抄録
This paper proposes a strategy to improve the performance of the LP method used for speech analysis. System identification methods are utilized to estimate the coefficients of the all-pole filter. For the input estimation, novel methods are proposed based on the derivation of an improved prediction error sequence. It is shown that by the use of the least squares type method, dependency on the pitch period of voiced speech is removed if the input sequence is accurately estimated as an impulse train from the output sequence being the observed speech data. It is also shown that the instrumental variable method aided by an instrumental model guarantees the stability of the estimated all-pole filter and provides better performance than the LP method m a noisy environment.
