This paper presents an adaptive patchwork algorithm for audio watermarking based on neural network. Two special subsets of the host signal features were selected to embed the watermark signal, adding a small constant value to one subset and subtracting the same from another patch. Then, choose some sample from the embedded audio signal to train a Neural Network. One the extract procedure, the neural network obtained before will be used to verify the watermark information. The method is based on wavelet domain and the watermark signals were embedded in approximation wavelet coefficients. The quality of the watermarked signal is evaluated by PSNR (Peak Signal Noise Ratio) method and extract ratio(ER) after various attacks. Experimental results show that the proposed adaptive patchwork method based on neural network is robust against various common attacks such as filtering, resampling and so on.