A new algorithm for parameter re-optimization in multi-pulse excitation LP synthesizer

A new algorithm is proposed to re-optimize the synthesis filter parameters and pulse amplitudes in the Multi-Pulse Excitation Linear Prediction (LP) synthesizer. It is based on minimizing the perceptional weighted mean square errors between the original and the reconstructed speech. Steepest gradient descent and conjugate gradient descent are employed. The LP filter parameters and all the pulse amplitudes are iteratively re-optimized. Compared with the conventional multi-pulse excitation LPC algorithm, the new algorithm improves the segmental signal-to-noise ratio (SEGSNR) from about 20 dB to 23 dB. The informal listening also shows that the quality of synthesized speech is improved greatly.