Effect of different parameter sets on reconstructed speech quality using temporal decomposition

Temporal Decomposition (TD) is a speech analysis procedure based on linear modeling of coarticulation which relates speech spectral parameters to phonetic events. The performance of the method is highly dependent on the dynamic behavior of the spectral parameters used. In [3], a comparative study of the effect of 9 different spectral parameter sets has been carried from the `phonetic relevance' and `parametric distance analysis' viewpoints. However, in speech coding applications, the phonetic relevance and parametric similarity are not the main goals. Instead, the quality of reconstructed speech using the approximated parameters is more crucial. In this paper, we compare four best candidates as spectral parameters in TD, through objective assessment of the reconstructed speech using a perceptually based spectral distance measure.