Participation in the 2006 Spring Rich Transcription Evaluation

For the RT06s evaluation a total of 23 systems were presented in the multiple tasks and subtasks proposed. Each system uses one or more of the improvements presented in this thesis. The common characteristics of all systems are:

In the following list the main characteristics of the systems presented are explained. Across tasks, systems with the same ID are equal or very similar, just differing on a few parameters. Their characteristics are:

This is the primary system presented this year for all multi-microphone conditions. It uses most of the proposed improvements of this thesis, and all changes in the diarization code from last year's evaluation.

This system is presented for the multi-microphone cases and is composed of RT05s evaluation code using the new filter&sum algorithm, this year's hybrid speech/non-speech detector and taking advantage of the delays for clustering. It uses a minimum duration of 3 seconds, 1/5 initial Gaussian mixtures for delays/acoustics and a split weight of 0.1/0.9 between the streams fixed for all meetings. It is intended to compare the improvements of using delays in the system compared to last year's performance.

This system is identical to p-wdels in all parts except the decision of the initial number of clusters, which is fixed to 16 and 10 clusters for conference and lecture rooms, respectively. It intends to compare the robustness of the initial number of clusters selection.

This system contains all of RT06s improvements with respect to filter&sum (when available, in MDM), speech/non-speech detection and other diarization algorithms except the inclusion of the delays as an extra feature stream.

This system uses all improvements in filter&sum (when available, in MDM) and speech/non-speech detection while using the RT05s core speaker diarization system. It is meant to serve as a baseline result for RT06s systems.

This system guesses one speaker for all of the show. In RT05s this was presented as the primary system for lecture room data, showing the need to beat this system in order to think of speaker diarization in the lecture data as a reasonable task. In RT06s it is also presented as a baseline lecture-room system to be compared with the other lecture-room systems.

user 2008-12-08