Robust speaker recognition system employing covariance matrix and Eigenvoice

This paper presents an original speaker recognition system that utilizes a quantized spectral covariance matrix on the input to a two-dimensional Principal Component Analysis (2DPCA) function. Eigenvoice algorithm is used as a classifying tool and is generated by the features of a group of speakers. The proposed system is selective in acquiring acoustic parameters and leads to a significant decrease in storage requirements. The system is robust in a noisy environment with recognition rates as high as 92% at 0dB SNR. Concatenated vowels that make up the speech signal are extracted from the TIMIT database and the noise environment is acquired from the NOIZEOUS database.