Meeting Recorder

1
Meeting Recorder

• Adam Janin
• janin_at_icsi.berkeley.edu

2
A Portable Meeting Recorder

• Record meetings in natural settings
• Support multiple speakers
• Allowing correction and annotation
• Support indexing and searching
• Self-contained (using IRAM)

3
Vector IRAM

• High absolute performance
• Floating point and integer vector units
• High bandwidth to memory
• High performance per watt
• Low instruction dispatch rate for vectors
• On-chip memory
• Well-suited for Meeting Recorder
• ASR is (mostly) easy to vectorize
• Supports floating point
• Compiler

4
(No Transcript)
5
Why Meeting Recorder?

• Why not written notes?
• Unreadable handwriting
• Detracts from listening
• Incomplete record
• Hard to search
• Why not a tape recorder?
• Hard to annotate
• Playback can be disruptive
• Too much data
• Hard to search

6
Natural Settings

• Record impromptu meetings
• Robust to noise, reverberation, etc.
• Large vocabulary
• Spontaneous speech

• Portable
• Self-contained
• Hands-free

7
Multiple Speakers

• Robust to accents, speaking styles
• Interrupts
• Mark speaker change
• Adapt to common participants

8
Correction and Annotation

• Correcting informing the system that it has
  made an error
• If the system has a good idea of alternatives, it
  may be faster to correct than to edit
• Recognizer can adapt to user and vocabulary
• Annotation Supplementing and modifying the
  output
• Thats not what I meant to say
• Pen vs. speech input
• General ink and text annotations

9
Indexing and searching

• Index by speaker and text content
• Store and replay audio
• Allow text and spoken queries

• Perfect speech recognition is not required!
• N-best lists
• Function vs. content words
• Audio record

10
Collaboration

• Multiple Meeting Recorders in a meeting connected
  via a wireless network
• Offline
• Share notes (à la NotePals)
• Improve transcript quality
• Online
• Chat
• Microphone array

11
Data Collection Phase

• Meeting room at ICSI
• 5 wireless head-mounted microphones
• 3 PZM omni-directional microphones
• HandSpring PDA (probably wired network)
• Primarily speech and NLP meetings
• Collect and hand-transcribe about 40 hours of
  meetings

12
Issues

• Will people object to being recorded?
• Will their speaking style change?
• Are the limitations of the data collection phase
  important?
• Is cross-domain training feasible?
• Lots of UI issues
• Lots of IR issues...
• Lots of ASR issues

13
Prototype of a Personal Dictation System

• Records a single user dictating text
• Allows correction and limited annotation

14
Prototype Screen Shots
Richard Posner, who heads the United States
Court of Appeals ...