Optimization Of Decision Threshold For Speaker Recognition

Norizan, Mohamad (2003) Optimization Of Decision Threshold For Speaker Recognition. Masters thesis, Universiti Teknologi Malaysia.

[img] PDF (24 Pages)
Optimization_Of_Decision_Threshold_For_Speaker_Recognition.pdf - Submitted Version
Restricted to Registered users only

Download (1MB)

Abstract

It is well known that human can differentiate and recognize a person by his or her voices. This can happen without even looking at his or her face such as through telephone conversation. This phenomena shows that human speech consists of complex information that can be used to identify and verify the speaker. Speaker recognition has an important usage in many biometric applications and security. There are several well established techniques to extract speech features, keep them in database form, and make comparison with the input speech. However, there is no well established techniques when the recognition system has to decide whether to accept or reject the input speech with the most minimum error. This project examines the use of cost function to set the threshold point such that an optimization of false acceptance and false rejection rate can be achieved. Experimental results show that minimum cost threshold outperforms equal error rate (EER) threshold by 3.6% reduction in total cost for speaker verification task and 2.44% reduction in speaker identification task. Other minimum cost thresholds with different setting of FA and FR prior probabilities and costs are also shown to be better than EER in terms of total cost. The experimental results on digits combination show a significant reduction in total cost for all threshold settings. However, the minimum cost threshold still outperforms EER threshold by having a slightly higher total cost reduction in both speaker verification and identification tasks.

Item Type: Thesis (Masters)
Uncontrolled Keywords: Signal processing -- Digital techniques, Automatic speech recognition
Subjects: T Technology > T Technology (General)
T Technology > TK Electrical engineering. Electronics Nuclear engineering
Divisions: Library > Tesis > FKEKK
Depositing User: Nor Aini Md. Jali
Date Deposited: 15 Sep 2014 14:46
Last Modified: 28 May 2015 04:30
URI: http://eprints.utem.edu.my/id/eprint/13190
Statistic Details: View Download Statistic

Actions (login required)

View Item View Item