Noncausal ARMA modeling of voiced speech

A noncausal autoregressive moving average (ARMA) source-filter model of voiced speech is proposed. Although the human speech-production mechanism is obviously causal, a noncausal model allows the simple source-filter approach to incorporate different parameters for the open-glottis and closed glottis portions of the pitch period (without explicit determination of the open- and closed-glottis regions). The noncausal impulse response is obtained by standard cepstral deconvolution. Separate ARMA models are then found for the causal and anticausal portions of the impulse response. Initial experiments show that very close approximations to the speech waveform and spectrum are produced by two twelfth-order ARMA models