Performance improvement of speaker identification system using speech soft-decision

Abstract

One of the ways to improve performance of speaker identification (SID) systems in noisy environments is only to use real speech segments of an utterance. Therefore, SID systems utilize voice activity detection (VAD) methods to detect speech presence regions, and it is possible to develop improved systems by utilizing more outstanding VAD. There are two widely used VAD methods for speech signal processing; energy based VAD (EVAD) and statistical model based VAD (SMVAD). However, those VAD methods have a weakness in noisy environments; it is that they use hard-decision logic as decision rules. Because the ordinary VAD methods-that use hard-decision-always clearly decide the given speech frames as speech presence regions or speech absence regions even though it is difficult to obviously split into the two regions in noisy environments, errors may occur near borders of the two regions in noisy environments. Then the errors may decrease performances of SID systems. Thus, we propose a technique utilizing speech soft-decision (SSD) method, which weighs to speech boundaries with a value between 0 and 1. We evaluated the performance of the proposed system with the accuracy of speaker identification and compared the results with the performances of SID systems using the conventional hard-decision based EVAD and SMVAD. We could see that the performance of the proposed system improved in noisy environments.