A computationally efficient algorithm is proposed to remove noise impulses from speech and audio signals while retaining its features and tonal quality. The proposed method is based on the SD-ROM (Signal Dependent Rank Order Mean) algorithm. This technique has successfully been used to remove impulse noise from images. It has the advantage of being relatively fast, simple and robust. The algorithm estimates the likelihood the sample under inspection is corrupt relative to the neighboring samples and replaces a sample detected as corrupted by a value based on the neighboring samples. This algorithm also has the advantage of being `configurable' to the type of noise impulses in the sample, as the thresholds used to detect noise impulses can be varied to suit the signal.