Autoregressive Guidance of Deep Spatially Selective Filters using Bayesian Tracking for Efficient Extraction of Moving Speakers

The researchers proposed a new method for depth-space selective filters that combines Bayesian tracking with autoregressive guidance, aiming to efficiently extract moving speakers in dynamic scenarios. This method utilizes the time feedback mechanism in frame-level causal processing to incorporate the enhanced voice signal into lightweight tracking algorithms to optimize performance. The study constructed a synthetic data generation framework based on social force models to enhance the realism of trajectory simulations. Experiments showed that autoregressive fusion significantly improved the accuracy of the Bayesian tracker, achieving excellent enhancement effects with minimal or no computational overhead. Verified through real recordings, this method demonstrated generalized ability to handle unseen acoustic conditions.