A Study on Variable Embedding Locations of Reversible Spectral Speech Watermarking

Rapid advancements in machine learning and speech synthesis have made it feasible to manipulate and tempering audio content. Although watermarking is a useful technique for guaranteeing the integrity of evidential audio data, distortion of sound quality must be prevented. A reversible digital watermarking technique based on the human auditory system was conventionally proposed to embed feature data for tamper detection in the high frequency coefficients. However, this technique has the drawback that spectral analysis can reveal the embedding locations through the borderline. This research explored variable embedding locations while preserving good sound quality. The experimental results show the effectiveness of the proposed to conceal the embedding location, and the distortion are controlled to be imperceptible.