Participation in the 2005 Spring Rich Transcription Evaluation

For the RT05s evaluation ICSI presented several systems combining different alternative algorithms. All combinations have in common:

• Frontend composed of a single acoustic stream with $19^{th}$ order MFCC, no deltas, 30 msec analysis window, 10 msec step size.

• Each initial cluster is modeled with a GMM with five Gaussian mixtures.

• Iterative segmentation/training.

• Segment level cluster purification.

ICSI participated in the speaker diarization task on conference room and lecture room data. The speech activity detection (SAD) algorithm was also ran on the data but it did not compete in the official SAD evaluation. Next the systems presented for each domain are presented.