A method of audio watermark embedding by sample bit modification directly in the time domain is proposed. Five maximum consecutive samples in each voiced or unvoiced segments are located and are embedded with a watermark using a key. Voiced segment embedding has better performance in robustness and imperceptibility while the unvoiced segment embedding has higher payload. Both of the proposed methods are better on imperceptibility, as measured by enhanced modified Bark spectrum distortion, and have higher robustness to embedded data under noise-added conditions compared to a similar bit modification method.