Improvement of analysis accuracy of high pitch speech utilizing residual of linear prediction analysis
Takeshi Fukabayashi · Electronics and Communications in Japan (Part III Fundamental Electronic Science) · 1992
Abstract This paper proposes a method to improve the accuracy of analysis in the traditional linear prediction analysis by utilizing the information contained in the residual due to the analysis error. In the proposed method, the linear analysis is applied first to the speech data and the residual is determined. At every interval almost equal to one pitch period, “0” sequences composed of more than one “0” are inserted into appropriate positions of the residual and the sequence is prolonged to the backward direction. The linear prediction analysis is applied to the resulting sequence. When the residual of the sequence with more inserted “0′s” than the order of the residual analysis is processed, the autocorrelation coefficient up to the delay needed for the analysis is equal to the sum of the independently determined autocorrelation coefficients between “0” sequences. Consequently, the result is almost equal to the analysis of the residual for almost one pitch period using the average autocorrelation function. This reduces the effect of the pitch period of the residual. Next, the impulse response sequence of the all‐pole filter constructed from the linear prediction coefficients of the residual is determined. The spectrum of the forementioned impulse response sequence is an approximation of the spectral envelope of the residual. The foregoing impulse response sequence is passed through the all‐pole filter determined previously from the linear prediction coefficients of the speech. The output, i.e., the synthesized waveform, is analyzed. It is shown by the analysis experiment of the synthesized speech that the accuracy of the analysis can be improved by this method.