Data-driven and physical model-based designs of probabilistic spatial dictionary for online meeting diarization and adaptive beamforming

In this paper, we comparatively study alternative dictionary designs for recently proposed meeting diarization and adaptive beamforming based on a probabilistic spatial dictionary. This dictionary models the feature distribution for each possible direction of arrival (DOA) of speech signals and the feature distribution for background noise. The dictionary enables online DOA detection, which in turn enables online diarization. Here we describe data-driven and physical model-based designs of the dictionary. Experiments on a meeting dataset showed that a physical model-based dictionary gave a word error rate (WER) of 24.9 %, which is close to that for the best-performing data-driven dictionary (24.1%). Therefore, the former has a significant advantage over the latter that it allows us to bypass the cumbersome measurement of training data without much degrading the performance of the automatic speech recognition (ASR).