抄録
In this paper we propose a method to improve the accuracy of linear prediction analysis of high pitch speech. Processing of the variable extension of pitch period is performed by inserting sequences of ‘0’s every pitch period of residual of linear prediction and synthesizing a waveform every pitch period independently. Processing of windowing is performed every pitch period of this synthesized waveform. These waveforms are separated by inserting ‘0’s of the number more than a linear prediction analysis order to the back of windowing, and then these waveforms is analyzed by linear prediction analysis collectively. We show that by this method, the influence of the short pitch period is reduced and the analysis accuracy is improved. Additionally, the problem of increase of the amplitude of the waveform synthesized by linear prediction coefficient and the pitch pulse obtained from residual of linear prediction is improved. And we investigate the tolerance to noise of the proposed method.
