Quantization of variable dimension spectral vectors

We present a comprehensive study of a quantization technique for variable dimension spectral vectors that we call weighted non-square transform VQ (WNSTVQ). This technique employs generalized perceptually weighted linear dimension conversion to a fixed dimension vector followed by vector quantization (VQ). We show that the total error can be separated into the weighted modeling error and the weighted quantization error. Perceptual weighting in the modeling and the quantization of the variable dimension spectral vectors is essential for harmonic coding systems where the spectral vectors are obtained by harmonic sampling of the spectrum of the LP residual signal. The paper presents codebook search procedures for the WNSTVQ and experimental results for different types of transforms.