Go To Program 20 A Distortionless Noise-Reduction Filter Based on the Widely Linear Estimation Theory Jingdong Chen Jacob Benesty Yiteng (Arden) Huang 1 WeVoice, Inc., Bridgewater, New Jersey, United States 2 INRS-EMT, University of Quebec, Montreal, Quebec, Canada Recently, we have developed a single-channel distortionless filter for noise reduction based on the widely linear (WL) estimation theory. In this paper, we continue to investigate the WL distortionless noise-reduction filter. We demonstrate through both theoretical analysis and experiments that this filter can increase the signal-to-noise ratio (SNR) without adding any distortion to the desired speech signal, which is different from the traditional single-channel approaches that always introduce speech distortion. We also address the issue of how to estimate the variance parameters and circularity quotients of the clean, noisy, and noise signals in practice. View pdf