Optimization of the CELP model in the LSP domain

This paper presents a new Analysis-by-Synthesis (AbS) technique for joint optimization of the excitation and model parameters based on minimizing the closed loop synthesis error instead of the linear prediction error. By minimizing the synthesis error, the analysis and synthesis stages become more compatible. Using a gradient descent algorithm, LSPs for a given excitation are optimized to minimize the error between the original and the synthesized speech. Since the optimization starts from the LPC solution, the synthesis error is guaranteed to be lower than that obtained using the LPC coefficients. For the ITU G.729 codec, there is about 1dB of improvement in the segmental SNR for male and female speakers over 4 to 6 second long sentences. By adding an extra optimization step, the technique can be incorporated into the LPC, multi-pulse LPC and CELP-type speech coders.