Filtering processing is a kind of typical and serious attack on the audio watermarking algorithm. Based on the computation in time domain before and after attacking on the audio, the mean and the standard deviation show good invariant statistical feature. The relation of four consecutive bins in a histogram can keep the change within ?5%. A data range selection approach, which meets the demands of normal distribution, is created generally for the watermark hiding. Based on segmenting idea, an audio watermarking algorithm is designed with statistical features by the stable relation of four consecutive bins. Experimental results show that the algorithm can resist on the attack of low pass filtering with large hiding bandwidth up to 17 bit/s. It shows much better than the existed histogram approach with relations of three consecutive bins.