Obs-TasNet: Online estimation of virtual sensing observation filters for active noise control

Local active noise control (ANC) with adaptive processing requires an accurate residual error signal at the point of cancellation, which is often obtained via virtual sensing. We propose Obs-TasNet, a neural approach that estimates observation filter coefficients for the remote microphone technique (RMT) in time-variant settings. By estimating coefficients during operation, the method eliminates the need for pre-optimizing filters for selected scenarios and subsequent interpolation. Obs-TasNet builds on a modified inter-channel Conv-TasNet. The raw waveform signals from remote microphones, as well as the coordinates of the virtual microphone, are embedded as latent representation using a learnable encoding. A temporal convolutional network (TCN), employing dilated depthwise separable convolutions and an output transformation, predicts the observation filter coefficients. Following a brief hyperparameter search, an ablation study demonstrates that the proposed architectural modifications lead to reduced estimation error while substantially reducing both the number of model parameters and computational cost. In simulation, the proposed ANC system with RMT and Obs-TasNet achieves superior noise reduction over a substantially wider frequency range than a multi-point ANC baseline, validating the effectiveness of observation filter estimation.