Computer Science & Engineering: An International Journal (CSEIJ), Vol.2, No.4, August 2012 DOI : 10.5121/cseij.2012.2401 1 ANALYSIS AND PROPOSING A ROBUST SOURCE FEATURES FOR AUTOMATIC TEXT-INDEPENDENT SPEAKER INDEXING SYSTEM V.Subba Ramaiah 1 and R.Rajeswara Rao 2 1,2 Department of Computer Science & Engineering, MGIT, Hyderabad, AP, India 1 subbubdl@gmail.com, 2 raob4u@yahoo.com ABSTRACT Speaker indexing (tracking) is the task of recognizing the multiple speakers from the given speech signal. Speaker indexing is a pattern recognition task. Every pattern recognition task is classified into three phases namely, feature extraction, training and testing phases. In this paper, we analyse and study about the major feature extraction techniques that exist and propose robust source features for text-independent speaker indexing system. A comparative study is carried out with each existing feature extraction techniques available with source features for speaker indexing task. Finally, we propose, source features are better than the other feature extraction techniques. KEYWORDS Speaker Indexing, Speaker Recognition, Source Feature, and Feature Extraction 1. INTRODUCTION Speaker recognition is the process of automatically recognizing who is speaking on the basis of individual information included in speech waves. Speaker recognition can be classified into two tasks, namely speaker identification and speaker verification [1]. Speaker identification is the process of recognizing the speaker from the given spoken utterance from the registered set of ‘N’ Speakers. Speaker verification is the process of accepting or rejecting the identity claim of the speaker. Speaker recognizing technology is used in biometric approaches. This technology is used in many commercial applications such as banking by telephone, voice mail, and other security related applications. Speaker recognition is one of the centre field of research. In existing systems, unknown speaker is recognized from the given speech signal. In many real- time conversations and news broad casting, the speech is continuous, beginning and end of speech segment of a speaker is unknown. Therefore, we need to index speech signals based on the speaker utterance. This process is called speaker segmentation or speaker indexing system. Speaker tracking is also essential in many applications, such as conference and meeting indexing [2], audio/video retrieval or browsing [3, 4], speaker adaptation for speech recognition [5,23], and video content analysis. Traditionally, the speaker recognition task supposes that training and testing are composed of mono-speaker records. Then, to handle this kind of multi-speaker recordings, some extensions of the speaker recognition task are needed, such as: