Adaptive compensation for robust speech recognition

Adaptation and compensation are two commonly adopted strategies to improve the robustness of speech recognition systems, especially in those cases when the testing data do not resemble the training data. In many ways, adaptation and compensation share similar goals and should be considered as a unified strategy for robust speech recognition. In this paper, we discuss adaptive compensation in which the compensation is accomplished through adaptive learning from the given testing data. Two major classes of adaptive compensation techniques can be considered, namely: (1) adaptive feature and model compensation, in which recognition features and/or model parameters are modified as needed; and (2) adaptive classifier compensation, in which the classifier structure and the corresponding parameters are modified as needed. We address the capabilities and limitations of these approaches.