Estimating clean speech thresholds for perceptual based speech enhancement

We propose a new method for the estimation of clean speech masking thresholds for speech enhancement. These thresholds are applied to a perceptually based spectral subtraction algorithm to enhance speech in a non-stationary noise environment. In contrast to other approaches we do not directly use an estimate of the clean speech to obtain the masking thresholds, but examine the relationship between those from the corrupted speech and corrupting noise. The thresholds from this algorithm are compared to those produced from a clean speech estimate from a variety of common spectral subtraction algorithms