To improve the robustness of watermarking, the spectral singularities resulting from the widely used noise-like author keystream are investigated with audio signals. Because such a keystream distributes its energy abruptly among all frequencies, either the Fourier spectrum or the wavelet spectrum of a watermarked version might become very rough and has many sharp positions, called singularities. The situation exists even in adaptive watermarking. Since the singularities can be viewed as the exposure of watermark signals, attackers can directly modify their values to more precisely damage the watermark without any other knowledge. A method that reshapes the spectrum of the keystream into a more ordinary signal before embedding is proposed to eliminate the phenomenon or temper its effect.