Eigenspace estimation with missing values and its application to eigenvoice adaptation for speech recognition

Eigenspace estimation via principal component analysis (PCA) has been used in many applications, e.g., in eigenvoice modeling for speaker adaptation. Here the data of interest are the speaker supervectors, where each supervector is a concatenation of all the mean vectors in the speakerpsilas speaker-dependent (SD) model. One problem is that we often do not have enough speaker-specific data to establish the individual SD models (having unseen phones). To address this issue, an approach to eigenspace estimation by expectation-maximization (EM) algorithm in situations where the training samples contain missing values is proposed, applied to eigenvoice adaptation, and experimentally evaluated in this paper.