Correction of distortion mask speech based on parameter estimation of AR model

Abstract

The generation model of speech signal has been regarded as an all-pole AR model. Distortion will happen when normal speech is disturbed or interfered. In this paper, we proposed a new signal model excited by the non-white noise signal to represent transfer function of a closed oxygen mask. Using LPC method to find the parameters of the all-pole signal model from the practical distortion signal, the prediction model is in accordance with the theoretical estimated of AR model Consequently, we can design the transfer function of the inverse filter with respect to the transfer function of the estimated model. The inverse filter is in series connection with the distortion filter in order to correct the distortion speech recorded by wearing the mask. By comparing the waveforms, normalized spectrums and spectrograms among the normal speech, the distortion speech, and the corrected speech using the proposed method, the experiment results indicate the feasibility and availability of the proposed method.