This thesis compares several speaker recognition algorithms in reverberant environments. The Gaussian Mixture Model, Auto-regressive vector model, covariance based models and Multi layer perceptron are compared.
The methods are compared when there is a mismatch between the training and test speech due to the non-reverberant nature of the training speech and the reverberant nature of the test speech.
In order to counteract the effects of reverberation, training was performed using reverberant speech. Average recognition accuracy improved by 9.8% for the GMM, 53% for AR-Itakura, 18.8% for sphericity measure, 18.1% for the divergence shape measure and 15.9% for AR-AGS.
A method was proposed to create a set of reverberant models for each speaker using speech reverberated to different degrees. A novel technique was proposed to determine which reverberant model for each speaker best matches the reverberant test speech. 98.7% classification accuracy was obtained using an Autoregressive vector model adapted for this purpose.