An environment structuring framework to facilitating suitable prior density estimation for MAPLR on robust speech recognition

In this paper, we propose using an environment structuring framework to facilitate suitable prior density estimation for maximum a posteriori linear regression (MAPLR) under adverse testing conditions. The framework is constructed in a two-stage hierarchical tree structure by performing two algorithms, environment clustering and environment partitioning. The constructed framework has good capability to characterize detailed regional information of various speaker and speaking environments. We intend to incorporate such information into prior density calculation for MAPLR and have designed three types of prior density, namely clustered prior, hierarchical prior, and integrated prior densities. We conduct experiments with the Aurora-2 task. From the testing results, we first observe that MAPLR provides improvements over baseline and maximum likelihood linear regression (MLLR) using either one of the three prior densities. Moreover, we find that by using the integrated prior density that combines the advantages of the other two, MAPLR can give the best performance. When using the best integrated prior density, MAPLR achieves a clear improvement of 10.72% word error rate reduction over the baseline result.