A two-threshold strategy, refined by simple criteria to merge adjacent disturbances and to reduce threshold value along iterations in a given block is presented as a way to improve the efficiency of threshold-based algorithms indented to impulsive noise detection in audio signals. Restoration applied to naturally or artificially corrupted real audio signals using the proposed detection scheme demonstrates an evident improvement in perceptual quality of the restored signals over those obtained by employing the conventional method. Quantitative measurements indicate a significant reduction of missing detections, justifying better subjective results, in spite of a slight increase in false detections. The modified detection algorithm seems to be more robust to variations in both audio and noise characteristics and does not imply any remarkable increase in computational complexity compared to the standard one.