https://arxiv.org/pdf/2204.11942

Adaptive filtering is a fundamental technique in signal processing, used across various domains such as audio processing. Adaptive filters relies on manually designed iterative optimization methods like Least Mean Squares (LMS) or Recursive Least Squares (RLS), which requires fine-tuning.

Adaptive Filter usage with META-AF:

✔ Dereverberation: Reduces room reflections, makes voices clearer.

✔ Beamforming: Focuses on one speaker, suppresses others.

✔ Source Separation (if added): Splits all voices into separate signals.

Minimum Variance Distortionless Response (MVDR) beamforming using the Generalized Sidelobe Canceller (GSC) formulation. This approach is designed for interference suppression and speech enhancement, used in automatic speech recognition (ASR).

Meta-AF beamforming was compared against: ✔ NLMS (Normalized Least Mean Squares)

✔ RLS (Recursive Least Squares)

✔ MVDR with fixed updates

Results:

• Better Speech Isolation: Meta-AF outperforms traditional beamforming by dynamically adjusting to changing environments.
• Robustness: Works even when the number of microphones varies from training data.