Telephone speech data corpus and performances of speaker independent recognition system using the corpus

Abstract

The authors first describe the speech data corpus they collected from 400 male and 400 female subjects over the phone. They then compare the performances of two types of triphone model based speaker independent recognition systems, in which they used the corpus for training models and testing. One system uses a normal continuous mixture density HMM, and the other uses a CDHMM with a tree structure of 2,064 Gaussian distributions, which needs only one thirtieth of the Gaussian computation of a normal one. As a result, the system with the tree-structure CDHMM performed as well as 3% less than the system using the normal CDHMM. This shows that tree-structure CDHMM are useful for telephone speech recognition.<>