Automatic speaker recognition is one of the major topics in the area of speech recognition. Speaker recognition is the general name for those technologies used to identify or verify who is speaking based on the voice she or he makes. Pratical applications include voice dialing, banking transaction via a telephone network, database access services, security control for confidential information areas etc....[ Read more ]

Automatic speaker recognition is one of the major topics in the area of speech recognition. Speaker recognition is the general name for those technologies used to identify or verify who is speaking based on the voice she or he makes. Pratical applications include voice dialing, banking transaction via a telephone network, database access services, security control for confidential information areas etc.

Unlike the text-dependent case, in text-independent speaker recognition, the contents of utterances spoken by a speaker to access the system are not restricted. This, although may lead to a degradation in performance, provids more flexibilities and conveniences to both the speaker and the application. Text-independent speaker recognition can be divided into text-independent speaker identification and speaker verification according to the nature of application.

In order to build systems which can be optimal in performance, we study different factors affecting the performance of text-independent speaker recognition systems. A set of text-independent baseline speaker identification and verification systems are built. Specifically, we propose a novel frame-selection algorithm based on Log Likelihood Ratio (LLR) aiming at selecting the frames which are useful for representing speaker characteristics from the speaker utterances, and apply it to our speaker recognition systems.