Robust statistical processing of TDOA estimates for distant speaker diarization

摘要

Speaker diarization systems aim to segment an audio signal into homogeneous sections with only one active speaker and answer the question "who spoke when?" We present a novel approach to speaker diarization exploiting spatial information through robust statistical modeling of Time Difference of Arrival (TDOA) estimates obtained using pairs of microphones. The TDOAs are modeled with Gaussian Mixture Models (GMM) trained in a robust manner with the expectation-conditional maximization algorithm and minorization-maximization approach. In situations of multiple microphone deployment, our method allows for the selection of the best microphone pair as part of the modeling and supports ad-hoc microphone placement. Such information can be useful for subsequent speech processing algorithms. We show that our method, which uses only spatial information, achieves up to 36.1% relative reduction in speaker error time compared to an open source toolkit using TDOA features and tested on the NIST RT05 multiparty meeting database.