On the use of higher frame rate in the training phase of ASR

Abstract

The number of observations which are the basis for parameter estimation plays an important role in the quality of acoustic models. HMM based automatic speech recognition (ASR) systems generally have to cope with an insufficient number of observations for a good estimate. One way of tackling this problem is a well known procedure of state-tying, which is performed in order to gather sufficient information for a reasonable estimate for a large number of models. This procedure introduces an additional bias into the estimates, often leading to poor recognition results. In this paper a simple alternative to that solution is offered. It should be noted that most existing ASR systems use the same frame step size of 10ms in the training of the acoustical models, justifying it with the fact that speech signals exhibit quasi-stationary behavior at shorter durations. We claim that it is fully acceptable to adopt a much smaller frame step size in the acoustical training, thus providing estimators with a significantly higher number of observations compared to the standard 10ms case. This results in better parameter estimates and consequently better recognition results. Beside being justifiable from a phonetical point of view, it is also supported by results of an experimental on a real ASR system.