Speaker verification for home security system

In this paper, we attempt to develop a reliable speaker verification algorithm that is suitable for use in a home security system. A phoneme-based hidden Markov model (HMM) has been adopted for the task of speaker verification with linear predictive cepstral coefficients (LPCC) as feature vectors for our model. Individual codebooks, designed to enhance performance, are also generated for all speakers in the test database. A simple way of combining the individual phoneme scores for text independent verification is also proposed. An equal error rate (ERR) of 10.5% has been achieved using the best phoneme model and 4.5% when using the combined scores of a 4-phoneme set