A new approach to noise-robust LPC

In this paper a new spectral estimate of noisy speech derived from linear predictive Autoregressive(AR) analysis is compared with the spectra produced by single order Linear Predictive Coding(LPC). The new estimate consists of a bin-by-bin average cf a set of linear prediction AR spectra with orders between 10 and 30. According to simulation results the improvement in peak centre frequency and bandwidth estimation resulting from using this model for noisy speech is equivalent to increasing the SNR by about 5dB when the SNR of the input signal is 20dB or less.