Multidimensional Model based Speech Signal Representations for Automatic Speaker Identification

P. Premakanthan and W.B. Mikhael (USA)


: Speaker Identification, Linear Prediction,Vector Quantization.


A novel Model based Automatic Speaker Identification (M-ASI) technique employing multidimensional representations of the Linear Prediction (LP) coefficients, and the LP residual is proposed. During the training mode, the LP coefficients, and the LP residuals extracted from the speech signal are projected into multiple domains, and vector quantized codebooks are obtained using energy based split vector quantization. During the running mode, a closest match is found by comparing the speech vectors of the unknown speaker, and the reconstructed speech employing each of the known codebooks stored in the database. Employing a normalized matching accuracy measure, the proposed technique is consistently found to obtain enhanced ASI accuracy in comparison with Vector Quantization (VQ) employing existing single dimensional LP based ASI approaches at the expense of a modest increase in computational complexity. 100% speaker identification accuracy is obtained with a low signal-coding rate of less than 2.91 bps.

Important Links:

Go Back