Local active noise control (ANC) systems often use adaptive filters that require an error signal to function. For the best results, this error signal should be measured as close as possible to the point where noise cancellation is desired. However, in many cases, this is not practical without intruding into the listener's space. To address this, virtual sensing methods like the remote microphone technique (RMT) can estimate the signal using nearby sensors and knowledge of the environment. The RMT uses filters that are determined during a training phase based on recorded scenarios. To maintain good performance in different acoustic environments and setups, the system must select appropriate filters from a pre-calculated database during operation. This paper introduces a new method for estimating the observation filter in the RMT using a convolutional neural network. By using correlation metrics and coordinates as input, the method enables efficient asynchronous processing on external hardware. This approach can handle various acoustic conditions and virtual microphone positions, eliminating the need for traditional filter selection processes.