This paper presents speech watermarking techniques based on modification of the duration of estimated silence regions, direct samples replacement within silence regions, dc component changes per short-time frame, and specific modification of low-amplitude samples ranges in a speech signal. These methods could be used to secretly embed information into speech data for applications such as copyright protection, annotations, and covert communications. Natural variability of the modified parameters within a sentence makes the algorithms imperceptible to the human listener while providing enough bandwidth to hide encoded information. Performance of the schemes is evaluated by comparing them to other data watermarking approaches. Experimental results for several speech signals verify the robustness of the algorithm.