Local active noise control (ANC) algorithms are often based on adaptive filters, requiring an error signal to operate. For optimal performance, this error signal should be captured at a position as close as possible to the desired point of cancellation. However, in many situations, this is not feasible without invading the listener’s space. In such cases, virtual sensing algorithms such as the remote microphone technique (RMT) can be employed to estimate the signal using nearby sensors and domain knowledge. The RMT relies on filters, calculated in a training phase based on recorded scenarios. To ensure sufficient performance under varied acoustic conditions and geometrical configurations, a suitable filter set must be selected during operation from a pre-calculated database encompassing all potential scenarios. This paper proposes a novel approach for an online estimation of the observation filter in the RMT, based on a convolutional neural network. By providing correlation metrics and coordinates as input features, efficient asynchronous computation on external processing units is possible. Handling various acoustic scenarios with variable virtual error microphone position, this approach renders the use of any filter selection logic obsolete.