Codebook Design using DCT Coder for Text-independent Speaker Recognition

S.-Y. Lung (Taiwan)


Speaker recognition, Gaussian mixture model, Discretecosine transform, Codebook design, Zero-treequantization


This paper presents text-independent speaker identification results for TALUNG databases for telephone speech. A speaker identification system based on Gaussian mixture speaker model (GMM) derived from codebook design using discrete cosine transform (DCT) coder is proposed. We have combined zero-tree quantization and the DCT coder to looking for the desired number of initial code vectors in a training space. Identification accuracies of 91% were achieved on the TALUNG databases for 100 Mandarin speakers.

Important Links:

Go Back