Reproduce “Hybrid AHS: A Hybrid of Kalman Filter and Deep Learning for Acoustic Howling Suppression” on OpenResearch

About this paper

Hao Zhang, Meng Yu, Yuzhong Wu, Tao Yu, Dong Yu

A method termed Hybrid AHS integrates a Kalman filter and a deep neural network (SARNN) for acoustic howling suppression. This approach mitigates the mismatch between offline training and real-time streaming inference, resulting in improved howling suppression and better speech quality compared to standalone deep learning or traditional methods across various amplification gains.

Key insights

• Integrating a traditional adaptive filter (Kalman filter) with a deep learning model (SARNN) effectively leverages the strengths of both, addressing the limitations of each standalone approach.
• Using Kalman filter pre-processed signals as auxiliary input for the DNN reduces the mismatch between offline training and real-time streaming inference in recursive feedback loops.
• The DNN architecture, incorporating deep filtering and covariance estimation, enables robust handling of nonlinear distortions and residual howling that traditional linear models struggle with.

Summary by alphaXiv.