Voxsecure : voice-based biometric speaker verification

Abstract

The acoustic features of speech have been found to differ between individuals. These acoustic patterns reflect both anatomy and learned behavioral patterns. This incorporation of learned patterns into the voice models has earned speaker verification its classification as a "behavioral biometric." The aim of this project is to design and implement a code toolkit that enables voiceprint scanning and analysis of different speakers. The toolkit will be tested on different speaker models from different recording sources, and will enable voice learning, voice identification and voice verification. Once the toolkit is in place the second objective is to build a reference implementation demo software, that utilizes the functionality of the VoxSecure toolkit for experimentation and evaluation. The task can be described as a classification problem, with two subtasks: (i) feature extraction and (ii) feature modeling and pattern matching. Feature extraction is done using one of the most successful methods in literature, Mel-Frequency Ceptrum Coefficients, which is obtained by a pipeline of digital signal processing techniques. The feature modeling is done by building a Gaussian Mixture Model (GMM), which is a probability density function to represent a speaker's voice model. The GMM parameters are first estimated using clustering algorithms such as K-Means, and then optimized using the Expectation-Maximization (EM) algorithm. Identification/Verification can then be performed using statistical inference on the GMM. With rigorous evaluation, we managed to configure VoxSecure for a 1003 identification rate on our corpus, and as an obvious parallel, an excellent verification system. These results augur well for the future application of similar systems in security scenarios outside of laboratory conditions. We are not yet capable, though, of utilizing such a system for real security scenarios, due to lack of training data, and due to the errors on short testing samples.