MULTIPLE SPEAKER RECOGNITION

Abstract

Today, the speakers can be identified from their voices by using different techniques of speaker recognition, but when multiple speakers are speaking simultaneously it becomes difficult to identify a particular speaker because, of the mixing of the.speech signals of different speakers. In this thesis work the well-known techniques used for speaker recognition are reviewed. The main approach here is to separate the individual speech signals from the mixed speech signal. The technique of separating the signals from an observed mixture of a group of signals is called blind source separation. In this work for doing blind source separation for speech signals Independent Component Analysis (ICA) has been used which is a powerful higher order statistical technique. The process of speaker recognition follows the separation of speech signals. Speaker recognition consists of two parts: feature extraction and modeling of extracted features. Feature extraction is an important step in recognition process to extract sufficient infor-mation as it reduces the data while retaining speaker discriminative information. Two types of features, namely MFCC and LPCC are used for such purpose, but here only MFCC features have been used. For speaker modeling mainly there are two techniques: Vector Quantization and Gaussian Mixture Modeling(GMM), GMM has been used here for modeling of the extracted features. After the modeling a decision based algorithm is implemented to recognize the speaker. A mixture of speeches of five different speakers have been used here for test and validation. All the five speakers have been successfully identified. 3