A content-based fragile watermark scheme for speech waveform authentication

To detect where the speech waveform is replaced, inserted, or deleted, a content-based fragile watermark scheme is proposed. First, to make sure that the speech features used to generate the watermarks are not affected during the watermark embedding process, an approximated speech waveform is generated by setting some least significant bits (LSBs) being zero. Then speech features, LSP and pitch, are extracted from the approximated speech frame. The conventional watermark scheme in which the features are extracted only from a single frame where the watermarks are embedded is easily attacked and insufficient to detect where and how the speech is faked. Hence, in this work, features from multiple frames and other messages are combined together to generate the watermark. Finally, the watermarks and the synchronization code are embedded into the LSBs of specific speech signals. The theoretical analysis shows the proposed scheme is safe even if the cheater knows how it works. Therefore, the content-based fragile watermark scheme proposed herein can effectively protect the integrity of speech waveforms.