This paper mainly introduced the basic knowledge of the digital watermarking and digital audio watermarking,including the definition of digital watermarking and digital audio watermarking,the embedding algorithm of digital audio watermarking and the common attack methods.Digital audio watermarking is very important in the watermark fields.This paper studies a non-linear detector based on Gaussian mixture model for detecting the audio watermark.Using the Expectation Maximization algorithm to calculate the mean and variance of discrete cosine transform coefficient in frequency domain,the mean and variance as the nonlinear detector parameters,and through offset coefficient to describe the detection performance.In attack test we have consider join Gaussian noise,re-sampling,re-quantization,MPEG compression,Chebyshev low-pass filter.Experiments prove this watermark embedding and detection algorithm can improve the robustness of watermarking and ability against the attack.