Speaker Recognition Using Gaussian Mixtures Models

Speech signal contains several levels of information. At first it contains information about the spoken message. At second level speech signal also gives information about the speaker identity, his emotional state and so on. The task of speaker recognition can be divided into two parts: speaker identification and speaker verification. Speaker identification is answering the question which one of the group of known voices best matches the input voice. Speaker verification is answering the question is really this person who claims to be. Also speaker recognition can be text dependent or text independent. In text dependent speaker recognition, speech recognition is performed too and there are used the same methods as in speech recognition. In speech and speaker recognition systems various features are used [1], calculated from the short intervals (named as frames) of the speech signal: coefficients of Linear prediction coding (LPC), cepstral coefficients, calculated from LPC model (LPCC), mel-cepstrum coefficients (MFCC), bark cepstrum coefficients, delta cepstrum and so on. Duration of the frame is about 25ms. These frames overlap one another. The same features are often used in speech and speaker recognition systems, however there are two completely different tasks. There are proposed a lot of methods for speaker modelling and recognition. In text dependent speaker recognition the most popular methods are dynamic time warping (DTW), Hidden Markov Models (HMM) [2]. In text independent speaker recognition the most popular methods are: Vector Quantization (VQ) [3], fully connected (ergodic) HMM‘s, artificial neural networks (ANN) [4], support vector machines (SVM) [5], and Gaussian Mixture Models (GMM) [6]. In this paper we would like to propose text independent speaker recognition method with new feature vectors, that consist of fundamental frequency and four formant frequencies, try to build Gaussian Mixture speaker models. Vector Quantization method was employed for initial parameters estimation of speakers GMM. Experiments of speaker recognition were performed and compared with experiments using Gaussian Mixture Models with mel – frequency cepstral coefficients, that is baseline in speaker recognition.