Text-Independent Speaker Identification System Using Different Pattern Matching Algorithms
Because words are not the only understandable information gathered from speech, voice biometry is a powerful measurement technology. Listeners can know gender, age, health situation, emotion state, and speaker identity. Automatic Speaker Recognition (ASR) represents the ability of a program or device to identify the speaker from its utterance. ASR is usually divided into Speaker Verification (SV) and Speaker Identification (SI). In this paper we propose a Speaker Identification System (SIS) that identifies effectively all registered speakers based on their speech. SIS is composed of two main modules: feature extraction and feature matching. Mel Frequency Cepstrum Coefficients (MFCCs) is used for extracting features from speech signal. For feature matching, we have applied four common ASR algorithms: Vector Quantization (VQ), Gaussian Mixture Model (GMM), Artificial Neural Networks (ANNs), and Decision Trees (DT). When evaluating performance, fusion method was the best accurate procedure for identifying true speakers and rejecting imposters. Fusion method has an accuracy rate of 99% for 100 speakers and 96% for 25 and 50 speakers. The proposed SIS in this study proved its ability to identify a text-independent closed set of speaker groups effectively.
Index Terms- Speaker Recognition, Speaker Identification, Vector Quantization, Gaussian Mixture Models, Artificial Neural Networks, Decision Trees, Fusion.