🤖 AI Summary
This work addresses the performance limitations of autoregressive (AR) models in audio inpainting. It systematically investigates the optimization mechanisms of key parameters—AR estimator selection, model order, and window size—in Janssen-type methods for high-fidelity reconstruction of missing segments (gaps). We propose the first gap-adapted Janssen algorithm, incorporating gap-aware modeling and adaptive spectral interpolation, and explicitly characterize fundamental differences among AR-based inpainting paradigms. Comprehensive evaluation—using objective metrics (PESQ, STOI) and subjective listening tests—demonstrates consistent superiority over state-of-the-art AR and deep learning baselines across diverse gap lengths (average PESQ gain: +1.23; MOS improvement: +0.8). Our core contribution is the identification of the AR estimator as the dominant factor governing reconstruction quality, coupled with a lightweight, reproducible, and computationally efficient optimization framework.
📝 Abstract
The paper presents an evaluation of popular audio inpainting methods based on autoregressive modeling, namely the extrapolation-based and Janssen methods. A novel variant of the Janssen method suitable for inpainting of gaps is also proposed. The main differences between the particular popular approaches are pointed out. In the experimental part, the importance of the choice of the AR model estimator is confirmed by objective metrics. Then, a mid-scale computational experiment is presented, and its results are confirmed by a listening test. All the experiments demonstrate the superiority of the new gap-wise Janssen method.