Improved Memoryless GMM VQ for Speech Line Spectral Frequencies

An improved memoryless vector quantization scheme is proposed for efficient quantization of speech line spectral frequencies (LSF) parameter. The algorithm is based on Gaussian mixture model with diagonal covariance matrices. Assume that the LSF parameter belongs to some of the Gaussian distribution of GMM, and then the LSF vector could be quantized using random Gaussian distribution quantization. Scheme of bit allocation for quantization is proposed and each dimension of the LSF parameter is quantized using non-uniform scalar quantization. Finally, performance of split-VQ, PDF VQ, and proposed memoryless GMM VQ are compared. Testing results show that the proposed memoryless GMM VQ outperforms the conventional split VQ and PDF VQ, i.e. memoryless GMM VQ could achieve comparable or even better quantization performance using less bit resources