This paper proposes a novel dual-domain audio watermarking approach based on flexible segmentation and adaptive embedding aimed to improve robustness and imperceptibility. Compared with conventional watermarking strategies, the proposed approach has two advantages. First, a novel audio beat detection approach is designed to flexibly segment the audio, which provides stronger robustness to synchronization attacks. The audio is decomposed by the discrete wavelet packet transform. Then, the covariance relationships of the decomposition coefficients at different time instants are calculated to determine the locations of the beats and to establish a flexible segmentation model. Second, a dual-domain embedding approach is proposed to realize better robustness to compression attacks while maintaining imperceptibility. In each segment, the psychoacoustic model is used to calculate the audio masking threshold, which divides the signals into the masking signal domain and masked signal domain. The signals in the masking signal domain are robust to compression attacks, and the signals in the masked signal domain have better imperceptibility. To combine these advantages, we embed the watermark into the two domains simultaneously by using the distortion-compensated dither modulation quantization approach. To reduce the impact of the watermark on the original audio, the frequency band with the lowest mask-to-noise ratio is selected as the embedding position for each domain. Moreover, the adaptive quantization steps are calculated to control the embedding strength according to the masking effect. The adaptive embedding will improve the robustness to compression attacks without significantly affecting the original audio quality. The effectiveness of our approach is verified through simulation experiments. INDEX TERMS Audio beats, dual-domain, DWPT, DC-DM, psychoacoustic model.