For the issue that the robustness of the traditional audio watermarking algorithm based on Empirical Mode Decomposition (EMD) is not strong, an blind audio watermarking algorithm based on the extremum of Intrinsic Mode Function (IMF) was presented. The original audio signal was segmented firstly, and the audio frame was decomposed to a series of IMFs by EMD. Watermarking bits and synchronization code were embedded in the extremum of the last IMF by mean quantization. The embedding payload of the proposed method was 46.9~50.3 b/s, and the watermarked audio signal keeps the perceptive quality of the original audio signal. Several signal attacks such as adding noise, MP3 compression, re-sampling, filtering and cropping were imposed on the watermarked audio. The extracted watermarking bit changed a little, which shows the robustness of the proposed scheme. Compared with time domain and wavelet domain methods, the proposed method can resist 32 kb/s MP3 compression attack with high embedding payload.