Audio Engineer in a Box

is a compact neural network for real-time live music enhancement on resource-constrained devices. Our framework integrates lightweight neural network control with differentiable digital signal processing (DDSP) modules. Unlike traditional methods using autoencoders, music signals bypass the need for embedding and reconstruction, allowing for direct processing. This reduces model size and streamlines the inference pipeline. DDSP modules are essential, enabling gradient descent for network training. We train on a synthetic dataset replicating a live music venue's acoustics using impulse responses (IRs) from a club's PA system and smartphone recordings, simulating a concert experience. This approach ensures the network generalises effectively to real-world scenarios with complex room acoustics and audience noise captured by smartphones. Take a look on GitHub!