Audio-visual biometric recognition by vector quantization

We present a Vector Quantization based bimodal (speech and face) biometric recognition method which delivers high performance amidst noise, illumination variations and occlusions (disguised mode) while requiring very little training data, memory storage and complexity of operation. A transform VQ method delivers good face-recognition performance and a Text Dependent VQ method provides good recognition performance using speech. Simple fusion of two leads to a wider separation between the user-clusters in the combined feature space, leading to high performance.