Optimization of temporal decomposition model with level promotion

The level promotion method is presented to optimize event functions of temporal decomposition of speech. This method takes into account not only the minimum accumulative reconstruction error between two adjacent event vectors, but also the well-shapedness property of event functions. In our work, the level promotion method is flexibly combined with the modified restricted temporal decomposition and the limited error based event localizing temporal decomposition, respectively. Compared with the conventional construction methods of event functions, our method reduces the spectral distortion of the line spectral frequency parameters in the temporal decomposition model.