NIST 2004 Rich Transcription Spring Meeting Evaluation

Within the NIST 2004 Spring Rich Transcription Evaluation (NIST Spring Rich Transcription Evaluation in Meetings website, http://www.nist.gov/speech/tests/rt/rt2005/spring, 2006) speaker diarization was evaluated in meeting recordings in two different conditions: Multiple Distant Microphones (MDM) and Single Distant Microphone (SDM). The MDM condition uses multiple microphones located in the center of a meetings table, and the SDM case uses only one of these microphones, normally the most centrally located. This is the first time that this task was performed for meetings environment on the MDM condition. A full description of the different tasks evaluated and the results of such evaluation can be found in Garofolo et al. (2004). Following are the approaches (in brief) that the participants proposed for the MDM and SDM conditions:

• Macquarie University in Cassidy (2004) proposes the same system for SDM than for MDM, using always the SDM channel. A BIC based speaker segmentation step is followed by an agglomerative clustering using Mahalanobis distance between clusters and BIC as stopping criterion.

• The ELISA consortium in Fredouille et al. (2004) proposes a two-axis merging strategy. An horizontal merging consists on the collapse and resegmentation of the clustering output of their two expert systems (based on BIC and EHMM) as proposed in the RT03 and SRE02 evaluations (Moraru, Meignier, Fredouille, Besacier and Bonastre, 2004). This is done for each individual MDM channel or for the SDM channel. The vertical merging is applied when processing multiple channels and unifies all the individual channels into one single resulting output by merging all channels at the output level. It uses an iterative process that searches for the longest speaker interventions that are common to all outputs and finally assigns to the closest speaker those segments of short duration where the different channels do not agree on.

• Carnegie Mellon University (CMU) in Jin et al. (2004) presents a clustering scheme based on GLR distance and BIC stopping criterion. In order to obtain the initial segmentation of the data it does a three steps process, first a Speech Activity Detection (SAD) is done over all the channels, then the resulting segments for all channels are collapsed into a single segmentation and the best channel (according to an energy/SNR metric) is chosen for each segment. Finally GLR change detection is applied on segments $>$5s to detect any missed change point. The speaker clustering is done using a global GMM trained on all the meeting excerpt data and adapted to each segment and uses GLR to compute the cluster pair distances to be used in an agglomerative clustering processing with BIC stopping criterion.