Compensation with Prior Knowledge

All methods analyzed and contrasted in this chapter have the unique attribute of exploiting prior knowledge about distortion in the training stage, in addition to training an HMM. They then use such prior knowledge as a guide to either remove noise or adapt models in the testing or deployment stage. Most methods which use prior knowledge about acoustic distortions as discussed in this chapter learn the nonlinear mapping functions between the clean and noisy speech features when they are available in the training phase as a pair of stereo data. By modeling the differences between the features or models of the stereo data, a distortion model can be learned accurately in training and subsequently used in testing to perform feature enhancement or model compensation. Another set of methods that also exploit prior knowledge operate by collecting and learning a set of simple models first, each corresponding to one specific acoustic environment in the training. These environment-specific models are then combined in the online fashion to form a new acoustic model that is aimed to fit the test environment in an optimal matter.