Abstract:

The joint algorithm of fuzzy clustering and LBG, which was used to train the VQ codebook, was brought forward on the basis of Fuzzy Clustering theory because of the defect that traditional algorithm LBG will emerge empty cell cavity and fall into local minimum and have larger number of iterations; This algorithm is have choice of the initial codebook by fuzzy clustering, and the traditional LBG find better codebooks based on the initial codebook. This paper present principles and methods of joint algorithm .The codebook of Log Area Ratio (LAR) of Coefficient of Linear Prediction and subband voiced strength was trained by this algorithm. The training process indicates that this algorithm could convergence with a faster speed and could have a strong expansion if the training samples are large enough and the wildpoints were removed. In the simulation experiment, the synthesized speech has better quality because of the codebook.

Key words: vector quantization(VQ);fuzzy clustering;LBG；codebook;the training sample set