Speaker Recognition Based On Adaptive Gaussian Mixture Model

With the growing prevalence of mobile devices, users are starting to expect a fullrange of computational as well as communication services from these devices. Furthermore,given that these devices are used in a wide variety of environments, by users with differingaccess and operational requirements, the need fornon-keyboard based (hands-free)interfaces is becoming increasingly apparent. Recent advances have significantly improvedthe robustness and accuracy of speech recognition technology, making speech-basedinterfaces for human-computer (mobile device) interaction viable. Speech may also beused in order to enable secure access to the mobile device, through the use of voiceauthentication and speaker verification. Distributed processing approach can also allow foradded levels of security, by limiting amount of valuable information (such as speakermodels) stored on the actual mobile device. State-of-the-art speaker verification systemsare built around the likelihood ratio test, using Orthogonal transform with VectorQuantization Gaussian Mixture Models (GMM) for likelihood functions, a UniversalBackground Model (UBM) for alternative speaker representation, and a form of Bayesianadaptation to derive speaker models from the UBM. This work tackles optimal quantizerdesign of the speech cepstral features (MFCCs) for such systems. The problem is posed asthe minimization of loss of log-likelihood ratio between the quantized and unquantizedspeech features.