This paper presents a new distance measure that is based on the spectral sensitivity of the line spectrum frequency parameters (LSFs) and its Karhunen-Loeve (KL) transformed coefficients. It is shown that the proposed distance measure achieves better performance of vector quantization (VQ) compared to other methods in the field of LSF coding. In most cases, the percentage of outliers is reduced when using the new one, compared to the best results of using other conventional weighting functions, The use of this distance as the weighting function of the LSF transformed parameters is also suggested.

We have developed a new general distance measure that not only can be used in a vector quantization (VQ) of the line spectrum frequency (LSF) parameters but performs well in the LSF transformed domain. The new distance is based on the spectral sensitivity of the LSF and their transformed coefficients. In addition, the fixed scaling factor is used to decrease the sensitivity of the spectral error at higher frequencies. Experimental results have shown that the proposed distance measure leads to as good as or better performance of VQ compared to other methods in the field of LSF coding. The use of this distance as the weighting function of the LSFs' transformed parameters is also suggested.

We propose a new transcoding method for MPEG video bit-streams. In our implementation, instead of using scalar quantization in the transcoder, a lattice vector quantizer (LVQ) is applied to exceed the MPEG compression capabilities while providing acceptable quality. In this way, the transcoded video is not MPEG-compatible anymore, hence a low cost user interface is needed prior to the MPEG decoding. However, the decoded pictures have higher subjective quality compared to those transcoded by an MPEG-compatible transcoder at the same bit-rate. Due to the LVQ the quantization noise is more uniform on the pictures, and less artifacts are visible around the edges. Beside this, a slight improvement in the PSNR was also experienced by using certain LVQ parameters.

The intraframe correlation properties of line spectrum pair (LSP) are used to develop an efficient encoding algorithm using the Karhunen-Loeve (KL) transformation. The important nonuniform statistical characteristics of LSP frequencies are investigated. Based upon this nonuniform property the neural network based techniques for generating the transform vectors via system training are studied. Using the principal component analysis (PCA) network to decorrelate LSP coefficients, we show that these new approaches lead to as good or better distortion as compared to other methods for speech analysis-synthesis.