home printable version contact information
description
Introduction
Browser
 
recogntion
Emotion
Speech
Speaker ID
 
tracking
Attention
Face
Body
Discourse
 
technologie
Distant microphone
Corpus
Publications
 
 
Interactive Systems Lab

At the Interactive System Labs of CMU, Pittsburgh, we have collected meetings since 1999 . Our database currently consists of more than 100 diverse meetings.
In the following, we will give an overview of how meetings are recorded (scenario, participants, environment, equipment) and prepared (transcription) at our lab. We will briefly describe the data we have collected to date. Finally, we will use statistical results to demonstrate how speaking styles change by choosing a variety of scenarios and situations.

We define a meeting as at least three individuals speaking to one another. A meeting recorded at ISL usually results in a maximum of eight mono audio files in WAV format, so-called speaker and recording protocol files containing information about the participants, equipment, environment and scenario, three video tapes, one transcription file of the entire meeting, a so-called marker file containing begin and end time stamps for conversation contributions, and a list of the meeting's vocabulary.

A meeting type can be controlled by a given scenario or topic, or meeting parties are invited to discuss a prearranged topic.

Creating the scenario provides the opportunity to satisfy specific research needs. For instance, to satisfy the needs of emotion recognition, we can create a scenario that may invoke anger or excitement, e.g. a controversial discussion about a political issue. Furnishing the scenario limits the spoken vocabulary to a specific domain. Giving a military strategy situation and military personal holding the meeting provides specific military vocabulary, acronyms and jargon.
We experimented with the following situations:
Project/Work Planning, Military Block Parties, Games, Chatting, and Topic Discussion.

104 meetings have been collected so far, generating a combined total of 103 hours (4.3 days). Each meeting lasted an average of 60 minutes. The recorded audio (since every speaker had his/her own channel) is 588.5 hours in 552 wav files, 77430.5 MB of data. The meetings have an average of 6.4 participants per meeting.

Related paper:
The ISL Meeting Corpus: The Impact of Meeting Type on Speech Style
Susanne Burger, Victoria MacLaren, Hua Yu
ICSLP-2002

 
top