Extending the Hierarchical Scheme: Inter and Intra Phonetic Information

Abstract

It could be seen in the previous chapter that the hierarchical scheme, consisting of two MLPs in tandem, significantly outperforms the single MLP approach. Additionally, proposed downsampling schemes implemented at the output of the first level improve system performance under memory and computational time considerations. In this chapter, an analysis of the hierarchical scheme is presented based on the downsampling schemes. The kind of information at the output of the first level is investigated. Additionally, it is shown how the second level uses this information for improving performance. Once the system is fully analyzed, an extension of this technique is more feasible in order to achieve higher recognition accuracies.