Digital speech watermarking to enhance the security using speech as a biometric for person authentication

摘要

This work presents the modules for enhancing the security of speaker authentication by embedding the watermark in a speech signal. Speaker is authenticated by speech as well as the extracted watermark from the watermarked speech. Firstly, the speech signal is converted into frames, and discrete wavelet transform is applied to each frame, and it is preferable to embed the watermark in detail coefficients. The segment for embedding the watermark is appropriately chosen based on the energy calculations. The approximation and the modified detail coefficients are used to generate the watermarked speech by inverse discrete wavelet transform. Imperceptibility of the watermark in a watermarked speech is purely depending on the embedding of the watermark. In the receiver, the watermarked speech will undergo wavelet decomposition, and the watermark bits are extracted from the detail coefficients and appropriately transformed into watermark speech/image. The performance the work is evaluated by using the metrics such as Peak signal to noise ratio (PSNR) between original watermark and extracted watermark, PSNR between original speech and watermarked speech and Bit error rate (BER) and Perceptual evaluation speech quality (PESQ). Speaker identification system is assessed by using extraction of the perceptual features and application of features to develop the models for the set of utterances about the speaker during the training phase of the work. Testing is done by applying the original and watermarked speech utterances to the feature extraction phase, followed by we have the testing phase which is used for computing the accuracy. Accuracy is 98.2% for the speaker identification with the set of original test utterances and 98.1% with watermarked set of test utterances and it is observed that there is the marginal difference in accuracy for using speech as a watermark. It is 97.85% for using the image as a watermark. Cover speech signals and watermark speech used in our work are continuous speech utterances chosen from “TIMIT” speech database. Image watermark is the Quick response (QR) code for the LOGO. This work also emphasizes the effectiveness of the algorithm in providing robustness for copyright protection to ownership of the data and authenticating persons using speech as a biometric.