Spectral quantization using statistics of static and dynamic features

We propose a vector quantization (VQ) method using statistics of static and dynamic features. The distortion between the input vector and codeword is measured by the likelihood defined by statistics. The output vector is determined so that its likelihood is maximized. The proposed method enables us to obtain the quantized spectra which vary quite smoothly. To evaluate the proposed method, we applied it to encoding of the LSP parameters. From the subjective test, it was shown that the use of the proposed method results in a significant improvement of subjective quality.