Deep learning ANC leverages 3D spatial cues to kill echo in rooms
New SF-GFANC method uses CRNN to track noise sources in 3D space...
Deep Dive
Spatial-frequency cued generative fixed-filter active noise control (SF-GFANC) is proposed, exploiting 3D spatial information (source distance, elevation, and azimuth) and frequency cues via a multi-task convolutional recurrent neural network (CRNN). The system generates control filters guided by these cues, outperforming representative ANC algorithms in reverberant environments on both simulated and measured acoustic paths.
Key Points
- CRNN estimates 3D spatial cues (distance, elevation, azimuth) plus frequency combination weights simultaneously.
- SF-GFANC outperforms prior GFANC and FxLMS methods in reverberant environments on both simulated and measured acoustic paths.
- Theoretical analysis proves that spatially conditioned filter design is essential for optimal noise cancellation in echoic rooms.
Why It Matters
Real-world noise cancellation in offices, cars, and homes finally works even when sound echoes and moves.