Based on the analysis of the audio features, a new audio watermarking algorithm using the discrete multiwavelet transform is proposed. Combined with time-frequency masking property of the human auditory system, the proposed algorithm analyses the zero-cross ratio and the short-time energy of each audio frame to choose the audio frames to embed the watermark. Using the features of sub-sampling and the advantages of multiwavelet in signal processing, each frame to embed the watermark is sub-sampled into two sub-audio frames, and these sub-audio frames are decomposed into multiwavelet domain respectively. According to the energies of two sub-audio frames in multiwavelet domain, the capacity of embedded watermark in audio signal is estimated, and then watermark embedding is accomplished based on the energy relationship between two sub-audio frames. The retrieval of embedded watermark can be considered as a classification problem with two-class that can be solved by support vector machines. The experimental results show that the proposed algorithm can find the suitable audio frames to embed watermark according to the features of the audio signal and can also adjust the embedding strength dynamically improving the robustness of watermarking system without losing auditory quality.