Abstract
A technique that uses a prediction error filter for reducing sinusoidal noises from a noisy speech has been proposed previously. Since the prediction error filter can estimate the sinu-soidal noise completely, the output becomes zero in a non-speech segment. After the prediction error filter converges, the update of the filter coefficients is stopped. Then the fixed prediction error filter can cancel the sinusoidal noises except for a speech signal in a speech segment. However, frequency characteristics of the filter depend on its prediction algorithm, and the coefficients may converge the values which gives degradation of the speech. In this paper, we propose a new noise reduction algorithm which is a kind of leaky LMS algorithm, so that the prediction error filter removes only the sinusoidal line spectrum without speech degradation.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have