The Purdue SAD system was originally designed to identify speech regions in multichannel meeting recordings with the goal of focusing transcription effort on regions containing speech. In the NIST RT-05S evaluation, this system was evaluated in the ihm condition of the speech activity detection task. The goal for this task condition is to separate the voice of the speaker on each channel from silence and crosstalk. Our system consists of several steps and does not require a training set. It starts with a simple silence detection algorithm that utilizes pitch and energy to roughly separate silence from speech and crosstalk. A global Bayesian Information Criterion (BIC) is integrated with a Viterbi segmentation algorithm that divides the concatenated stream of local speech and crosstalk into homogeneous portions, which allows an energy based clustering process to then separate local speech and crosstalk. The second step makes use of the obtained segment information to iteratively train a Gaussian mixture model for each speech activity category and decode the whole sequence over an ergodic network to refine the segmentation. The final step first uses a cross-correlation analysis to eliminate crosstalk, and then applies a batch of post-processing operations to adjust the segments to the evaluation scenario. In this paper, we describe our system and discuss various issues related to its evaluation.
展开▼
