Noise Reduction Algorithm using the Linear Regression(LR) Estimator Thresholds for Speech Enhancement

다국어 초록 (Multilingual Abstract)

A new noise reduction algorithm is based on calculating the ratio of the average of noisy power spectrum in time-frequency bin and its a time-frequency averaged normalized. If the ratio is greater than the linear regression(LR) threshold, it is taken as speech present else it is taken as speech absent. The algorithm has the advantage of being computationally simple and clearly lower residual noise. It is evaluated using the segment SNR under various noise conditions. We show that this algorithm is generally superior to that of the conventional method.