Speech Enhancement Using Gain Function of Noisy Power Estimates and Linear Regression

Speech enhancement algorithm is proposed based on calculating the ratio of the average noisy power spectrum in time-frequency bins to the normalized time-frequency average. If the ratio is greater than the linear regression estimator threshold, speech is considered to be present. The algorithm has the advantage of being computationally simple and it produces significantly less residual noise. We evaluated the algorithm under various noise conditions to demonstrate its general superiority to conventional methods.