Speech enhancement by ARUI model and a new type of adaptive filter with sequential processing

Abstract

In the present study, we attempt to sequentially estimate clean speech from noisy speech in various environments with neither a priori training nor estimation of noise power, and to improve quality of speech in both subjective and objective evaluations, respectively. The time varying auto-regressive with unknown input (ARUI) model and a new type of adaptive filter (called the “adaptive filter with bias”), proposed in the present paper, are applied for enhancement of speech that has been degraded by white/colored additive interference. The ARUI model is used to reduce high-frequency noise components. The adaptive filter is designed in order to reduce low-frequency additive noise components. The Kalman filter is used to estimate the parameters of the ARUI model and the adaptive filter. We confirmed that the quality of the enhanced speech is improved by comparison via original noisy speech or spectral subtraction in both objective and subjective evaluations, respectively.