Memory-Based Approximation of the Gaussian Mixture Model Framework for Bandwidth Extension of Narrowband Speech.

摘要: In this paper, we extend our previous work on exploiting speech temporal properties to improve Bandwidth Extension (BWE) of narrowband using Gaussian Mixture Models (GMMs). By quantifying through information theoretic measures and delta features, have shown that memory significantly increases certainty about highband parameters. However, as features are non-invertible, they can not be directly used reconstruct frequency content. the presented herein, embed indirectly into GMM structure a memorydependent tree-based approach representation narrow band. particular, sequences past frames progressively grow in tree-like fashion. This growth results reliable estimates for parameters such Maximum Likelihood estimation is no longer necessary, thus circumventing complexity accompanying high-dimensionality training.