A watermarking scheme based on SVM is implemented and analyzed for two different forms of copyright information, a 34 × 45 binary image and a sound signal for utterance of a word. In this work, the generalization ability of SVM using a basic linear kernel for two different forms of copyright information is analyzed. During extraction, SVM is trained using an extra reference watermark. The perceptual legibility of image and the sound are taken as the objective parameter for analysis. The copyright information is embedded in the positions where the values of the pixels are well within the limit decided by the mean of the surrounding pixels. The perceptual legibility is found to be increased by the above step. The experimental results show the perceptual legibility of sound is better even when the watermarked image is heavily distorted.