Audio Stream Segmentation Based on Nonlinear Diffusion Filtering (Perona–Malik Approach)

The segmentation of continuous audio streams into distinct units, such as individual news stories, is a critical step for efficient indexing and retrieval in multimedia applications. Existing linear filtering and energy-based approaches are inadequate for preserving sharp amplitude changes, resulting in reduced sensitivity to the speech pauses that define segment boundaries in audio streams. This paper proposes the integration of the Perona-Malik nonlinear diffusion model into the preprocessing phase of an audio segmentation pipeline. Unlike linear smoothing methods, the Perona-Malik scheme adaptively reduces noise while preserving sharp transitions, enabling more accurate detection of speech pauses. The method was evaluated on a dataset comprising twenty manually annotated television news recordings. Pause detection was based on normalized root mean square (RMS) energy computed over diffused signals. Candidate segmentation points were selected by applying duration and energy-based thresholds. Experimental results show that nonlinear diffusion improves the robustness of pause detection, providing an effective alternative preprocessing strategy for broadcast news segmentation. The results confirm that partial differential equation (PDE)-based techniques are suitable for audio stream structuring tasks, offering flexibility and strong performance without requiring complex supervised models.