
• NIST will create a pilot meeting room data
  collection
• Data storage will be a significant issue
• NIST will create evaluation software for the new
  domain
• Update SCLITE detection-based scoring software
• If feasible, NIST will coordinate an experimental
  evaluation
• Late summer/early Fall 2001
• NIST will host a workshop (October 2001)
• to discuss research issues
• to introduce the pilot corpus to the wider
  research community
• to discuss evaluation metrics and the dry-run
  evaluation
• to plan for future efforts (kickoff for larger
  DARPA program?)

7
21st Century Automatic Speech Recognition
Meeting Rooms and Beyond

• John Garofolo
• John.Garofolo_at_nist.gov
• NIST Speech Group
• http//www.nist.gov/speech
• NIST Smart Space Lab
• http//www.nist.gov/smartspace/

ASR 2000 September 20, 2000