Comparison of some algorithms for endpoint detection for speech recognition device used in cars

Abstract

A comparison of four algorithms form endpoint detection will be presented according to two different comparison techniques tests in different conditions. For the experiments two corpora were collected: one consisting of records with additive noise and other containing files recorded on an ordinary embedded microphone. The tests were conducted with assumption that hand-made boundaries may be considered as “gold standard”. Experiments have showed that the best results are achieved when only features based on short-time energy are used. Also it has revealed that pre-emphasis of the source signal may have negative effect on the result of the analysis.