The goal of any digital audio system is to sample and reconstruct an analog audio signal, without noticeable changes to the original signal. Currently, two major types of reconstruction filters, brickwall and monotonic filters, are used to smooth a sampled analog audio signal during its reconstruction. Brickwall filters work best on reconstruction of smooth signals and the monotonic filters are best for reconstruction of transient signals. Since audio is composed of mixed transient and smooth signals, both of these filters will introduce undesirable artifacts to the signal during its reconstruction. The paper presents a new neural network based dynamic reconstruction filter that can change its behavior to best match the type of signal that is being filtered.
展开▼
