Speech quality evaluation using digital watermarking

Speech quality evaluation is a very important research topic. The Mean Opinion Score (MOS) is reliable but the listening test is very expensive, time consuming, and even impractical for some applications. Objective quality evaluation methods require either the original speech or a complicated computation model, which makes some applications of quality evaluation impossible. Different from the perceptual model used by the Perceptual Evaluation of Speech Quality (PESQ), in this thesis, we propose to use digital audio watermarking to evaluate the quality of speech. Based on quantization, watermark bits are embedded and extracted in the Discrete Wavelet Transform (DWT) domain. By comparing the original and the extracted watermark, we predict the quality of speech that has undergone MP3 compression, Gaussian noise addition, low-pass filtering, or packet loss. Our quality evaluation method does not need the original signal or a computation model. For the quality evaluation, we use the PESQ MOS as a reference. We predict the speech quality from the PCEW (Percentage of Correctly Extracted Watermark bits) based on the mapping between ITU-T P.862 PESQ MOS and the PCEW. To evaluate the performance of our objective quality evaluation method, we introduce the correlation coefficient and residual error to evaluate the correlation between the predicted MOS and the PESQ MOS. The experiments show that the method yields very promising evaluation results which are very close to the results of the PESQ