Word endpoint detection/extraction method for speaker identification purposes

In the paper, a method for the word endpoints detection in the continuous speech is presented. The key element of the method is an adaptive finding of two thresholds to discriminate the time instants/sections in which the useful signal occurs. An algorithm for the threshold waveform computing based on the averaging of the speech signal envelope is proposed, and the envelope is found as a correlation between the square of the speech signal instant value and the Gaussian function. The cooperation of a preliminary gate (acoustic background-useful speech signal discrimination) and final discriminative gate (useful signal endpoint depicting) is explained. Experiments for the word beginning detection in continuous read utterances are reported and the test results concluded regarding method flexibility and application to speaker ID.