On prefiltering and endpoint detection of speech signal

Abstract

In speech recognition systems, it is a fundamental task to detect endpoints of speech signals. The efficiency of endpoint detection affects the performance of the entire recognition system. This is especially important to an isolated word recognition system. Ordinarily, the rate of zero crossing and short time energy are used for endpoint detection. However, this is not a trivial process. In fact, there are many details we need to pay attention to. Through experiments we found that by selecting a moderate algorithm for ZCR and energy, and cooperating with the preemphasis filter (prefiltering), better accuracy of endpoint detection can be obtained.