A digital audio watermarking based on artificial neural networks is proposed in this paper. Utilizing the learning and adaptive capabilities of artificial neural networks, the relationship between audio signals and embedded watermark is established by using the important characters of audio signals as the input vector of artificial neural networks, and the watermark were embedded into original audio signals without modifying the audio data. The experimental results show that the embedded watermark is robust to audio signal processing, and the watermarking method does not require the original audio signals for watermarking extraction.