Effect of decision rule on speaker recognition performance
Loh Mun Yee, Abdul Manan Ahmad, Fatimah Mohamad · 2009
Speaker recognition from speech signal is still an ongoing research in forensics and biometrics area. Speaker recognition is the process to enable machine to recognize speaker's identity from their speech. The applications of speaker recognition technologies include access control system, security control for confidential information, and telephone banking. As a subset of speaker recognition, speaker identification task needed in forensics is to decide how likely it is that two recordings are from the same person. In these circumstances, pattern classification engine for speaker identification should capable to manage and process huge speaker data sets in a short time limit. Meanwhile, current works for the production of speaker recognition are almost directed towards accuracy problems, not time processing problems. How to make a decision is a critical question to reduce the time consuming problem. In this paper, we propose a novel technique that applying decision tree method towards better decision making for speaker identification system. In our proposed modeling, we take the superiority of decision tree theory, which is simplicity computation to distinguish a group of speaker into smaller subgroup. We believe that smaller group of training data will decrease time processing in run times. To evaluate the performance of our method, we have done simulations based on well known pattern classification method, Vector Quantization. A comparison about time processing has made between proposed method and baseline Vector Quantization model. Comparative results show that our decision tree method always yielded better improvements in time processing. Thus, our method provides an alternative way for real time identification system which time is the important issue. In this paper, a new, robust and simplicity computation method of pattern classification technique for speaker identification system is proposed. Consequently, this research is intended to develop a fully optimize ways speaker identification approach from decision tree theory.