The measurement of acoustic similarity and its applications

Abstract

It' s always helpful if we can predict how well a recognition system can perform for different tasks and evaluate the quality of different acoustic models. This paper presents a way to compute the acoustic similarity between words of a recognition task from the statistic models directly based on 2 kinds of definitions of the acoustic distance between recogniton units. It' s applications for command and control task complexity prediction and acoustic model evaluation are discussed. The experimental results shows it's a useful measurement.